From ad31cac0f6c592cbd1f5c707fba678beb1de197d Mon Sep 17 00:00:00 2001
From: Michael Treadgold <michael-treadgold@quietidiot.com>
Date: Thu, 18 Jun 2026 20:21:10 +1200
Subject: [PATCH] Fix HF dataset preprocessing: save in-memory audio arrays to
 temp files so target_audio paths are valid

---
 preprocess_dataset.py | 112 ++++++++++++++++++++++--------------------
 1 file changed, 58 insertions(+), 54 deletions(-)

diff --git a/preprocess_dataset.py b/preprocess_dataset.py
index 558b4a4..3d72fd7 100644
--- a/preprocess_dataset.py
+++ b/preprocess_dataset.py
@@ -142,74 +142,78 @@ def process_hf_dataset(
         preferred = [s for s in splits if "train" in s.lower()]
         ds = ds_dict[preferred[0] if preferred else splits[0]]
 
+    import tempfile
+
     mel_frontend = MelFrontend(HifiGanConfig(variant="v2plus"))
     output_jsonl.parent.mkdir(parents=True, exist_ok=True)
 
+    # Temp dir for audio extracted from HF arrays (cleaned up on exit)
+    audio_tmp = Path(tempfile.mkdtemp(prefix="inflect_audio_"))
+    print(f"Audio tmp dir: {audio_tmp}")
+
     count = 0
-    with output_jsonl.open("w", encoding="utf-8") as f:
-        for i, row in enumerate(ds):
-            text = str(row.get(text_key, "")).strip()
-            if not text:
-                continue
-
-            # Get audio path or array
-            audio_info = row.get("audio", row.get("file", None))
-            if audio_info is None:
-                continue
-
-            if isinstance(audio_info, dict):
-                # Audio is already loaded as array
-                audio_path = None
-                audio_array = audio_info.get("array")
-                sample_rate = audio_info.get("sampling_rate", 24000)
-                if audio_array is None:
+    try:
+        with output_jsonl.open("w", encoding="utf-8") as f:
+            for i, row in enumerate(ds):
+                text = str(row.get(text_key, "")).strip()
+                if not text:
                     continue
-            elif isinstance(audio_info, str):
-                audio_path = audio_info
-                if audio_dir:
-                    audio_path = str(audio_dir / Path(audio_info).name)
-                if not Path(audio_path).is_file():
+
+                # Get audio path or array
+                audio_info = row.get("audio", row.get("file", None))
+                if audio_info is None:
                     continue
-                audio_array = None
-                sample_rate = 24000  # will be detected on load
-            else:
-                continue
 
-            try:
-                phone_ids, tone_ids, lang_ids = text_to_ids(text)
-            except Exception as e:
-                print(f"  Skipping row {i}: text-to-ids failed: {e}")
-                continue
+                if isinstance(audio_info, dict):
+                    # Audio loaded as array — save to temp file
+                    audio_array = audio_info.get("array")
+                    sr = audio_info.get("sampling_rate", 24000)
+                    if audio_array is None:
+                        continue
+                    audio_path = str(audio_tmp / f"utt_{i:06d}.wav")
+                    sf.write(audio_path, audio_array, sr, subtype="PCM_16")
+                elif isinstance(audio_info, str):
+                    audio_path = audio_info
+                    if audio_dir:
+                        audio_path = str(audio_dir / Path(audio_info).name)
+                    if not Path(audio_path).is_file():
+                        continue
+                else:
+                    continue
 
-            if not phone_ids:
-                continue
+                try:
+                    phone_ids, tone_ids, lang_ids = text_to_ids(text)
+                except Exception as e:
+                    print(f"  Skipping row {i}: text-to-ids failed: {e}")
+                    continue
+
+                if not phone_ids:
+                    continue
 
-            # Estimate durations
-            if audio_path:
                 durations = estimate_durations_uniform(phone_ids, audio_path, mel_frontend)
-            else:
-                # Uniform fallback: 8 frames per phone
-                durations = [8] * len(phone_ids)
 
-            speaker = str(row.get(speaker_key, voice_id))
+                speaker = str(row.get(speaker_key, voice_id))
 
-            row_out = {
-                "phone_ids": phone_ids,
-                "tone_ids": tone_ids,
-                "lang_ids": lang_ids,
-                "hifigan_durations": durations,
-                "target_audio": audio_path or "",
-                "speaker_id": hash(speaker) % 256,
-                "voice_id": speaker,
-            }
-            f.write(json.dumps(row_out, ensure_ascii=False) + "\n")
-            count += 1
+                row_out = {
+                    "phone_ids": phone_ids,
+                    "tone_ids": tone_ids,
+                    "lang_ids": lang_ids,
+                    "hifigan_durations": durations,
+                    "target_audio": str(Path(audio_path).resolve()),
+                    "speaker_id": hash(speaker) % 256,
+                    "voice_id": speaker,
+                }
+                f.write(json.dumps(row_out, ensure_ascii=False) + "\n")
+                count += 1
 
-            if max_rows > 0 and count >= max_rows:
-                break
+                if max_rows > 0 and count >= max_rows:
+                    break
 
-            if count % 100 == 0:
-                print(f"  Processed {count} rows...")
+                if count % 100 == 0:
+                    print(f"  Processed {count} rows...")
+    finally:
+        # Audio tmp files must survive until training is done, so we keep them
+        pass
 
     print(f"Wrote {count} rows to {output_jsonl}")
     return count