It is built around one question: how usable can a complete local TTS stack get when the entire inference model is smaller than many single embeddings tables in modern speech systems?

This is not a SOTA voice model. The point is the size-to-functionality ratio: a complete text-to-waveform TTS pipeline that is small enough for tiny local experiments, offline assistants, embedded prototypes, browser/WASM-style work, and efficient inference research.

4.632M total params | 3.465M acoustic | 1.167M vocoder | 24 kHz | English

Listen First

These examples cover punctuation, short questions, numbers, hard words, longer phrasing, and transition-heavy prompts.

Prompt	Audio
"Did the timing change?" she answered. "Then why did Logan leave?"
Who puts a parking meter next to an ER label?
Please say neighborhood, statistics, and anesthesiologist clearly, without rushing through the middle syllables.
I said 91, not 306, which is a very different number.
The inference path looked natural, but the decoder still needed a smoother transition before Marcus approved the final test.
The appointment moved to 1:25, the invoice was $674.96, and the archive was labeled 1998.
If Logan sounded uneasy, then it happened near Long Beach, and the pause has to carry that.
The word aluminum should not steal attention from the softer ending after entrepreneur.

Why This Exists

Most strong TTS models are large. Many "small" TTS demos also depend on a larger external vocoder, which hides a lot of the real inference cost. Inflect-Nano-v1 counts both sides:

text -> acoustic model -> mel spectrogram -> neural vocoder -> waveform

The model is intentionally constrained:

Capability	Inflect-Nano-v1
Complete text-to-waveform stack	Yes
External vocoder required	No
Total inference parameters	4.632M
Sample rate	24 kHz
Language	English
Voice count	1
Voice cloning	No
Multilingual	No

Size Context

Approximate parameter comparison:

Model / class	Params	Relative to Inflect-Nano-v1
Inflect-Nano-v1	4.63M	1.0x
Kokoro-82M	~82M	~17.7x larger
Chatterbox-sized TTS	~500M	~108x larger
Multi-billion TTS systems	1B+	216x+ larger

The comparison is about model scale, not quality parity. Larger models are expected to sound better.

Quickstart

git clone https://huggingface.co/owensong/Inflect-Nano-v1
cd Inflect-Nano-v1
pip install -r requirements.txt

Generate audio:

python inference.py \
  --text "Wait, are you actually being for real now? I can't believe it!" \
  --out sample.wav

CPU:

python inference.py \
  --device cpu \
  --text "Please say neighborhood, statistics, and anesthesiologist clearly." \
  --out sample_cpu.wav

Optional controls:

python inference.py \
  --text "No, seriously, did Jordan leave the receipt in Albuquerque?" \
  --length-scale 1.03 \
  --pitch-scale 1.00 \
  --energy-scale 1.00 \
  --out sample_controlled.wav

Local Gradio demo:

python app.py

The text frontend may download tokenizer files on first run.

Model Summary

Component	Parameters	Notes
Acoustic model	3,465,125	Compact non-autoregressive FastSpeech-style model
Vocoder generator	1,167,077	Snake V2Mid HiFi-GAN-style generator
Total inference model	4,632,202	Acoustic + vocoder

Model files:

weights/inflect_nano_v1_acoustic.pt
weights/inflect_nano_v1_vocoder.pt

Architecture

Inflect-Nano-v1 is a two-part TTS stack:

Text
  -> English normalization + G2P frontend
  -> Compact FastSpeech-style acoustic model
  -> 80-bin mel spectrogram
  -> Snake V2Mid HiFi-GAN-style vocoder
  -> 24 kHz waveform

Acoustic Model

The acoustic model predicts duration, energy, brightness, and pitch, expands token states into frame states, and decodes 80-bin mel spectrograms.

Main config:

{
  "hidden": 168,
  "encoder_layers": 5,
  "decoder_layers": 6,
  "decoder_ff_mult": 3,
  "kernel_size": 7,
  "speaker_dim": 64,
  "dropout": 0.08,
  "n_mels": 80,
  "sample_rate": 24000,
  "max_frames": 1400,
  "postnet_scale": 0.1,
  "use_frame_pitch": true,
  "abs_frame_bins": 512
}

Approximate acoustic split:

total acoustic: 3.465M
encoder:        1.292M
decoder:        1.211M
postnet:        0.276M
local context:  0.226M
frame GRU:      0.128M
heads/embeds/projections: remainder

Vocoder

The vocoder is a custom Snake-activation HiFi-GAN-style generator. Discriminators are training-only and are not included in inference parameter counts.

Main config:

{
  "variant": "snake_v2mid",
  "sample_rate": 24000,
  "n_fft": 1024,
  "hop_size": 256,
  "win_size": 1024,
  "num_mels": 80,
  "fmax": 12000.0,
  "upsample_rates": [8, 8, 2, 2],
  "upsample_kernel_sizes": [16, 16, 4, 4],
  "upsample_initial_channel": 144,
  "resblock_kernel_sizes": [3, 7, 11],
  "resblock_dilation_sizes": [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
  "activation": "snake"
}

Intended Use

Good fits:

Tiny local TTS experiments
Offline assistant prototypes
Embedded or low-resource speech demos
Efficient inference research
Browser/WASM-style exploration
Baseline for sub-5M TTS work

Poor fits:

Production narration
Accessibility-critical output
Voice cloning
Multilingual TTS
High-fidelity studio speech
Long-form audiobook generation

Limitations

This model is intentionally tiny and has clear quality limits:

Unseen text can stumble or sound unstable.
The voice can sound robotic, buzzy, or artifacted.
Long or unusual prompts are less reliable.
The vocoder is a major quality bottleneck.
It is not a voice cloning model.
It is not multilingual.
It is not suitable for safety-critical or production accessibility use.

Recommended Framing

Use this as:

An experimental 4.63M-parameter English TTS model exploring the quality/size tradeoff for ultra-small local speech synthesis.

Do not present it as SOTA or production-quality.

License

Apache-2.0.

This repository includes a small third-party English text frontend derived from TinyTTS-style code for tokenization/G2P compatibility. Its license is included as TINY_TTS_LICENSE.