michsethowusu/fast-ghana-voice
Domain:
natural language processingRecord type:
softwaremodel
Creator:
mic
Host:
Offline Twi/Ewe/Dagbani text-to-speech on ONNX Runtime (no PyTorch). ~14-20x realtime on CPU.
# Fast Ghana Voice 🗣️🇬🇭
**Offline Twi / Ewe / Dagbani text-to-speech on ONNX Runtime — no PyTorch.**
▶️ **Try the live demo:** Fast Ghana Voice on Hugging Face Spaces (runs the same full ONNX pipeline on CPU).
Lightweight (`onnxruntime + numpy + soundfile`), fast on CPU, and fully offline after the
first model download. Built from a StableTTS
flow-matching model finetuned on the
GhanaNLP Bible dataset,
exported to ONNX (text encoder + duration + flow-matching decoder baked into one acoustic
graph; a 44.1 kHz Vocos vocoder with a real-op iSTFT).
## Install
```bash
pip install fast-ghana-voice
# or from source:
pip install git+github.com
```
## Usage
### Python
```python
from fast_ghana_voice import FastGhanaVoice
tts = FastGhanaVoice() # downloads + caches ONNX models on first run
wav, sr = tts.synthesize("Awurade ne me hwɛfo.", language="twi")
tts.save("out.wav", wav, sr)
```
### Command line
```bash
fast-ghana-voice "Naawuni n-nyɛ n lana." --language dagbani -o out.wav
```
Languages: `twi`, `ewe`, `dagbani` (each is a distinct baked voice).
### Fully offline
Download the model files once (`model-steps-4.onnx`, `vocos.onnx`, `tokens.txt` from the
model repo)
into a folder, then point the library at it — no network needed:
```python
tts = FastGhanaVoice(model_dir="./models")
```
## Knobs
- `length_scale` — `>1` slower, `<1` faster speech.
- `noise_scale` — sampling temperature (default 0.667).
- `steps` — baked ODE-solver steps variant (default 4; more steps = smoother, slightly slower).
## How it works
1. **Tokenizer** (pure Python): text → grapheme characters → ids → blank-interspersed.
2. **Acoustic ONNX** (`model-steps-N.onnx`): tokens + speaker id → mel. The text encoder,
duration predictor, length regulation, and the euler ODE loop are all baked in.
3. **Vocoder ONNX** (`vocos.onnx`): mel → 44.1 kHz waveform (Vocos; the complex iSTFT was
reimplemented with real ops so it runs in ONNX …