Repackage frontend as third-party dependency

third_party/tiny_tts_frontend/tiny_tts/__init__.py

@@ -0,0 +1,90 @@
+import os
+import torch
+import soundfile as sf
+from tiny_tts.text.english import normalize_text, grapheme_to_phoneme
+from tiny_tts.text import phonemes_to_ids
+from tiny_tts.nn import commons
+from tiny_tts.models.synthesizer import VoiceSynthesizer
+from tiny_tts.text.symbols import symbols
+from tiny_tts.utils.config import (
+    SAMPLING_RATE, SEGMENT_FRAMES, ADD_BLANK, SPEC_CHANNELS,
+    N_SPEAKERS, SPK2ID, MODEL_PARAMS,
+)
+from tiny_tts.infer import load_engine
+
+class TinyTTS:
+    def __init__(self, checkpoint_path=None, device=None):
+        if device is None:
+            self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        else:
+            self.device = device
+            
+        if checkpoint_path is None:
+            # Look for default checkpoint in pacakage
+            pkg_dir = os.path.dirname(os.path.abspath(__file__))
+            default_ckpt = os.path.join(os.path.dirname(pkg_dir), "checkpoints", "G.pth")
+            # 2. Check HuggingFace Cache / Download
+            if not os.path.exists(default_ckpt):
+                try:
+                    from huggingface_hub import hf_hub_download
+                    print("Downloading/Loading checkpoint from Hugging Face Hub (backtracking/tiny-tts)...")
+                    default_ckpt = hf_hub_download(repo_id="backtracking/tiny-tts", filename="G.pth")
+                except ImportError:
+                    raise ImportError("huggingface_hub is required to auto-download the model. Run: pip install huggingface_hub")
+                except Exception as e:
+                    raise ValueError(f"Failed to download checkpoint from Hugging Face: {e}")
+
+            checkpoint_path = default_ckpt
+                
+        self.model = load_engine(checkpoint_path, self.device)
+
+    def speak(self, text, output_path="output.wav", speaker="MALE", speed=1.0):
+        """Synthesize text to speech and save to output_path."""
+        print(f"Synthesizing: {text}")
+
+        # Normalize text
+        normalized = normalize_text(text)
+
+        # Phonemize
+        phones, tones, word2ph = grapheme_to_phoneme(normalized)
+
+        # Convert to sequence
+        phone_ids, tone_ids, lang_ids = phonemes_to_ids(phones, tones, "EN")
+
+        # Add blanks
+        if ADD_BLANK:
+            phone_ids = commons.insert_blanks(phone_ids, 0)
+            tone_ids = commons.insert_blanks(tone_ids, 0)
+            lang_ids = commons.insert_blanks(lang_ids, 0)
+
+        x = torch.LongTensor(phone_ids).unsqueeze(0).to(self.device)
+        x_lengths = torch.LongTensor([len(phone_ids)]).to(self.device)
+        tone = torch.LongTensor(tone_ids).unsqueeze(0).to(self.device)
+        language = torch.LongTensor(lang_ids).unsqueeze(0).to(self.device)
+
+        # Speaker ID
+        if speaker not in SPK2ID:
+            print(f"Warning: Speaker '{speaker}' not found, using ID 0. Available: {list(SPK2ID.keys())}")
+            sid = torch.LongTensor([0]).to(self.device)
+        else:
+            sid = torch.LongTensor([SPK2ID[speaker]]).to(self.device)
+
+        # BERT features (disabled - using zero tensors)
+        bert = torch.zeros(1024, len(phone_ids)).to(self.device).unsqueeze(0)
+        ja_bert = torch.zeros(768, len(phone_ids)).to(self.device).unsqueeze(0)
+
+        # speed > 1.0 = faster speech, < 1.0 = slower speech
+        length_scale = 1.0 / speed
+
+        with torch.no_grad():
+            audio, *_ = self.model.infer(
+                x, x_lengths, sid, tone, language, bert, ja_bert,
+                noise_scale=0.667,
+                noise_scale_w=0.8,
+                length_scale=length_scale
+            )
+
+        audio_np = audio[0, 0].cpu().numpy()
+        sf.write(output_path, audio_np, SAMPLING_RATE)
+        print(f"Saved audio to {output_path}")
+        return audio_np

third_party/tiny_tts_frontend/tiny_tts/alignment/__init__.py

@@ -0,0 +1,16 @@
+from numpy import zeros, int32, float32
+from torch import from_numpy
+
+from .core import viterbi_decode_kernel
+
+
+def viterbi_decode(neg_cent, mask):
+    device = neg_cent.device
+    dtype = neg_cent.dtype
+    neg_cent = neg_cent.data.cpu().numpy().astype(float32)
+    path = zeros(neg_cent.shape, dtype=int32)
+
+    t_t_max = mask.sum(1)[:, 0].data.cpu().numpy().astype(int32)
+    t_s_max = mask.sum(2)[:, 0].data.cpu().numpy().astype(int32)
+    viterbi_decode_kernel(path, neg_cent, t_t_max, t_s_max)
+    return from_numpy(path).to(device=device, dtype=dtype)

third_party/tiny_tts_frontend/tiny_tts/alignment/core.py

@@ -0,0 +1,46 @@
+import numba
+
+
+@numba.jit(
+    numba.void(
+        numba.int32[:, :, ::1],
+        numba.float32[:, :, ::1],
+        numba.int32[::1],
+        numba.int32[::1],
+    ),
+    nopython=True,
+    nogil=True,
+)
+def viterbi_decode_kernel(paths, values, t_ys, t_xs):
+    b = paths.shape[0]
+    max_neg_val = -1e9
+    for i in range(int(b)):
+        path = paths[i]
+        value = values[i]
+        t_y = t_ys[i]
+        t_x = t_xs[i]
+
+        v_prev = v_cur = 0.0
+        index = t_x - 1
+
+        for y in range(t_y):
+            for x in range(max(0, t_x + y - t_y), min(t_x, y + 1)):
+                if x == y:
+                    v_cur = max_neg_val
+                else:
+                    v_cur = value[y - 1, x]
+                if x == 0:
+                    if y == 0:
+                        v_prev = 0.0
+                    else:
+                        v_prev = max_neg_val
+                else:
+                    v_prev = value[y - 1, x - 1]
+                value[y, x] += max(v_prev, v_cur)
+
+        for y in range(t_y - 1, -1, -1):
+            path[y, index] = 1
+            if index != 0 and (
+                index == y or value[y - 1, index] < value[y - 1, index - 1]
+            ):
+                index = index - 1

third_party/tiny_tts_frontend/tiny_tts/infer.py

@@ -0,0 +1,191 @@
+import os
+import sys
+import re
+import torch
+import soundfile as sf
+import argparse
+from tiny_tts.text.english import normalize_text, grapheme_to_phoneme
+from tiny_tts.text import phonemes_to_ids
+from tiny_tts.nn import commons
+from tiny_tts.models import VoiceSynthesizer
+from tiny_tts.text.symbols import symbols
+from tiny_tts.utils import (
+    SAMPLING_RATE, SEGMENT_FRAMES, ADD_BLANK, SPEC_CHANNELS,
+    N_SPEAKERS, SPK2ID, MODEL_PARAMS,
+)
+
+
+def load_engine(checkpoint_path, device='cuda'):
+    print(f"Loading model from {checkpoint_path}")
+    net_g = VoiceSynthesizer(
+        len(symbols),
+        SPEC_CHANNELS,
+        SEGMENT_FRAMES,
+        n_speakers=N_SPEAKERS,
+        **MODEL_PARAMS
+    ).to(device)
+
+    # Count model parameters
+    total_params = sum(p.numel() for p in net_g.parameters())
+    trainable_params = sum(p.numel() for p in net_g.parameters() if p.requires_grad)
+    print(f"Model parameters: {total_params/1e6:.2f}M total, {trainable_params/1e6:.2f}M trainable")
+
+    checkpoint = torch.load(checkpoint_path, map_location=device)
+    state_dict = checkpoint['model']
+
+    # Remove module. prefix and filter shape mismatches
+    model_state = net_g.state_dict()
+    new_state_dict = {}
+    skipped = []
+    for k, v in state_dict.items():
+        key = k[7:] if k.startswith('module.') else k
+        if key in model_state:
+            if v.shape == model_state[key].shape:
+                new_state_dict[key] = v
+            else:
+                skipped.append(f"{key}: ckpt{v.shape} vs model{model_state[key].shape}")
+        else:
+            new_state_dict[key] = v
+
+    if skipped:
+        print(f"Skipped {len(skipped)} mismatched keys:")
+        for s in skipped[:5]:
+            print(f"  {s}")
+        if len(skipped) > 5:
+            print(f"  ... and {len(skipped)-5} more")
+
+    net_g.load_state_dict(new_state_dict, strict=False)
+    net_g.eval()
+
+    # Fold weight_norm into weight tensors for faster inference (~18% speedup)
+    net_g.dec.remove_weight_norm()
+
+    return net_g
+
+
+def synthesize(text, output_path, model, speaker="MALE", device='cuda', speed=1.0):
+    print(f"Synthesizing: {text}")
+
+    # Normalize text
+    normalized = normalize_text(text)
+
+    # Phonemize
+    phones, tones, word2ph = grapheme_to_phoneme(normalized)
+
+    # Convert to sequence
+    phone_ids, tone_ids, lang_ids = phonemes_to_ids(phones, tones, "EN")
+
+    # Add blanks
+    if ADD_BLANK:
+        phone_ids = commons.insert_blanks(phone_ids, 0)
+        tone_ids = commons.insert_blanks(tone_ids, 0)
+        lang_ids = commons.insert_blanks(lang_ids, 0)
+
+    x = torch.LongTensor(phone_ids).unsqueeze(0).to(device)
+    x_lengths = torch.LongTensor([len(phone_ids)]).to(device)
+    tone = torch.LongTensor(tone_ids).unsqueeze(0).to(device)
+    language = torch.LongTensor(lang_ids).unsqueeze(0).to(device)
+
+    # Speaker ID
+    if speaker not in SPK2ID:
+        print(f"Warning: Speaker {speaker} not found, using ID 0")
+        sid = torch.LongTensor([0]).to(device)
+    else:
+        sid = torch.LongTensor([SPK2ID[speaker]]).to(device)
+
+    # BERT features (disabled - using zero tensors)
+    bert = torch.zeros(1024, len(phone_ids)).to(device).unsqueeze(0)
+    ja_bert = torch.zeros(768, len(phone_ids)).to(device).unsqueeze(0)
+
+    # speed > 1.0 = faster speech, < 1.0 = slower speech
+    length_scale = 1.0 / speed
+
+    with torch.no_grad():
+        audio, *_ = model.infer(
+            x, x_lengths, sid, tone, language, bert, ja_bert,
+            noise_scale=0.667,
+            noise_scale_w=0.8,
+            length_scale=length_scale
+        )
+
+    audio = audio[0, 0].cpu().numpy()
+    sf.write(output_path, audio, SAMPLING_RATE)
+    print(f"Saved audio to {output_path}")
+
+
+def get_latest_checkpoint(checkpoint_dir):
+    """Finds the latest G_*.pth checkpoint in the given directory."""
+    checkpoints = [f for f in os.listdir(checkpoint_dir) if f.startswith('G_') and f.endswith('.pth')]
+    if not checkpoints:
+        return None
+
+    def get_step(filename):
+        match = re.search(r'_(\d+)\.pth', filename)
+        return int(match.group(1)) if match else -1
+
+    latest_ckpt = max(checkpoints, key=get_step)
+    return os.path.join(checkpoint_dir, latest_ckpt)
+
+
+def main():
+    parser = argparse.ArgumentParser(description="TinyTTS — English Text-to-Speech Inference")
+    parser.add_argument("--text", "-t", type=str, default="The weather is nice today, and I feel very relaxed.", help="Text to synthesize")
+    parser.add_argument("--checkpoint", "-c", type=str, default=None, help="Path to checkpoint. Auto-downloads if not provided.")
+    parser.add_argument("--output", "-o", type=str, default="output.wav", help="Output audio file path")
+    parser.add_argument("--speaker", "-s", type=str, default="MALE", help="Speaker ID")
+    parser.add_argument("--speed", type=float, default=1.0, help="Speech speed (1.0=normal, 1.5=faster, 0.7=slower)")
+    parser.add_argument("--device", type=str, default="cuda", help="Device to use (cuda or cpu)")
+
+    args = parser.parse_args()
+
+    if args.checkpoint is None:
+        try:
+            from huggingface_hub import hf_hub_download
+            print("Downloading/Loading checkpoint from Hugging Face Hub (backtracking/tiny-tts)...")
+            args.checkpoint = hf_hub_download(repo_id="backtracking/tiny-tts", filename="G.pth")
+        except ImportError:
+            print("Error: huggingface_hub is required for auto-download. Run: pip install huggingface_hub")
+            sys.exit(1)
+        except Exception as e:
+            print(f"Error downloading checkpoint: {e}")
+            sys.exit(1)
+
+    if not os.path.exists(args.checkpoint):
+        print(f"Error: Checkpoint or directory not found at {args.checkpoint}")
+        sys.exit(1)
+
+    if os.path.isdir(args.checkpoint):
+        latest_ckpt = get_latest_checkpoint(args.checkpoint)
+        if not latest_ckpt:
+            print(f"Error: No G_*.pth checkpoints found in directory {args.checkpoint}")
+            sys.exit(1)
+        args.checkpoint = latest_ckpt
+        print(f"Auto-detected latest checkpoint: {args.checkpoint}")
+
+    # Extract step from checkpoint filename
+    ckpt_basename = os.path.basename(args.checkpoint)
+    match = re.search(r'_(\d+)\.pth', ckpt_basename)
+    step_str = match.group(1) if match else "unknown"
+
+    # Save to output folder
+    out_dir = "infer_outputs"
+    os.makedirs(out_dir, exist_ok=True)
+
+    out_name = os.path.basename(args.output)
+    name, ext = os.path.splitext(out_name)
+    model = load_engine(args.checkpoint, args.device)
+
+    if args.speaker.lower() == "all":
+        if not SPK2ID:
+            print("Error: No speakers found")
+            sys.exit(1)
+        print(f"Synthesizing for all {len(SPK2ID)} speakers...")
+        for spk in SPK2ID.keys():
+            final_output = os.path.join(out_dir, f"{name}_step{step_str}_spk{spk}{ext}")
+            synthesize(args.text, final_output, model, speaker=spk, device=args.device, speed=args.speed)
+    else:
+        final_output = os.path.join(out_dir, f"{name}_step{step_str}_spk{args.speaker}{ext}")
+        synthesize(args.text, final_output, model, speaker=args.speaker, device=args.device, speed=args.speed)
+
+if __name__ == "__main__":
+    main()

third_party/tiny_tts_frontend/tiny_tts/infer_onnx.py

@@ -0,0 +1,199 @@
+"""
+ONNX Runtime inference engine for TinyTTS.
+
+Replaces the PyTorch VoiceSynthesizer.infer() with equivalent
+ONNX Runtime sessions + NumPy ops for the non-exported parts
+(alignment path computation).
+"""
+import os
+import numpy as np
+import soundfile as sf
+
+from tiny_tts.text.english import normalize_text, grapheme_to_phoneme
+from tiny_tts.text import phonemes_to_ids
+from tiny_tts.nn import commons
+from tiny_tts.utils.config import (
+    SAMPLING_RATE, ADD_BLANK, SPK2ID,
+)
+
+try:
+    import onnxruntime as ort
+except ImportError:
+    raise ImportError("onnxruntime is required. Run: pip install onnxruntime")
+
+
+def _build_session(path: str, use_gpu: bool = False):
+    """Create an ORT InferenceSession with optional GPU support."""
+    providers = (
+        ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        if use_gpu else
+        ["CPUExecutionProvider"]
+    )
+    opts = ort.SessionOptions()
+    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+    opts.intra_op_num_threads = os.cpu_count() or 4
+    return ort.InferenceSession(path, sess_options=opts, providers=providers)
+
+
+def _create_length_mask_np(lengths, max_len=None):
+    """NumPy equivalent of commons.create_length_mask."""
+    if max_len is None:
+        max_len = int(lengths.max())
+    ids = np.arange(max_len, dtype=np.float32)     # [T]
+    mask = (ids[None, :] < lengths[:, None]).astype(np.float32)  # [B, T]
+    return mask
+
+
+def _compute_alignment_path_np(w_ceil, attn_mask):
+    """
+    Monotonic alignment path - vectorized via cumsum (much faster than Python loops).
+    w_ceil:    [B, 1, T_x] — integer duration per phone
+    attn_mask: [B, 1, T_y, T_x] — joint mask
+    Returns attn: [B, 1, T_y, T_x]
+    """
+    B, _, T_x = w_ceil.shape
+    T_y = attn_mask.shape[2]
+
+    # Build duration matrix: for each phone column expand the duration
+    # cumulative sum of durations gives us the end frame index for each phone
+    dur = w_ceil[:, 0, :]                            # [B, T_x]
+    cum_dur = np.cumsum(dur, axis=1)                 # [B, T_x]  — end frame (1-indexed)
+    cum_dur_prev = np.pad(cum_dur[:, :-1], ((0,0),(1,0)))  # [B, T_x] — start frame
+
+    # Frame indices: [1, T_y, 1]
+    frame_idx = np.arange(T_y, dtype=np.float32)[None, :, None]   # [1, T_y, 1]
+    # For each phone, mark frames [start, end)
+    # cum_dur_prev: [B,1,T_x], cum_dur: [B,1,T_x]
+    start = cum_dur_prev[:, None, :]                 # [B, 1, T_x]
+    end   = cum_dur[:, None, :]                      # [B, 1, T_x]
+    attn  = ((frame_idx >= start) & (frame_idx < end)).astype(np.float32)  # [B, T_y, T_x]
+    attn  = attn[:, None, :, :]                      # [B, 1, T_y, T_x]
+    return attn * attn_mask
+
+
+class OnnxTinyTTS:
+    """
+    Inference using ONNX Runtime.
+
+    Args:
+        onnx_dir: directory containing the 4 .onnx files
+        use_gpu:  if True, try CUDAExecutionProvider
+    """
+
+    def __init__(self, onnx_dir: str = "onnx", use_gpu: bool = False):
+        onnx_dir = os.path.abspath(onnx_dir)
+        print(f"Loading ONNX sessions from: {onnx_dir}")
+
+        self._enc  = _build_session(os.path.join(onnx_dir, "text_encoder.onnx"), use_gpu)
+        self._dp   = _build_session(os.path.join(onnx_dir, "duration_predictor.onnx"), use_gpu)
+        self._flow = _build_session(os.path.join(onnx_dir, "flow.onnx"), use_gpu)
+        self._dec  = _build_session(os.path.join(onnx_dir, "decoder.onnx"), use_gpu)
+
+        print("ONNX sessions ready ✅")
+
+    def _text_to_ids(self, text: str):
+        normalized = normalize_text(text)
+        phones, tones, _ = grapheme_to_phoneme(normalized)
+        phone_ids, tone_ids, lang_ids = phonemes_to_ids(phones, tones, "EN")
+
+        if ADD_BLANK:
+            phone_ids = commons.insert_blanks(phone_ids, 0)
+            tone_ids  = commons.insert_blanks(tone_ids, 0)
+            lang_ids  = commons.insert_blanks(lang_ids, 0)
+
+        return phone_ids, tone_ids, lang_ids
+
+    def speak(
+        self,
+        text: str,
+        output_path: str = "onnx_output.wav",
+        speaker: str = "female",
+        noise_scale: float = 0.667,
+        noise_scale_w: float = 0.8,
+        length_scale: float = 1.0,
+        output_sr: int = None,
+    ) -> np.ndarray:
+        """Synthesize speech and save to output_path.
+        
+        Args:
+            output_sr: If set (e.g. 22050), resample the output from 44100 Hz.
+                       Useful to reduce file size while keeping quality.
+        """
+        print(f"[ONNX] Synthesizing: {text}")
+
+        phone_ids, tone_ids, lang_ids = self._text_to_ids(text)
+        T = len(phone_ids)
+
+        # Prepare inputs as float32 / int64 arrays
+        x       = np.array(phone_ids, dtype=np.int64)[None, :]        # [1, T]
+        x_len   = np.array([T], dtype=np.int64)                        # [1]
+        tone    = np.array(tone_ids, dtype=np.int64)[None, :]          # [1, T]
+        lang    = np.array(lang_ids, dtype=np.int64)[None, :]          # [1, T]
+        bert    = np.zeros((1, 1024, T), dtype=np.float32)
+        ja_bert = np.zeros((1, 768,  T), dtype=np.float32)
+        sid_val = SPK2ID.get(speaker, 0)
+        sid     = np.array([sid_val], dtype=np.int64)                  # [1]
+
+        # ── 1. Text Encoder ──────────────────────────────────────────────
+        x_enc, m_p, logs_p, x_mask, g = self._enc.run(
+            None,
+            {
+                "phone_ids":    x,
+                "phone_lengths":x_len,
+                "tone_ids":     tone,
+                "language_ids": lang,
+                "bert":         bert,
+                "ja_bert":      ja_bert,
+                "speaker_id":   sid,
+            },
+        )
+
+        # ── 2. Duration Predictor ─────────────────────────────────────────
+        logw = self._dp.run(None, {"x": x_enc, "x_mask": x_mask, "g": g})[0]
+
+        # ── 3. Alignment Path (NumPy) ─────────────────────────────────────
+        w        = np.exp(logw) * x_mask * length_scale          # [1, 1, T]
+        w_ceil   = np.ceil(w)                                    # [1, 1, T]
+        y_len    = max(1, int(w_ceil.sum()))
+        y_lens   = np.array([y_len], dtype=np.int64)
+
+        y_mask   = _create_length_mask_np(y_lens, y_len)         # [1, T_y]
+        y_mask   = y_mask[:, None, :]                            # [1, 1, T_y]
+        # attn_mask: [1, 1, T_y, T_x]  (outer product of frame mask and phone mask)
+        attn_mask = y_mask[:, :, :, None] * x_mask[:, :, None, :]   # [1,1,T_y,T_x]
+        attn     = _compute_alignment_path_np(w_ceil, attn_mask) # [1, 1, T_y, T_x]
+
+        # Expand prior stats via alignment
+        m_p_exp    = np.matmul(attn[:, 0], m_p.transpose(0, 2, 1)).transpose(0, 2, 1)
+        logs_p_exp = np.matmul(attn[:, 0], logs_p.transpose(0, 2, 1)).transpose(0, 2, 1)
+
+        # ── 4. Sample z_p ─────────────────────────────────────────────────
+        z_p = m_p_exp + np.random.randn(*m_p_exp.shape).astype(np.float32) * \
+              np.exp(logs_p_exp) * noise_scale
+
+        # ── 5. Flow (reverse) ─────────────────────────────────────────────
+        z = self._flow.run(
+            None,
+            {"z_p": z_p, "y_mask": y_mask.astype(np.float32), "g": g},
+        )[0]
+
+        # ── 6. Decoder ────────────────────────────────────────────────────
+        z_masked = (z * y_mask).astype(np.float32)
+        audio = self._dec.run(None, {"z": z_masked, "g": g})[0]  # [1, 1, samples]
+
+        audio_np = audio[0, 0]
+        save_sr = SAMPLING_RATE
+        if output_sr is not None and output_sr != SAMPLING_RATE:
+            try:
+                import torchaudio
+                import torch
+                wav_t = torch.from_numpy(audio_np).unsqueeze(0)
+                resampler = torchaudio.transforms.Resample(SAMPLING_RATE, output_sr)
+                audio_np = resampler(wav_t).squeeze(0).numpy()
+                save_sr = output_sr
+            except Exception as e:
+                print(f"[ONNX] Resampling failed ({e}), saving at {SAMPLING_RATE}Hz")
+
+        sf.write(output_path, audio_np, save_sr)
+        print(f"[ONNX] Saved: {output_path} ({save_sr}Hz)")
+        return audio_np

third_party/tiny_tts_frontend/tiny_tts/models/__init__.py

@@ -0,0 +1 @@
+from .synthesizer import VoiceSynthesizer

third_party/tiny_tts_frontend/tiny_tts/models/synthesizer.py

@@ -0,0 +1,718 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from tiny_tts.nn import commons
+from tiny_tts.nn import modules
+from tiny_tts.nn import attentions
+
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+from tiny_tts.nn.commons import initialize_weights, compute_padding
+import tiny_tts.alignment as alignment
+
+
+class AttentionFlowBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+        share_parameter=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+
+        self.wn = (
+            attentions.FeedForward(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=self.gin_channels,
+            )
+            if share_parameter
+            else None
+        )
+
+        for i in range(n_flows):
+            self.flows.append(
+                modules.TransformerCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    n_layers,
+                    n_heads,
+                    p_dropout,
+                    filter_channels,
+                    mean_only=True,
+                    wn_sharing_parameter=self.wn,
+                    gin_channels=self.gin_channels,
+                )
+            )
+            self.flows.append(modules.FlipTransform())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class VariationalDurationModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        filter_channels = in_channels
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.log_flow = modules.LogTransform()
+        self.flows = nn.ModuleList()
+        self.flows.append(modules.AffineCoupling(2))
+        for i in range(n_flows):
+            self.flows.append(
+                modules.ConvolutionalFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.flows.append(modules.FlipTransform())
+
+        self.post_pre = nn.Conv1d(1, filter_channels, 1)
+        self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.post_convs = modules.DepthwiseSepConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        self.post_flows = nn.ModuleList()
+        self.post_flows.append(modules.AffineCoupling(2))
+        for i in range(4):
+            self.post_flows.append(
+                modules.ConvolutionalFlow(2, filter_channels, kernel_size, n_layers=3)
+            )
+            self.post_flows.append(modules.FlipTransform())
+
+        self.pre = nn.Conv1d(in_channels, filter_channels, 1)
+        self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
+        self.convs = modules.DepthwiseSepConv(
+            filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout
+        )
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
+
+    def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
+        x = torch.detach(x)
+        x = self.pre(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.convs(x, x_mask)
+        x = self.proj(x) * x_mask
+
+        if not reverse:
+            flows = self.flows
+            assert w is not None
+
+            logdet_tot_q = 0
+            h_w = self.post_pre(w)
+            h_w = self.post_convs(h_w, x_mask)
+            h_w = self.post_proj(h_w) * x_mask
+            e_q = (
+                torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype)
+                * x_mask
+            )
+            z_q = e_q
+            for flow in self.post_flows:
+                z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
+                logdet_tot_q += logdet_q
+            z_u, z1 = torch.split(z_q, [1, 1], 1)
+            u = torch.sigmoid(z_u) * x_mask
+            z0 = (w - u) * x_mask
+            logdet_tot_q += torch.sum(
+                (F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2]
+            )
+            logq = (
+                torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q**2)) * x_mask, [1, 2])
+                - logdet_tot_q
+            )
+
+            logdet_tot = 0
+            z0, logdet = self.log_flow(z0, x_mask)
+            logdet_tot += logdet
+            z = torch.cat([z0, z1], 1)
+            for flow in flows:
+                z, logdet = flow(z, x_mask, g=x, reverse=reverse)
+                logdet_tot = logdet_tot + logdet
+            nll = (
+                torch.sum(0.5 * (math.log(2 * math.pi) + (z**2)) * x_mask, [1, 2])
+                - logdet_tot
+            )
+            return nll + logq
+        else:
+            flows = list(reversed(self.flows))
+            flows = flows[:-2] + [flows[-1]]
+            z = (
+                torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype)
+                * noise_scale
+            )
+            for flow in flows:
+                z = flow(z, x_mask, g=x, reverse=reverse)
+            z0, z1 = torch.split(z, [1, 1], 1)
+            logw = z0
+            return logw
+
+
+class DurationEstimator(nn.Module):
+    def __init__(
+        self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0
+    ):
+        super().__init__()
+
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+
+        self.drop = nn.Dropout(p_dropout)
+        self.conv_1 = nn.Conv1d(
+            in_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_1 = modules.ChannelNorm(filter_channels)
+        self.conv_2 = nn.Conv1d(
+            filter_channels, filter_channels, kernel_size, padding=kernel_size // 2
+        )
+        self.norm_2 = modules.ChannelNorm(filter_channels)
+        self.proj = nn.Conv1d(filter_channels, 1, 1)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, in_channels, 1)
+
+    def forward(self, x, x_mask, g=None):
+        x = torch.detach(x)
+        if g is not None:
+            g = torch.detach(g)
+            x = x + self.cond(g)
+        x = self.conv_1(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_1(x)
+        x = self.drop(x)
+        x = self.conv_2(x * x_mask)
+        x = torch.relu(x)
+        x = self.norm_2(x)
+        x = self.drop(x)
+        x = self.proj(x * x_mask)
+        return x * x_mask
+
+
+class PhonemeEncoder(nn.Module):
+    def __init__(
+        self,
+        n_vocab,
+        out_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        gin_channels=0,
+        num_languages=None,
+        num_tones=None,
+    ):
+        super().__init__()
+        if num_languages is None:
+            from tiny_tts.text import num_languages
+        if num_tones is None:
+            from tiny_tts.text import num_tones
+        self.n_vocab = n_vocab
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.gin_channels = gin_channels
+        self.emb = nn.Embedding(n_vocab, hidden_channels)
+        nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
+        self.tone_emb = nn.Embedding(num_tones, hidden_channels)
+        nn.init.normal_(self.tone_emb.weight, 0.0, hidden_channels**-0.5)
+        self.language_emb = nn.Embedding(num_languages, hidden_channels)
+        nn.init.normal_(self.language_emb.weight, 0.0, hidden_channels**-0.5)
+        self.bert_proj = nn.Conv1d(1024, hidden_channels, 1)
+        self.ja_bert_proj = nn.Conv1d(768, hidden_channels, 1)
+
+        self.encoder = attentions.TransformerBlock(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, tone, language, bert, ja_bert, g=None):
+        bert_emb = self.bert_proj(bert).transpose(1, 2)
+        ja_bert_emb = self.ja_bert_proj(ja_bert).transpose(1, 2)
+        x = (
+            self.emb(x)
+            + self.tone_emb(tone)
+            + self.language_emb(language)
+            + bert_emb
+            + ja_bert_emb
+        ) * math.sqrt(
+            self.hidden_channels
+        )
+        x = torch.transpose(x, 1, -1)
+        x_mask = torch.unsqueeze(commons.create_length_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+
+        x = self.encoder(x * x_mask, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        return x, m, logs, x_mask
+
+
+class FlowBlock(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        n_flows=4,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.n_flows = n_flows
+        self.gin_channels = gin_channels
+
+        self.flows = nn.ModuleList()
+        for i in range(n_flows):
+            self.flows.append(
+                modules.FlowCouplingLayer(
+                    channels,
+                    hidden_channels,
+                    kernel_size,
+                    dilation_rate,
+                    n_layers,
+                    gin_channels=gin_channels,
+                    mean_only=True,
+                )
+            )
+            self.flows.append(modules.FlipTransform())
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        if not reverse:
+            for flow in self.flows:
+                x, _ = flow(x, x_mask, g=g, reverse=reverse)
+        else:
+            for flow in reversed(self.flows):
+                x = flow(x, x_mask, g=g, reverse=reverse)
+        return x
+
+
+class LatentEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+
+        self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+        self.enc = modules.WaveNet(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            gin_channels=gin_channels,
+        )
+        self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+    def forward(self, x, x_lengths, g=None, tau=1.0):
+        x_mask = torch.unsqueeze(commons.create_length_mask(x_lengths, x.size(2)), 1).to(
+            x.dtype
+        )
+        x = self.pre(x) * x_mask
+        x = self.enc(x, x_mask, g=g)
+        stats = self.proj(x) * x_mask
+        m, logs = torch.split(stats, self.out_channels, dim=1)
+        z = (m + torch.randn_like(m) * tau * torch.exp(logs)) * x_mask
+        return z, m, logs, x_mask
+
+
+class WaveformDecoder(torch.nn.Module):
+    def __init__(
+        self,
+        initial_channel,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        gin_channels=0,
+    ):
+        super(WaveformDecoder, self).__init__()
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        self.conv_pre = Conv1d(
+            initial_channel, upsample_initial_channel, 7, 1, padding=3
+        )
+        resblock = modules.ConvResBlock if resblock == "1" else modules.ConvResBlockLight
+
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(resblock_kernel_sizes, resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(ch, k, d))
+
+        self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+        self.ups.apply(initialize_weights)
+
+        if gin_channels != 0:
+            self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+    def forward(self, x, g=None):
+        x = self.conv_pre(x)
+        if g is not None:
+            x = x + self.cond(g)
+
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for layer in self.ups:
+            remove_weight_norm(layer)
+        for layer in self.resblocks:
+            layer.remove_weight_norm()
+
+
+class StyleEncoder(nn.Module):
+    def __init__(self, spec_channels, gin_channels=0, layernorm=False):
+        super().__init__()
+        self.spec_channels = spec_channels
+        ref_enc_filters = [32, 32, 64, 64, 128, 128]
+        K = len(ref_enc_filters)
+        filters = [1] + ref_enc_filters
+        convs = [
+            weight_norm(
+                nn.Conv2d(
+                    in_channels=filters[i],
+                    out_channels=filters[i + 1],
+                    kernel_size=(3, 3),
+                    stride=(2, 2),
+                    padding=(1, 1),
+                )
+            )
+            for i in range(K)
+        ]
+        self.convs = nn.ModuleList(convs)
+
+        out_channels = self.calculate_channels(spec_channels, 3, 2, 1, K)
+        self.gru = nn.GRU(
+            input_size=ref_enc_filters[-1] * out_channels,
+            hidden_size=256 // 2,
+            batch_first=True,
+        )
+        self.proj = nn.Linear(128, gin_channels)
+        if layernorm:
+            self.layernorm = nn.LayerNorm(self.spec_channels)
+        else:
+            self.layernorm = None
+
+    def forward(self, inputs, mask=None):
+        N = inputs.size(0)
+
+        out = inputs.view(N, 1, -1, self.spec_channels)
+        if self.layernorm is not None:
+            out = self.layernorm(out)
+
+        for conv in self.convs:
+            out = conv(out)
+            out = F.relu(out)
+
+        out = out.transpose(1, 2)
+        T = out.size(1)
+        N = out.size(0)
+        out = out.contiguous().view(N, T, -1)
+
+        self.gru.flatten_parameters()
+        memory, out = self.gru(out)
+
+        return self.proj(out.squeeze(0))
+
+    def calculate_channels(self, L, kernel_size, stride, pad, n_convs):
+        for i in range(n_convs):
+            L = (L - kernel_size + 2 * pad) // stride + 1
+        return L
+
+
+class VoiceSynthesizer(nn.Module):
+    """Voice synthesis model for inference."""
+
+    def __init__(
+        self,
+        n_vocab,
+        spec_channels,
+        segment_size,
+        inter_channels,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        resblock,
+        resblock_kernel_sizes,
+        resblock_dilation_sizes,
+        upsample_rates,
+        upsample_initial_channel,
+        upsample_kernel_sizes,
+        n_speakers=256,
+        gin_channels=256,
+        use_sdp=True,
+        n_flow_layer=4,
+        n_layers_trans_flow=6,
+        flow_share_parameter=False,
+        use_transformer_flow=True,
+        use_vc=False,
+        num_languages=None,
+        num_tones=None,
+        norm_refenc=False,
+        **kwargs
+    ):
+        super().__init__()
+        self.n_vocab = n_vocab
+        self.spec_channels = spec_channels
+        self.inter_channels = inter_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.resblock = resblock
+        self.resblock_kernel_sizes = resblock_kernel_sizes
+        self.resblock_dilation_sizes = resblock_dilation_sizes
+        self.upsample_rates = upsample_rates
+        self.upsample_initial_channel = upsample_initial_channel
+        self.upsample_kernel_sizes = upsample_kernel_sizes
+        self.segment_size = segment_size
+        self.n_speakers = n_speakers
+        self.gin_channels = gin_channels
+        self.n_layers_trans_flow = n_layers_trans_flow
+        self.use_spk_conditioned_encoder = kwargs.get(
+            "use_spk_conditioned_encoder", True
+        )
+        self.use_sdp = use_sdp
+        self.use_noise_scaled_mas = kwargs.get("use_noise_scaled_mas", False)
+        self.mas_noise_scale_initial = kwargs.get("mas_noise_scale_initial", 0.01)
+        self.noise_scale_delta = kwargs.get("noise_scale_delta", 2e-6)
+        self.current_mas_noise_scale = self.mas_noise_scale_initial
+        if self.use_spk_conditioned_encoder and gin_channels > 0:
+            self.enc_gin_channels = gin_channels
+        else:
+            self.enc_gin_channels = 0
+        self.enc_p = PhonemeEncoder(
+            n_vocab,
+            inter_channels,
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout,
+            gin_channels=self.enc_gin_channels,
+            num_languages=num_languages,
+            num_tones=num_tones,
+        )
+        self.dec = WaveformDecoder(
+            inter_channels,
+            resblock,
+            resblock_kernel_sizes,
+            resblock_dilation_sizes,
+            upsample_rates,
+            upsample_initial_channel,
+            upsample_kernel_sizes,
+            gin_channels=gin_channels,
+        )
+        self.enc_q = LatentEncoder(
+            spec_channels,
+            inter_channels,
+            hidden_channels,
+            5,
+            1,
+            16,
+            gin_channels=gin_channels,
+        )
+        if use_transformer_flow:
+            self.flow = AttentionFlowBlock(
+                inter_channels,
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers_trans_flow,
+                5,
+                p_dropout,
+                n_flow_layer,
+                gin_channels=gin_channels,
+                share_parameter=flow_share_parameter,
+            )
+        else:
+            self.flow = FlowBlock(
+                inter_channels,
+                hidden_channels,
+                5,
+                1,
+                n_flow_layer,
+                gin_channels=gin_channels,
+            )
+        self.sdp = VariationalDurationModel(
+            hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels
+        )
+        self.dp = DurationEstimator(
+            hidden_channels, 256, 3, 0.5, gin_channels=gin_channels
+        )
+
+        if n_speakers > 0:
+            self.emb_g = nn.Embedding(n_speakers, gin_channels)
+        else:
+            self.ref_enc = StyleEncoder(spec_channels, gin_channels, layernorm=norm_refenc)
+        self.use_vc = use_vc
+
+    def infer(
+        self,
+        x,
+        x_lengths,
+        sid,
+        tone,
+        language,
+        bert,
+        ja_bert,
+        noise_scale=0.667,
+        length_scale=1,
+        noise_scale_w=0.8,
+        max_len=None,
+        sdp_ratio=0,
+        y=None,
+        g=None,
+    ):
+        if g is None:
+            if self.n_speakers > 0:
+                g = self.emb_g(sid).unsqueeze(-1)
+            else:
+                g = self.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
+        if self.use_vc:
+            g_p = None
+        else:
+            g_p = g
+        x, m_p, logs_p, x_mask = self.enc_p(
+            x, x_lengths, tone, language, bert, ja_bert, g=g_p
+        )
+        logw = self.sdp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w) * (
+            sdp_ratio
+        ) + self.dp(x, x_mask, g=g) * (1 - sdp_ratio)
+        w = torch.exp(logw) * x_mask * length_scale
+
+        w_ceil = torch.ceil(w)
+        y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
+        y_mask = torch.unsqueeze(commons.create_length_mask(y_lengths, None), 1).to(
+            x_mask.dtype
+        )
+        attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
+        attn = commons.compute_alignment_path(w_ceil, attn_mask)
+
+        m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(
+            1, 2
+        )
+        logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(
+            1, 2
+        )
+
+        z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
+        z = self.flow(z_p, y_mask, g=g, reverse=True)
+        o = self.dec((z * y_mask)[:, :, :max_len], g=g)
+        return o, attn, y_mask, (z, z_p, m_p, logs_p)

third_party/tiny_tts_frontend/tiny_tts/nn/__init__.py

@@ -0,0 +1 @@
+# Neural network building blocks

third_party/tiny_tts_frontend/tiny_tts/nn/attentions.py

@@ -0,0 +1,424 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from . import commons
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class ChannelLayerNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+@torch.jit.script
+def gated_activation(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+class TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        window_size=4,
+        isflow=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+
+        self.cond_layer_idx = self.n_layers
+        if "gin_channels" in kwargs:
+            self.gin_channels = kwargs["gin_channels"]
+            if self.gin_channels != 0:
+                self.spk_emb_linear = nn.Linear(self.gin_channels, self.hidden_channels)
+                self.cond_layer_idx = (
+                    kwargs["cond_layer_idx"] if "cond_layer_idx" in kwargs else 2
+                )
+                assert (
+                    self.cond_layer_idx < self.n_layers
+                ), "cond_layer_idx should be less than n_layers"
+        self.drop = nn.Dropout(p_dropout)
+        self.attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+
+        for i in range(self.n_layers):
+            self.attn_layers.append(
+                MultiHeadSelfAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    window_size=window_size,
+                )
+            )
+            self.norm_layers_1.append(ChannelLayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FeedForward(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                )
+            )
+            self.norm_layers_2.append(ChannelLayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, g=None):
+        attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            if i == self.cond_layer_idx and g is not None:
+                g = self.spk_emb_linear(g.transpose(1, 2))
+                g = g.transpose(1, 2)
+                x = x + g
+                x = x * x_mask
+            y = self.attn_layers[i](x, x, attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size=1,
+        p_dropout=0.0,
+        proximal_bias=False,
+        proximal_init=True,
+        **kwargs
+    ):
+        super().__init__()
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+
+        self.drop = nn.Dropout(p_dropout)
+        self.self_attn_layers = nn.ModuleList()
+        self.norm_layers_0 = nn.ModuleList()
+        self.encdec_attn_layers = nn.ModuleList()
+        self.norm_layers_1 = nn.ModuleList()
+        self.ffn_layers = nn.ModuleList()
+        self.norm_layers_2 = nn.ModuleList()
+        for i in range(self.n_layers):
+            self.self_attn_layers.append(
+                MultiHeadSelfAttention(
+                    hidden_channels,
+                    hidden_channels,
+                    n_heads,
+                    p_dropout=p_dropout,
+                    proximal_bias=proximal_bias,
+                    proximal_init=proximal_init,
+                )
+            )
+            self.norm_layers_0.append(ChannelLayerNorm(hidden_channels))
+            self.encdec_attn_layers.append(
+                MultiHeadSelfAttention(
+                    hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout
+                )
+            )
+            self.norm_layers_1.append(ChannelLayerNorm(hidden_channels))
+            self.ffn_layers.append(
+                FeedForward(
+                    hidden_channels,
+                    hidden_channels,
+                    filter_channels,
+                    kernel_size,
+                    p_dropout=p_dropout,
+                    causal=True,
+                )
+            )
+            self.norm_layers_2.append(ChannelLayerNorm(hidden_channels))
+
+    def forward(self, x, x_mask, h, h_mask):
+        self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(
+            device=x.device, dtype=x.dtype
+        )
+        encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+        x = x * x_mask
+        for i in range(self.n_layers):
+            y = self.self_attn_layers[i](x, x, self_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_0[i](x + y)
+
+            y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+            y = self.drop(y)
+            x = self.norm_layers_1[i](x + y)
+
+            y = self.ffn_layers[i](x, x_mask)
+            y = self.drop(y)
+            x = self.norm_layers_2[i](x + y)
+        x = x * x_mask
+        return x
+
+
+class MultiHeadSelfAttention(nn.Module):
+    def __init__(
+        self,
+        channels,
+        out_channels,
+        n_heads,
+        p_dropout=0.0,
+        window_size=None,
+        heads_share=True,
+        block_length=None,
+        proximal_bias=False,
+        proximal_init=False,
+    ):
+        super().__init__()
+        assert channels % n_heads == 0
+
+        self.channels = channels
+        self.out_channels = out_channels
+        self.n_heads = n_heads
+        self.p_dropout = p_dropout
+        self.window_size = window_size
+        self.heads_share = heads_share
+        self.block_length = block_length
+        self.proximal_bias = proximal_bias
+        self.proximal_init = proximal_init
+        self.attn = None
+
+        self.k_channels = channels // n_heads
+        self.conv_q = nn.Conv1d(channels, channels, 1)
+        self.conv_k = nn.Conv1d(channels, channels, 1)
+        self.conv_v = nn.Conv1d(channels, channels, 1)
+        self.conv_o = nn.Conv1d(channels, out_channels, 1)
+        self.drop = nn.Dropout(p_dropout)
+
+        if window_size is not None:
+            n_heads_rel = 1 if heads_share else n_heads
+            rel_stddev = self.k_channels**-0.5
+            self.emb_rel_k = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+            self.emb_rel_v = nn.Parameter(
+                torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+                * rel_stddev
+            )
+
+        nn.init.xavier_uniform_(self.conv_q.weight)
+        nn.init.xavier_uniform_(self.conv_k.weight)
+        nn.init.xavier_uniform_(self.conv_v.weight)
+        if proximal_init:
+            with torch.no_grad():
+                self.conv_k.weight.copy_(self.conv_q.weight)
+                self.conv_k.bias.copy_(self.conv_q.bias)
+
+    def forward(self, x, c, attn_mask=None):
+        q = self.conv_q(x)
+        k = self.conv_k(c)
+        v = self.conv_v(c)
+
+        x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+        x = self.conv_o(x)
+        return x
+
+    def attention(self, query, key, value, mask=None):
+        b, d, t_s, t_t = (*key.size(), query.size(2))
+        query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+        key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+        value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+        scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+        if self.window_size is not None:
+            assert (
+                t_s == t_t
+            ), "Relative attention is only available for self-attention."
+            key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+            rel_logits = self._matmul_with_relative_keys(
+                query / math.sqrt(self.k_channels), key_relative_embeddings
+            )
+            scores_local = self._relative_position_to_absolute_position(rel_logits)
+            scores = scores + scores_local
+        if self.proximal_bias:
+            assert t_s == t_t, "Proximal bias is only available for self-attention."
+            scores = scores + self._attention_bias_proximal(t_s).to(
+                device=scores.device, dtype=scores.dtype
+            )
+        if mask is not None:
+            scores = scores.masked_fill(mask == 0, -1e4)
+            if self.block_length is not None:
+                assert (
+                    t_s == t_t
+                ), "Local attention is only available for self-attention."
+                block_mask = (
+                    torch.ones_like(scores)
+                    .triu(-self.block_length)
+                    .tril(self.block_length)
+                )
+                scores = scores.masked_fill(block_mask == 0, -1e4)
+        p_attn = F.softmax(scores, dim=-1)
+        p_attn = self.drop(p_attn)
+        output = torch.matmul(p_attn, value)
+        if self.window_size is not None:
+            relative_weights = self._absolute_position_to_relative_position(p_attn)
+            value_relative_embeddings = self._get_relative_embeddings(
+                self.emb_rel_v, t_s
+            )
+            output = output + self._matmul_with_relative_values(
+                relative_weights, value_relative_embeddings
+            )
+        output = (
+            output.transpose(2, 3).contiguous().view(b, d, t_t)
+        )
+        return output, p_attn
+
+    def _matmul_with_relative_values(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0))
+        return ret
+
+    def _matmul_with_relative_keys(self, x, y):
+        ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+        return ret
+
+    def _get_relative_embeddings(self, relative_embeddings, length):
+        2 * self.window_size + 1
+        pad_length = max(length - (self.window_size + 1), 0)
+        slice_start_position = max((self.window_size + 1) - length, 0)
+        slice_end_position = slice_start_position + 2 * length - 1
+        if pad_length > 0:
+            padded_relative_embeddings = F.pad(
+                relative_embeddings,
+                commons.flatten_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
+            )
+        else:
+            padded_relative_embeddings = relative_embeddings
+        used_relative_embeddings = padded_relative_embeddings[
+            :, slice_start_position:slice_end_position
+        ]
+        return used_relative_embeddings
+
+    def _relative_position_to_absolute_position(self, x):
+        batch, heads, length, _ = x.size()
+        x = F.pad(x, commons.flatten_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
+
+        x_flat = x.view([batch, heads, length * 2 * length])
+        x_flat = F.pad(
+            x_flat, commons.flatten_pad_shape([[0, 0], [0, 0], [0, length - 1]])
+        )
+
+        x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+            :, :, :length, length - 1 :
+        ]
+        return x_final
+
+    def _absolute_position_to_relative_position(self, x):
+        batch, heads, length, _ = x.size()
+        x = F.pad(
+            x, commons.flatten_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
+        )
+        x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
+        x_flat = F.pad(x_flat, commons.flatten_pad_shape([[0, 0], [0, 0], [length, 0]]))
+        x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+        return x_final
+
+    def _attention_bias_proximal(self, length):
+        r = torch.arange(length, dtype=torch.float32)
+        diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+        return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        filter_channels,
+        kernel_size,
+        p_dropout=0.0,
+        activation=None,
+        causal=False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.activation = activation
+        self.causal = causal
+
+        if causal:
+            self.padding = self._causal_padding
+        else:
+            self.padding = self._same_padding
+
+        self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+        self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+        self.drop = nn.Dropout(p_dropout)
+
+    def forward(self, x, x_mask):
+        x = self.conv_1(self.padding(x * x_mask))
+        if self.activation == "gelu":
+            x = x * torch.sigmoid(1.702 * x)
+        else:
+            x = torch.relu(x)
+        x = self.drop(x)
+        x = self.conv_2(self.padding(x * x_mask))
+        return x * x_mask
+
+    def _causal_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = self.kernel_size - 1
+        pad_r = 0
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.flatten_pad_shape(padding))
+        return x
+
+    def _same_padding(self, x):
+        if self.kernel_size == 1:
+            return x
+        pad_l = (self.kernel_size - 1) // 2
+        pad_r = self.kernel_size // 2
+        padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+        x = F.pad(x, commons.flatten_pad_shape(padding))
+        return x

third_party/tiny_tts_frontend/tiny_tts/nn/commons.py

@@ -0,0 +1,151 @@
+import math
+import torch
+from torch.nn import functional as F
+
+
+def initialize_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def compute_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def flatten_pad_shape(pad_shape):
+    layer = pad_shape[::-1]
+    pad_shape = [item for sublist in layer for item in sublist]
+    return pad_shape
+
+
+def insert_blanks(lst, item):
+    result = [item] * (len(lst) * 2 + 1)
+    result[1::2] = lst
+    return result
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+    kl = (logs_q - logs_p) - 0.5
+    kl += (
+        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
+    )
+    return kl
+
+
+def rand_gumbel(shape):
+    uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
+    return -torch.log(-torch.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
+    return g
+
+
+def extract_segments(x, ids_str, segment_size=4):
+    ret = torch.zeros_like(x[:, :, :segment_size])
+    for i in range(x.size(0)):
+        idx_str = max(0, ids_str[i].item())
+        idx_end = idx_str + segment_size
+        available = x.size(2) - idx_str
+        if available >= segment_size:
+            ret[i] = x[i, :, idx_str:idx_end]
+        elif available > 0:
+            ret[i, :, :available] = x[i, :, idx_str:idx_str + available]
+    return ret
+
+
+def random_segments(x, x_lengths=None, segment_size=4):
+    b, d, t = x.size()
+    if x_lengths is None:
+        x_lengths = t
+    ids_str_max = torch.clamp(x_lengths - segment_size + 1, min=0)
+    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+    ret = extract_segments(x, ids_str, segment_size)
+    return ret, ids_str
+
+
+def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
+    position = torch.arange(length, dtype=torch.float)
+    num_timescales = channels // 2
+    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
+        num_timescales - 1
+    )
+    inv_timescales = min_timescale * torch.exp(
+        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
+    )
+    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
+    signal = F.pad(signal, [0, 0, 0, channels % 2])
+    signal = signal.view(1, channels, length)
+    return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return x + signal.to(dtype=x.dtype, device=x.device)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+    b, channels, length = x.size()
+    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
+
+
+def subsequent_mask(length):
+    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
+    return mask
+
+
+@torch.jit.script
+def gated_activation(input_a, input_b, n_channels):
+    n_channels_int = n_channels[0]
+    in_act = input_a + input_b
+    t_act = torch.tanh(in_act[:, :n_channels_int, :])
+    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
+    acts = t_act * s_act
+    return acts
+
+
+def shift_1d(x):
+    x = F.pad(x, flatten_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+    return x
+
+
+def create_length_mask(length, max_length=None):
+    if max_length is None:
+        max_length = length.max()
+    x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+    return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def compute_alignment_path(duration, mask):
+    b, _, t_y, t_x = mask.shape
+    cum_duration = torch.cumsum(duration, -1)
+
+    cum_duration_flat = cum_duration.view(b * t_x)
+    path = create_length_mask(cum_duration_flat, t_y).to(mask.dtype)
+    path = path.view(b, t_x, t_y)
+    path = path - F.pad(path, flatten_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+    path = path.unsqueeze(1).transpose(2, 3) * mask
+    return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+    if isinstance(parameters, torch.Tensor):
+        parameters = [parameters]
+    parameters = list(filter(lambda p: p.grad is not None, parameters))
+    norm_type = float(norm_type)
+    if clip_value is not None:
+        clip_value = float(clip_value)
+
+    total_norm = 0
+    for p in parameters:
+        param_norm = p.grad.data.norm(norm_type)
+        total_norm += param_norm.item() ** norm_type
+        if clip_value is not None:
+            p.grad.data.clamp_(min=-clip_value, max=clip_value)
+    total_norm = total_norm ** (1.0 / norm_type)
+    return total_norm

third_party/tiny_tts_frontend/tiny_tts/nn/modules.py

@@ -0,0 +1,578 @@
+import math
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.nn import Conv1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+
+from . import commons
+from .commons import initialize_weights, compute_padding
+from .transforms import spline_transform
+from .attentions import TransformerBlock
+
+LRELU_SLOPE = 0.1
+
+
+class ChannelNorm(nn.Module):
+    def __init__(self, channels, eps=1e-5):
+        super().__init__()
+        self.channels = channels
+        self.eps = eps
+
+        self.gamma = nn.Parameter(torch.ones(channels))
+        self.beta = nn.Parameter(torch.zeros(channels))
+
+    def forward(self, x):
+        x = x.transpose(1, -1)
+        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+        return x.transpose(1, -1)
+
+
+class ConvReluNorm(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels,
+        out_channels,
+        kernel_size,
+        n_layers,
+        p_dropout,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.hidden_channels = hidden_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+        assert n_layers > 1, "Number of layers should be larger than 0."
+
+        self.conv_layers = nn.ModuleList()
+        self.norm_layers = nn.ModuleList()
+        self.conv_layers.append(
+            nn.Conv1d(
+                in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
+            )
+        )
+        self.norm_layers.append(ChannelNorm(hidden_channels))
+        self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
+        for _ in range(n_layers - 1):
+            self.conv_layers.append(
+                nn.Conv1d(
+                    hidden_channels,
+                    hidden_channels,
+                    kernel_size,
+                    padding=kernel_size // 2,
+                )
+            )
+            self.norm_layers.append(ChannelNorm(hidden_channels))
+        self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask):
+        x_org = x
+        for i in range(self.n_layers):
+            x = self.conv_layers[i](x * x_mask)
+            x = self.norm_layers[i](x)
+            x = self.relu_drop(x)
+        x = x_org + self.proj(x)
+        return x * x_mask
+
+
+class DepthwiseSepConv(nn.Module):
+    """Dilated and Depth-Separable Convolution"""
+
+    def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
+        super().__init__()
+        self.channels = channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.p_dropout = p_dropout
+
+        self.drop = nn.Dropout(p_dropout)
+        self.convs_sep = nn.ModuleList()
+        self.convs_1x1 = nn.ModuleList()
+        self.norms_1 = nn.ModuleList()
+        self.norms_2 = nn.ModuleList()
+        for i in range(n_layers):
+            dilation = kernel_size**i
+            padding = (kernel_size * dilation - dilation) // 2
+            self.convs_sep.append(
+                nn.Conv1d(
+                    channels,
+                    channels,
+                    kernel_size,
+                    groups=channels,
+                    dilation=dilation,
+                    padding=padding,
+                )
+            )
+            self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
+            self.norms_1.append(ChannelNorm(channels))
+            self.norms_2.append(ChannelNorm(channels))
+
+    def forward(self, x, x_mask, g=None):
+        if g is not None:
+            x = x + g
+        for i in range(self.n_layers):
+            y = self.convs_sep[i](x * x_mask)
+            y = self.norms_1[i](y)
+            y = F.gelu(y)
+            y = self.convs_1x1[i](y)
+            y = self.norms_2[i](y)
+            y = F.gelu(y)
+            y = self.drop(y)
+            x = x + y
+        return x * x_mask
+
+
+class WaveNet(torch.nn.Module):
+    def __init__(
+        self,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        gin_channels=0,
+        p_dropout=0,
+    ):
+        super(WaveNet, self).__init__()
+        assert kernel_size % 2 == 1
+        self.hidden_channels = hidden_channels
+        self.kernel_size = (kernel_size,)
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.gin_channels = gin_channels
+        self.p_dropout = p_dropout
+
+        self.in_layers = torch.nn.ModuleList()
+        self.res_skip_layers = torch.nn.ModuleList()
+        self.drop = nn.Dropout(p_dropout)
+
+        if gin_channels != 0:
+            cond_layer = torch.nn.Conv1d(
+                gin_channels, 2 * hidden_channels * n_layers, 1
+            )
+            self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+        for i in range(n_layers):
+            dilation = dilation_rate**i
+            padding = int((kernel_size * dilation - dilation) / 2)
+            in_layer = torch.nn.Conv1d(
+                hidden_channels,
+                2 * hidden_channels,
+                kernel_size,
+                dilation=dilation,
+                padding=padding,
+            )
+            in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+            self.in_layers.append(in_layer)
+
+            if i < n_layers - 1:
+                res_skip_channels = 2 * hidden_channels
+            else:
+                res_skip_channels = hidden_channels
+
+            res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+            res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+            self.res_skip_layers.append(res_skip_layer)
+
+    def forward(self, x, x_mask, g=None, **kwargs):
+        output = torch.zeros_like(x)
+        n_channels_tensor = torch.IntTensor([self.hidden_channels])
+
+        if g is not None:
+            g = self.cond_layer(g)
+
+        for i in range(self.n_layers):
+            x_in = self.in_layers[i](x)
+            if g is not None:
+                cond_offset = i * 2 * self.hidden_channels
+                g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+            else:
+                g_l = torch.zeros_like(x_in)
+
+            acts = commons.gated_activation(x_in, g_l, n_channels_tensor)
+            acts = self.drop(acts)
+
+            res_skip_acts = self.res_skip_layers[i](acts)
+            if i < self.n_layers - 1:
+                res_acts = res_skip_acts[:, : self.hidden_channels, :]
+                x = (x + res_acts) * x_mask
+                output = output + res_skip_acts[:, self.hidden_channels :, :]
+            else:
+                output = output + res_skip_acts
+        return output * x_mask
+
+    def remove_weight_norm(self):
+        if self.gin_channels != 0:
+            torch.nn.utils.remove_weight_norm(self.cond_layer)
+        for l in self.in_layers:
+            torch.nn.utils.remove_weight_norm(l)
+        for l in self.res_skip_layers:
+            torch.nn.utils.remove_weight_norm(l)
+
+
+class ConvResBlock(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ConvResBlock, self).__init__()
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=compute_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=compute_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=compute_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(initialize_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=compute_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=compute_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=compute_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(initialize_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ConvResBlockLight(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ConvResBlockLight, self).__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=compute_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=compute_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(initialize_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class LogTransform(nn.Module):
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
+            logdet = torch.sum(-y, [1, 2])
+            return y, logdet
+        else:
+            x = torch.exp(x) * x_mask
+            return x
+
+
+class FlipTransform(nn.Module):
+    def forward(self, x, *args, reverse=False, **kwargs):
+        x = torch.flip(x, [1])
+        if not reverse:
+            logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+            return x, logdet
+        else:
+            return x
+
+
+class AffineCoupling(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.channels = channels
+        self.m = nn.Parameter(torch.zeros(channels, 1))
+        self.logs = nn.Parameter(torch.zeros(channels, 1))
+
+    def forward(self, x, x_mask, reverse=False, **kwargs):
+        if not reverse:
+            y = self.m + torch.exp(self.logs) * x
+            y = y * x_mask
+            logdet = torch.sum(self.logs * x_mask, [1, 2])
+            return y, logdet
+        else:
+            x = (x - self.m) * torch.exp(-self.logs) * x_mask
+            return x
+
+
+class FlowCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        dilation_rate,
+        n_layers,
+        p_dropout=0,
+        gin_channels=0,
+        mean_only=False,
+    ):
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = WaveNet(
+            hidden_channels,
+            kernel_size,
+            dilation_rate,
+            n_layers,
+            p_dropout=p_dropout,
+            gin_channels=gin_channels,
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x
+
+
+class ConvolutionalFlow(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        filter_channels,
+        kernel_size,
+        n_layers,
+        num_bins=10,
+        tail_bound=5.0,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.filter_channels = filter_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.num_bins = num_bins
+        self.tail_bound = tail_bound
+        self.half_channels = in_channels // 2
+
+        self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
+        self.convs = DepthwiseSepConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
+        self.proj = nn.Conv1d(
+            filter_channels, self.half_channels * (num_bins * 3 - 1), 1
+        )
+        self.proj.weight.data.zero_()
+        self.proj.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0)
+        h = self.convs(h, x_mask, g=g)
+        h = self.proj(h) * x_mask
+
+        b, c, t = x0.shape
+        h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)
+
+        unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
+        unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(
+            self.filter_channels
+        )
+        unnormalized_derivatives = h[..., 2 * self.num_bins :]
+
+        x1, logabsdet = spline_transform(
+            x1,
+            unnormalized_widths,
+            unnormalized_heights,
+            unnormalized_derivatives,
+            inverse=reverse,
+            tails="linear",
+            tail_bound=self.tail_bound,
+        )
+
+        x = torch.cat([x0, x1], 1) * x_mask
+        logdet = torch.sum(logabsdet * x_mask, [1, 2])
+        if not reverse:
+            return x, logdet
+        else:
+            return x
+
+
+class TransformerCouplingLayer(nn.Module):
+    def __init__(
+        self,
+        channels,
+        hidden_channels,
+        kernel_size,
+        n_layers,
+        n_heads,
+        p_dropout=0,
+        filter_channels=0,
+        mean_only=False,
+        wn_sharing_parameter=None,
+        gin_channels=0,
+    ):
+        assert n_layers == 3, n_layers
+        assert channels % 2 == 0, "channels should be divisible by 2"
+        super().__init__()
+        self.channels = channels
+        self.hidden_channels = hidden_channels
+        self.kernel_size = kernel_size
+        self.n_layers = n_layers
+        self.half_channels = channels // 2
+        self.mean_only = mean_only
+
+        self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+        self.enc = (
+            TransformerBlock(
+                hidden_channels,
+                filter_channels,
+                n_heads,
+                n_layers,
+                kernel_size,
+                p_dropout,
+                isflow=True,
+                gin_channels=gin_channels,
+            )
+            if wn_sharing_parameter is None
+            else wn_sharing_parameter
+        )
+        self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+        self.post.weight.data.zero_()
+        self.post.bias.data.zero_()
+
+    def forward(self, x, x_mask, g=None, reverse=False):
+        x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+        h = self.pre(x0) * x_mask
+        h = self.enc(h, x_mask, g=g)
+        stats = self.post(h) * x_mask
+        if not self.mean_only:
+            m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+        else:
+            m = stats
+            logs = torch.zeros_like(m)
+
+        if not reverse:
+            x1 = m + x1 * torch.exp(logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            logdet = torch.sum(logs, [1, 2])
+            return x, logdet
+        else:
+            x1 = (x1 - m) * torch.exp(-logs) * x_mask
+            x = torch.cat([x0, x1], 1)
+            return x

third_party/tiny_tts_frontend/tiny_tts/nn/transforms.py

@@ -0,0 +1,209 @@
+import torch
+from torch.nn import functional as F
+
+import numpy as np
+
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def spline_transform(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails=None,
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if tails is None:
+        spline_fn = quadratic_spline
+        spline_kwargs = {}
+    else:
+        spline_fn = unbounded_spline
+        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+    outputs, logabsdet = spline_fn(
+        inputs=inputs,
+        unnormalized_widths=unnormalized_widths,
+        unnormalized_heights=unnormalized_heights,
+        unnormalized_derivatives=unnormalized_derivatives,
+        inverse=inverse,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+        **spline_kwargs
+    )
+    return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+    bin_locations[..., -1] += eps
+    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unbounded_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    tails="linear",
+    tail_bound=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+    outside_interval_mask = ~inside_interval_mask
+
+    outputs = torch.zeros_like(inputs)
+    logabsdet = torch.zeros_like(inputs)
+
+    if tails == "linear":
+        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+        constant = np.log(np.exp(1 - min_derivative) - 1)
+        unnormalized_derivatives[..., 0] = constant
+        unnormalized_derivatives[..., -1] = constant
+
+        outputs[outside_interval_mask] = inputs[outside_interval_mask]
+        logabsdet[outside_interval_mask] = 0
+    else:
+        raise RuntimeError("{} tails are not implemented.".format(tails))
+
+    (
+        outputs[inside_interval_mask],
+        logabsdet[inside_interval_mask],
+    ) = quadratic_spline(
+        inputs=inputs[inside_interval_mask],
+        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+        inverse=inverse,
+        left=-tail_bound,
+        right=tail_bound,
+        bottom=-tail_bound,
+        top=tail_bound,
+        min_bin_width=min_bin_width,
+        min_bin_height=min_bin_height,
+        min_derivative=min_derivative,
+    )
+
+    return outputs, logabsdet
+
+
+def quadratic_spline(
+    inputs,
+    unnormalized_widths,
+    unnormalized_heights,
+    unnormalized_derivatives,
+    inverse=False,
+    left=0.0,
+    right=1.0,
+    bottom=0.0,
+    top=1.0,
+    min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+    min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+    if torch.min(inputs) < left or torch.max(inputs) > right:
+        raise ValueError("Input to a transform is not within its domain")
+
+    num_bins = unnormalized_widths.shape[-1]
+
+    if min_bin_width * num_bins > 1.0:
+        raise ValueError("Minimal bin width too large for the number of bins")
+    if min_bin_height * num_bins > 1.0:
+        raise ValueError("Minimal bin height too large for the number of bins")
+
+    widths = F.softmax(unnormalized_widths, dim=-1)
+    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+    cumwidths = torch.cumsum(widths, dim=-1)
+    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+    cumwidths = (right - left) * cumwidths + left
+    cumwidths[..., 0] = left
+    cumwidths[..., -1] = right
+    widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+    derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+    heights = F.softmax(unnormalized_heights, dim=-1)
+    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+    cumheights = torch.cumsum(heights, dim=-1)
+    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+    cumheights = (top - bottom) * cumheights + bottom
+    cumheights[..., 0] = bottom
+    cumheights[..., -1] = top
+    heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+    if inverse:
+        bin_idx = searchsorted(cumheights, inputs)[..., None]
+    else:
+        bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+    delta = heights / widths
+    input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+    input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+    if inverse:
+        a = (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        ) + input_heights * (input_delta - input_derivatives)
+        b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+            input_derivatives + input_derivatives_plus_one - 2 * input_delta
+        )
+        c = -input_delta * (inputs - input_cumheights)
+
+        discriminant = b.pow(2) - 4 * a * c
+        assert (discriminant >= 0).all()
+
+        root = (2 * c) / (-b - torch.sqrt(discriminant))
+        outputs = root * input_bin_widths + input_cumwidths
+
+        theta_one_minus_theta = root * (1 - root)
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * root.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - root).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, -logabsdet
+    else:
+        theta = (inputs - input_cumwidths) / input_bin_widths
+        theta_one_minus_theta = theta * (1 - theta)
+
+        numerator = input_heights * (
+            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+        )
+        denominator = input_delta + (
+            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+            * theta_one_minus_theta
+        )
+        outputs = input_cumheights + numerator / denominator
+
+        derivative_numerator = input_delta.pow(2) * (
+            input_derivatives_plus_one * theta.pow(2)
+            + 2 * input_delta * theta_one_minus_theta
+            + input_derivatives * (1 - theta).pow(2)
+        )
+        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+        return outputs, logabsdet

third_party/tiny_tts_frontend/tiny_tts/text/__init__.py

@@ -0,0 +1,19 @@
+from .symbols import *
+
+
+_symbol_to_id = {s: i for i, s in enumerate(symbols)}
+
+
+def phonemes_to_ids(cleaned_text, tones, language, symbol_to_id=None):
+    """Converts a list of phoneme symbols to a sequence of integer IDs."""
+    symbol_to_id_map = symbol_to_id if symbol_to_id else _symbol_to_id
+    unk_id = symbol_to_id_map.get("UNK")
+    if unk_id is None:
+        phones = [symbol_to_id_map[symbol] for symbol in cleaned_text]
+    else:
+        phones = [symbol_to_id_map.get(symbol, unk_id) for symbol in cleaned_text]
+    tone_start = language_tone_start_map[language]
+    tones = [i + tone_start for i in tones]
+    lang_id = language_id_map[language]
+    lang_ids = [lang_id for _ in phones]
+    return phones, tones, lang_ids

third_party/tiny_tts_frontend/tiny_tts/text/english.py

@@ -0,0 +1,173 @@
+import pickle
+import os
+import re
+from g2p_en import G2p
+
+from . import symbols
+
+from .english_utils.abbreviations import expand_abbreviations
+from .english_utils.time_norm import expand_time_english
+from .english_utils.number_norm import normalize_numbers
+
+
+def distribute_phone(n_phone, n_word):
+    phones_per_word = [0] * n_word
+    for task in range(n_phone):
+        min_tasks = min(phones_per_word)
+        min_indices = [
+            i for i, x in enumerate(phones_per_word) if x == min_tasks
+        ]
+        chosen_index = min_indices[len(min_indices) // 2]
+        phones_per_word[chosen_index] += 1
+    return phones_per_word
+
+
+from transformers import AutoTokenizer
+
+current_file_path = os.path.dirname(__file__)
+CMU_DICT_PATH = os.path.join(current_file_path, "cmudict.rep")
+CACHE_PATH = os.path.join(current_file_path, "cmudict_cache.pickle")
+_g2p = G2p()
+
+arpa = {
+    "AH0", "S", "AH1", "EY2", "AE2", "EH0", "OW2", "UH0", "NG", "B",
+    "G", "AY0", "M", "AA0", "F", "AO0", "ER2", "UH1", "IY1", "AH2",
+    "DH", "IY0", "EY1", "IH0", "K", "N", "W", "IY2", "T", "AA1",
+    "ER1", "EH2", "OY0", "UH2", "UW1", "Z", "AW2", "AW1", "V", "UW2",
+    "AA2", "ER", "AW0", "UW0", "R", "OW1", "EH1", "ZH", "AE0", "IH2",
+    "IH", "Y", "JH", "P", "AY1", "EY0", "OY2", "TH", "HH", "D",
+    "ER0", "CH", "AO1", "AE1", "AO2", "OY1", "AY2", "IH1", "OW0", "L", "SH",
+}
+
+
+def map_phoneme(ph):
+    rep_map = {
+        "：": ",", "；": ",", "，": ",", "。": ".", "！": "!",
+        "？": "?", "\n": ".", "·": ",", "、": ",", "...": "…", "v": "V",
+    }
+    if ph in rep_map.keys():
+        ph = rep_map[ph]
+    if ph in symbols:
+        return ph
+    if ph not in symbols:
+        ph = "UNK"
+    return ph
+
+
+def read_dict():
+    g2p_dict = {}
+    start_line = 49
+    with open(CMU_DICT_PATH) as f:
+        line = f.readline()
+        line_index = 1
+        while line:
+            if line_index >= start_line:
+                line = line.strip()
+                word_split = line.split("  ")
+                word = word_split[0]
+
+                syllable_split = word_split[1].split(" - ")
+                g2p_dict[word] = []
+                for syllable in syllable_split:
+                    phone_split = syllable.split(" ")
+                    g2p_dict[word].append(phone_split)
+
+            line_index = line_index + 1
+            line = f.readline()
+
+    return g2p_dict
+
+
+def cache_dict(g2p_dict, file_path):
+    with open(file_path, "wb") as pickle_file:
+        pickle.dump(g2p_dict, pickle_file)
+
+
+def get_dict():
+    if os.path.exists(CACHE_PATH):
+        with open(CACHE_PATH, "rb") as pickle_file:
+            g2p_dict = pickle.load(pickle_file)
+    else:
+        g2p_dict = read_dict()
+        cache_dict(g2p_dict, CACHE_PATH)
+
+    return g2p_dict
+
+
+eng_dict = get_dict()
+
+
+def parse_phoneme(phn):
+    tone = 0
+    if re.search(r"\d$", phn):
+        tone = int(phn[-1]) + 1
+        phn = phn[:-1]
+    return phn.lower(), tone
+
+
+def parse_syllables(syllables):
+    tones = []
+    phonemes = []
+    for phn_list in syllables:
+        for i in range(len(phn_list)):
+            phn = phn_list[i]
+            phn, tone = parse_phoneme(phn)
+            phonemes.append(phn)
+            tones.append(tone)
+    return phonemes, tones
+
+
+def normalize_text(text):
+    text = text.lower()
+    text = expand_time_english(text)
+    text = normalize_numbers(text)
+    text = expand_abbreviations(text)
+    return text
+
+
+model_id = 'bert-base-uncased'
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+
+def grapheme_to_phoneme(text, pad_start_end=True, tokenized=None):
+    if tokenized is None:
+        tokenized = tokenizer.tokenize(text)
+    ph_groups = []
+    for t in tokenized:
+        if not t.startswith("#"):
+            ph_groups.append([t])
+        else:
+            ph_groups[-1].append(t.replace("#", ""))
+
+    phones = []
+    tones = []
+    word2ph = []
+    for group in ph_groups:
+        w = "".join(group)
+        phone_len = 0
+        word_len = len(group)
+        if w.upper() in eng_dict:
+            phns, tns = parse_syllables(eng_dict[w.upper()])
+            phones += phns
+            tones += tns
+            phone_len += len(phns)
+        else:
+            phone_list = list(filter(lambda p: p != " ", _g2p(w)))
+            for ph in phone_list:
+                if ph in arpa:
+                    ph, tn = parse_phoneme(ph)
+                    phones.append(ph)
+                    tones.append(tn)
+                else:
+                    phones.append(ph)
+                    tones.append(0)
+                phone_len += 1
+        aaa = distribute_phone(phone_len, word_len)
+        word2ph += aaa
+    phones = [map_phoneme(i) for i in phones]
+
+    if pad_start_end:
+        phones = ["_"] + phones + ["_"]
+        tones = [0] + tones + [0]
+        word2ph = [1] + word2ph + [1]
+    return phones, tones, word2ph

third_party/tiny_tts_frontend/tiny_tts/text/english_utils/abbreviations.py

@@ -0,0 +1,35 @@
+import re
+
+# List of (regular expression, replacement) pairs for abbreviations in english:
+abbreviations_en = [
+    (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+    for x in [
+        ("mrs", "misess"),
+        ("mr", "mister"),
+        ("dr", "doctor"),
+        ("st", "saint"),
+        ("co", "company"),
+        ("jr", "junior"),
+        ("maj", "major"),
+        ("gen", "general"),
+        ("drs", "doctors"),
+        ("rev", "reverend"),
+        ("lt", "lieutenant"),
+        ("hon", "honorable"),
+        ("sgt", "sergeant"),
+        ("capt", "captain"),
+        ("esq", "esquire"),
+        ("ltd", "limited"),
+        ("col", "colonel"),
+        ("ft", "fort"),
+    ]
+]
+
+def expand_abbreviations(text, lang="en"):
+    if lang == "en":
+        _abbreviations = abbreviations_en
+    else:
+        raise NotImplementedError()
+    for regex, replacement in _abbreviations:
+        text = re.sub(regex, replacement, text)
+    return text

third_party/tiny_tts_frontend/tiny_tts/text/english_utils/number_norm.py

@@ -0,0 +1,97 @@
+""" from https://github.com/keithito/tacotron """
+
+import re
+from typing import Dict
+
+import inflect
+
+_inflect = inflect.engine()
+_comma_number_re = re.compile(r"([0-9][0-9\,]+[0-9])")
+_decimal_number_re = re.compile(r"([0-9]+\.[0-9]+)")
+_currency_re = re.compile(r"(£|\$|¥)([0-9\,\.]*[0-9]+)")
+_ordinal_re = re.compile(r"[0-9]+(st|nd|rd|th)")
+_number_re = re.compile(r"-?[0-9]+")
+
+
+def _remove_commas(m):
+    return m.group(1).replace(",", "")
+
+
+def _expand_decimal_point(m):
+    return m.group(1).replace(".", " point ")
+
+
+def __expand_currency(value: str, inflection: Dict[float, str]) -> str:
+    parts = value.replace(",", "").split(".")
+    if len(parts) > 2:
+        return f"{value} {inflection[2]}"  # Unexpected format
+    text = []
+    integer = int(parts[0]) if parts[0] else 0
+    if integer > 0:
+        integer_unit = inflection.get(integer, inflection[2])
+        text.append(f"{integer} {integer_unit}")
+    fraction = int(parts[1]) if len(parts) > 1 and parts[1] else 0
+    if fraction > 0:
+        fraction_unit = inflection.get(fraction / 100, inflection[0.02])
+        text.append(f"{fraction} {fraction_unit}")
+    if len(text) == 0:
+        return f"zero {inflection[2]}"
+    return " ".join(text)
+
+
+def _expand_currency(m: "re.Match") -> str:
+    currencies = {
+        "$": {
+            0.01: "cent",
+            0.02: "cents",
+            1: "dollar",
+            2: "dollars",
+        },
+        "€": {
+            0.01: "cent",
+            0.02: "cents",
+            1: "euro",
+            2: "euros",
+        },
+        "£": {
+            0.01: "penny",
+            0.02: "pence",
+            1: "pound sterling",
+            2: "pounds sterling",
+        },
+        "¥": {
+            # TODO rin
+            0.02: "sen",
+            2: "yen",
+        },
+    }
+    unit = m.group(1)
+    currency = currencies[unit]
+    value = m.group(2)
+    return __expand_currency(value, currency)
+
+
+def _expand_ordinal(m):
+    return _inflect.number_to_words(m.group(0))
+
+
+def _expand_number(m):
+    num = int(m.group(0))
+    if 1000 < num < 3000:
+        if num == 2000:
+            return "two thousand"
+        if 2000 < num < 2010:
+            return "two thousand " + _inflect.number_to_words(num % 100)
+        if num % 100 == 0:
+            return _inflect.number_to_words(num // 100) + " hundred"
+        return _inflect.number_to_words(num, andword="", zero="oh", group=2).replace(", ", " ")
+    return _inflect.number_to_words(num, andword="")
+
+
+def normalize_numbers(text):
+    text = re.sub(_comma_number_re, _remove_commas, text)
+    text = re.sub(_currency_re, _expand_currency, text)
+    text = re.sub(_decimal_number_re, _expand_decimal_point, text)
+    text = re.sub(_ordinal_re, _expand_ordinal, text)
+    text = re.sub(_number_re, _expand_number, text)
+    return text

third_party/tiny_tts_frontend/tiny_tts/text/english_utils/time_norm.py

@@ -0,0 +1,47 @@
+import re
+
+import inflect
+
+_inflect = inflect.engine()
+
+_time_re = re.compile(
+    r"""\b
+                          ((0?[0-9])|(1[0-1])|(1[2-9])|(2[0-3]))  # hours
+                          :
+                          ([0-5][0-9])                            # minutes
+                          \s*(a\\.m\\.|am|pm|p\\.m\\.|a\\.m|p\\.m)? # am/pm
+                          \b""",
+    re.IGNORECASE | re.X,
+)
+
+
+def _expand_num(n: int) -> str:
+    return _inflect.number_to_words(n)
+
+
+def _expand_time_english(match: "re.Match") -> str:
+    hour = int(match.group(1))
+    past_noon = hour >= 12
+    time = []
+    if hour > 12:
+        hour -= 12
+    elif hour == 0:
+        hour = 12
+        past_noon = True
+    time.append(_expand_num(hour))
+
+    minute = int(match.group(6))
+    if minute > 0:
+        if minute < 10:
+            time.append("oh")
+        time.append(_expand_num(minute))
+    am_pm = match.group(7)
+    if am_pm is None:
+        time.append("p m" if past_noon else "a m")
+    else:
+        time.extend(list(am_pm.replace(".", "")))
+    return " ".join(time)
+
+
+def expand_time_english(text: str) -> str:
+    return re.sub(_time_re, _expand_time_english, text)

third_party/tiny_tts_frontend/tiny_tts/text/symbols.py

@@ -0,0 +1,293 @@
+# punctuation = ["!", "?", "…", ",", ".", "'", "-"]
+punctuation = ["!", "?", "…", ",", ".", "'", "-", "¿", "¡"]
+pu_symbols = punctuation + ["SP", "UNK"]
+pad = "_"
+
+# chinese
+zh_symbols = [
+    "E",
+    "En",
+    "a",
+    "ai",
+    "an",
+    "ang",
+    "ao",
+    "b",
+    "c",
+    "ch",
+    "d",
+    "e",
+    "ei",
+    "en",
+    "eng",
+    "er",
+    "f",
+    "g",
+    "h",
+    "i",
+    "i0",
+    "ia",
+    "ian",
+    "iang",
+    "iao",
+    "ie",
+    "in",
+    "ing",
+    "iong",
+    "ir",
+    "iu",
+    "j",
+    "k",
+    "l",
+    "m",
+    "n",
+    "o",
+    "ong",
+    "ou",
+    "p",
+    "q",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "u",
+    "ua",
+    "uai",
+    "uan",
+    "uang",
+    "ui",
+    "un",
+    "uo",
+    "v",
+    "van",
+    "ve",
+    "vn",
+    "w",
+    "x",
+    "y",
+    "z",
+    "zh",
+    "AA",
+    "EE",
+    "OO",
+]
+num_zh_tones = 6
+
+# japanese
+ja_symbols = [
+    "N",
+    "a",
+    "a:",
+    "b",
+    "by",
+    "ch",
+    "d",
+    "dy",
+    "e",
+    "e:",
+    "f",
+    "g",
+    "gy",
+    "h",
+    "hy",
+    "i",
+    "i:",
+    "j",
+    "k",
+    "ky",
+    "m",
+    "my",
+    "n",
+    "ny",
+    "o",
+    "o:",
+    "p",
+    "py",
+    "q",
+    "r",
+    "ry",
+    "s",
+    "sh",
+    "t",
+    "ts",
+    "ty",
+    "u",
+    "u:",
+    "w",
+    "y",
+    "z",
+    "zy",
+]
+num_ja_tones = 1
+
+# English
+en_symbols = [
+    "aa",
+    "ae",
+    "ah",
+    "ao",
+    "aw",
+    "ay",
+    "b",
+    "ch",
+    "d",
+    "dh",
+    "eh",
+    "er",
+    "ey",
+    "f",
+    "g",
+    "hh",
+    "ih",
+    "iy",
+    "jh",
+    "k",
+    "l",
+    "m",
+    "n",
+    "ng",
+    "ow",
+    "oy",
+    "p",
+    "r",
+    "s",
+    "sh",
+    "t",
+    "th",
+    "uh",
+    "uw",
+    "V",
+    "w",
+    "y",
+    "z",
+    "zh",
+]
+num_en_tones = 4
+
+# Korean
+kr_symbols = ['ᄌ', 'ᅥ', 'ᆫ', 'ᅦ', 'ᄋ', 'ᅵ', 'ᄅ', 'ᅴ', 'ᄀ', 'ᅡ', 'ᄎ', 'ᅪ', 'ᄑ', 'ᅩ', 'ᄐ', 'ᄃ', 'ᅢ', 'ᅮ', 'ᆼ', 'ᅳ', 'ᄒ', 'ᄆ', 'ᆯ', 'ᆷ', 'ᄂ', 'ᄇ', 'ᄉ', 'ᆮ', 'ᄁ', 'ᅬ', 'ᅣ', 'ᄄ', 'ᆨ', 'ᄍ', 'ᅧ', 'ᄏ', 'ᆸ', 'ᅭ', '(', 'ᄊ', ')', 'ᅲ', 'ᅨ', 'ᄈ', 'ᅱ', 'ᅯ', 'ᅫ', 'ᅰ', 'ᅤ', '~', '\\', '[', ']', '/', '^', ':', 'ㄸ', '*']
+num_kr_tones = 1
+
+# Spanish
+es_symbols = [
+        "N",
+        "Q",
+        "a",
+        "b",
+        "d",
+        "e",
+        "f",
+        "g",
+        "h",
+        "i",
+        "j",
+        "k",
+        "l",
+        "m",
+        "n",
+        "o",
+        "p",
+        "s",
+        "t",
+        "u",
+        "v",
+        "w",
+        "x",
+        "y",
+        "z",
+        "ɑ",
+        "æ",
+        "ʃ",
+        "ʑ",
+        "ç",
+        "ɯ",
+        "ɪ",
+        "ɔ",
+        "ɛ",
+        "ɹ",
+        "ð",
+        "ə",
+        "ɫ",
+        "ɥ",
+        "ɸ",
+        "ʊ",
+        "ɾ",
+        "ʒ",
+        "θ",
+        "β",
+        "ŋ",
+        "ɦ",
+        "ɡ",
+        "r",
+        "ɲ",
+        "ʝ",
+        "ɣ",
+        "ʎ",
+        "ˈ",
+        "ˌ",
+        "ː"
+    ]
+num_es_tones = 1
+
+# French
+fr_symbols = [
+    "\u0303",
+    "œ",
+    "ø",
+    "ʁ",
+    "ɒ",
+    "ʌ",
+    "ɜ",
+    "ɐ"
+]
+num_fr_tones = 1
+
+# German
+de_symbols = [
+    "ʏ",
+    "̩"
+]
+num_de_tones = 1
+
+# Russian
+ru_symbols = [
+    "ɭ",
+    "ʲ",
+    "ɕ",
+    "\"",
+    "ɵ",
+    "^",
+    "ɬ"
+]
+num_ru_tones = 1
+
+# combine all symbols
+normal_symbols = sorted(set(zh_symbols + ja_symbols + en_symbols + kr_symbols + es_symbols + fr_symbols + de_symbols + ru_symbols))
+symbols = [pad] + normal_symbols + pu_symbols
+sil_phonemes_ids = [symbols.index(i) for i in pu_symbols]
+
+# combine all tones
+num_tones = num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones + num_ru_tones
+
+# language maps
+language_id_map = {"ZH": 0, "JP": 1, "EN": 2, "ZH_MIX_EN": 3, 'KR': 4, 'ES': 5, 'SP': 5, 'FR': 6, 'DE': 7, 'RU': 8, 'VI': 9}
+num_languages = 10
+
+language_tone_start_map = {
+    "ZH": 0,
+    "ZH_MIX_EN": 0,
+    "JP": num_zh_tones,
+    "EN": num_zh_tones + num_ja_tones,
+    'KR': num_zh_tones + num_ja_tones + num_en_tones,
+    "ES": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "SP": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones,
+    "FR": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones,
+    "DE": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones,
+    "RU": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones,
+    "VI": num_zh_tones + num_ja_tones + num_en_tones + num_kr_tones + num_es_tones + num_fr_tones + num_de_tones + num_ru_tones,
+}
+
+if __name__ == "__main__":
+    a = set(zh_symbols)
+    b = set(en_symbols)
+    print(sorted(a & b))

third_party/tiny_tts_frontend/tiny_tts/utils/__init__.py

@@ -0,0 +1,5 @@
+from .config import (
+    SAMPLING_RATE, FILTER_LENGTH, HOP_LENGTH, SEGMENT_FRAMES,
+    ADD_BLANK, SPEC_CHANNELS, N_SPEAKERS, SPK2ID,
+    MODEL_PARAMS, NUM_LANGUAGES, NUM_TONES,
+)

third_party/tiny_tts_frontend/tiny_tts/utils/config.py

@@ -0,0 +1,42 @@
+# Audio
+SAMPLING_RATE = 44100
+FILTER_LENGTH = 2048
+HOP_LENGTH = 512
+SEGMENT_FRAMES = 32
+ADD_BLANK = True
+SPEC_CHANNELS = FILTER_LENGTH // 2 + 1  # 1025
+N_MEL_CHANNELS = 128                      # updated in new checkpoint
+
+# Speakers
+N_SPEAKERS = 1
+SPK2ID = {"MALE": 0}
+
+# Model — matches config.json for G_150000.pth (lighter version)
+MODEL_PARAMS = dict(
+    use_spk_conditioned_encoder=True,
+    use_noise_scaled_mas=True,
+    inter_channels=32,
+    hidden_channels=32,
+    filter_channels=128,
+    n_heads=2,
+    n_layers=3,
+    n_layers_trans_flow=3,
+    kernel_size=3,
+    p_dropout=0.1,
+    resblock="1",
+    resblock_kernel_sizes=[3, 7, 11],
+    resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+    upsample_rates=[8, 8, 2, 2, 2],
+    upsample_initial_channel=64,
+    upsample_kernel_sizes=[16, 16, 8, 2, 2],
+    n_layers_q=3,
+    use_spectral_norm=False,
+    gin_channels=128,
+    use_sdp=True,
+    mas_noise_scale_initial=0.01,
+    noise_scale_delta=2e-06,
+)
+
+# Language / Tone
+NUM_LANGUAGES = 1
+NUM_TONES = 6