nano-twi — Asante Twi Text-to-Speech

Fast, offline Asante Twi TTS: a Matcha-TTS acoustic model + Vocos vocoder, exported to ONNX and served with sherpa-onnx. Runs on CPU at several× realtime — desktop, Android, iOS, and web (WASM).

➡️ Full docs, code, and Python / web / mobile examples: https://github.com/michsethowusu/nano-twi 🔊 Live demo: https://huggingface.co/spaces/ghananlpcommunity/nano-twi

Files (sherpa-onnx/)

File	Role
twi_ep045_steps4.onnx	acoustic model — default, 4 ODE steps, best quality
twi_ep045_steps2.onnx	acoustic model — fast, 2 ODE steps, lower quality
vocos-22khz-univ.onnx	Vocos universal vocoder
tokens.txt	phoneme → id table
espeak-ng-data/	espeak-ng lfn-voice data (phonemization)

Quick start (only needs sherpa-onnx)

pip install -U sherpa-onnx
curl -L -o nano-twi.zip https://github.com/michsethowusu/nano-twi/releases/download/v1.0/nano-twi-sherpa-onnx.zip
unzip nano-twi.zip
curl -O https://raw.githubusercontent.com/michsethowusu/nano-twi/main/examples/python/synthesize.py
python3 synthesize.py --model-dir ./nano-twi \
  --text "Awurade ne me hwɛfoɔ, biribiara renhia me." --out twi.wav
# fast model: add  --acoustic twi_ep045_steps2.onnx

(Or fetch individual files with hf download ghananlpcommunity/nano-twi --include "sherpa-onnx/*".)

Notes

• Input is Asante Twi text in normal orthography (incl. ɛ ɔ); phonemized to IPA via espeak-ng lfn.
• Single voice — not a voice-cloning model.
• Use the bundled espeak-ng-data/ as-is (built with eSpeak NG 1.52.0) and sherpa-onnx ≥ 1.12; swapping in different espeak data can change pronunciation.

How it was built

Finetuned from the English LJSpeech Matcha-TTS checkpoint on the Asante Twi subset of michsethowusu/ghana-speech, phonemized with the espeak-ng lfn (Lingua Franca Nova) voice (which maps Twi orthography to clean IPA).

License

MIT. Matcha-TTS, Vocos, sherpa-onnx, and espeak-ng retain their respective licenses.