michael-treadgold/speech-nano
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Simplify README and add final banner image

Browse Source
This commit is contained in:
Owen Song
2026-06-17 22:42:06 +00:00
committed by system
parent 1a5f1ddfb1
commit b0996d2caf
3 changed files with 65 additions and 158 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.gitattributes
+1
View File
@@ -1,3 +1,4 @@
*.pt filter=lfs diff=lfs merge=lfs -text *.pt filter=lfs diff=lfs merge=lfs -text
*.wav filter=lfs diff=lfs merge=lfs -text *.wav filter=lfs diff=lfs merge=lfs -text
assets/inflect-nano-banner.png filter=lfs diff=lfs merge=lfs -text
tiny_tts/text/cmudict_cache.pickle filter=lfs diff=lfs merge=lfs -text tiny_tts/text/cmudict_cache.pickle filter=lfs diff=lfs merge=lfs -text
README.md
+61 -158
View File
@@ -16,26 +16,27 @@ library_name: pytorch
--- ---
<p align="center"> <p align="center">
<img src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/assets/inflect-nano-banner.svg" alt="Inflect-Nano-v1 banner" width="100%"> <img src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/assets/inflect-nano-banner.png" alt="Inflect-Nano banner" width="100%">
</p> </p>
# Inflect-Nano-v1 # Inflect-Nano-v1
**Inflect-Nano-v1 is a 4.63M-parameter English text-to-speech model, including its neural vocoder.** **Inflect-Nano-v1 is a tiny English text-to-speech model with 4.63M total inference parameters, including its vocoder.**
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? It is not trying to beat large TTS models. It is a small, local, complete text-to-waveform stack built to test how far ultra-lightweight speech synthesis can go.
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. ## Highlights
<p align="center"> - **4.63M parameters total**
<b>4.632M total params</b> | <b>3.465M acoustic</b> | <b>1.167M vocoder</b> | <b>24 kHz</b> | <b>English</b> - **Includes the vocoder**
</p> - **24 kHz audio**
- **Single English male voice**
- **Runs locally with PyTorch**
- Built for tiny-model experiments, local assistants, embedded demos, and efficient inference research
## Listen First ## Listen
These examples cover punctuation, short questions, numbers, hard words, longer phrasing, and transition-heavy prompts. | Text | Audio |
| Prompt | Audio |
|---|---| |---|---|
| "Did the timing change?" she answered. "Then why did Logan leave?" | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_01.wav"></audio> | | "Did the timing change?" she answered. "Then why did Logan leave?" | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_01.wav"></audio> |
| Who puts a parking meter next to an ER label? | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_02.wav"></audio> | | Who puts a parking meter next to an ER label? | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_02.wav"></audio> |
@@ -46,41 +47,7 @@ These examples cover punctuation, short questions, numbers, hard words, longer p
| If Logan sounded uneasy, then it happened near Long Beach, and the pause has to carry that. | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_07.wav"></audio> | | If Logan sounded uneasy, then it happened near Long Beach, and the pause has to carry that. | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_07.wav"></audio> |
| The word aluminum should not steal attention from the softer ending after entrepreneur. | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_08.wav"></audio> | | The word aluminum should not steal attention from the softer ending after entrepreneur. | <audio controls preload="none" src="https://huggingface.co/owensong/Inflect-Nano-v1/resolve/main/examples/example_08.wav"></audio> |
## Why This Exists ## Install
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
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
```bash ```bash
git clone https://huggingface.co/owensong/Inflect-Nano-v1 git clone https://huggingface.co/owensong/Inflect-Nano-v1
@@ -88,28 +55,23 @@ cd Inflect-Nano-v1
pip install -r requirements.txt pip install -r requirements.txt
``` ```
Generate audio: ## Generate Speech
```bash ```bash
python inference.py \ python inference.py --text "Wait, are you actually being for real now?" --out sample.wav
--text "Wait, are you actually being for real now? I can't believe it!" \
--out sample.wav
``` ```
CPU: CPU:
```bash ```bash
python inference.py \ python inference.py --device cpu --text "Please say neighborhood clearly." --out sample_cpu.wav
--device cpu \
--text "Please say neighborhood, statistics, and anesthesiologist clearly." \
--out sample_cpu.wav
``` ```
Optional controls: With simple controls:
```bash ```bash
python inference.py \ python inference.py \
--text "No, seriously, did Jordan leave the receipt in Albuquerque?" \ --text "The appointment moved to 1:25." \
--length-scale 1.03 \ --length-scale 1.03 \
--pitch-scale 1.00 \ --pitch-scale 1.00 \
--energy-scale 1.00 \ --energy-scale 1.00 \
@@ -122,138 +84,79 @@ Local Gradio demo:
python app.py python app.py
``` ```
The text frontend may download tokenizer files on first run. ## Model Size
## Model Summary | Part | Parameters |
|---|---:|
| Acoustic model | **3.465M** |
| Vocoder generator | **1.167M** |
| Total inference stack | **4.632M** |
| Component | Parameters | Notes | The model files are:
|---|---:|---|
| 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:
```text ```text
weights/inflect_nano_v1_acoustic.pt weights/inflect_nano_v1_acoustic.pt
weights/inflect_nano_v1_vocoder.pt weights/inflect_nano_v1_vocoder.pt
``` ```
## What Makes It Different
Many small TTS projects depend on a separate larger vocoder. Inflect-Nano-v1 includes the vocoder in the published inference stack, so the full text-to-waveform path stays under 5M parameters.
Pipeline:
```text
text
-> English text frontend
-> compact FastSpeech-style acoustic model
-> 80-bin mel spectrogram
-> small Snake HiFi-GAN-style vocoder
-> 24 kHz waveform
```
## Architecture ## Architecture
Inflect-Nano-v1 is a two-part TTS stack: The acoustic model is a compact non-autoregressive FastSpeech-style network. It predicts duration, pitch, energy, and brightness, then decodes an 80-bin mel spectrogram.
```text The vocoder is a small Snake-activation HiFi-GAN-style generator trained for 24 kHz waveform reconstruction.
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 Main settings:
The acoustic model predicts duration, energy, brightness, and pitch, expands token states into frame states, and decodes 80-bin mel spectrograms. | Setting | Value |
|---|---:|
| Sample rate | 24 kHz |
| Mel bins | 80 |
| Acoustic hidden size | 168 |
| Encoder layers | 5 |
| Decoder layers | 6 |
| Vocoder upsample rates | 8, 8, 2, 2 |
Main config: ## Good For
```json
{
"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:
```text
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:
```json
{
"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 - Tiny local TTS experiments
- Offline assistant prototypes - Offline assistant prototypes
- Embedded or low-resource speech demos
- Efficient inference research - Efficient inference research
- Embedded speech demos
- Browser/WASM-style exploration - Browser/WASM-style exploration
- Baseline for sub-5M TTS work - A baseline for sub-5M TTS work
Poor fits: ## Not Good For
- Production narration - Production narration
- Accessibility-critical output - Accessibility-critical output
- Voice cloning - Voice cloning
- Multilingual TTS - Multilingual speech
- High-fidelity studio speech - High-fidelity audiobook generation
- Long-form audiobook generation - Matching large modern TTS systems
## Limitations ## Limitations
This model is intentionally tiny and has clear quality limits: This is a very small experimental model. It can sound robotic, buzzy, or unstable, especially on difficult unseen text. Long prompts and unusual phrasing are less reliable. The vocoder is also a clear quality bottleneck.
- Unseen text can stumble or sound unstable. Use it as a tiny-model research/demo release, not as a production TTS engine.
- 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 ## License
Apache-2.0. 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`. This repository includes a small third-party English text frontend for tokenization/G2P compatibility. Its license is included as `TINY_TTS_LICENSE`.
assets/inflect-nano-banner.png
LFS
BIN
View File
Binary file not shown.