Ghana Fast Voice: base model
A ZipVoice model (123M parameters) fine-tuned on 542 hours of speech in 134 African languages plus Ghanaian English. It is meant to be a starting point for fine-tuning: language-specific or single-voice models, such as ghanaopenai/ghana-fast-voice, are trained from it. It is not packaged for deployment.
- Architecture: ZipVoice (flow matching, Zipformer text encoder and decoder, 24 kHz Vocos mel features), fine-tuned from k2-fsa's pre-trained ZipVoice. Zero-shot voice cloning from a speech prompt is kept.
- Text input: africa-g2p's universal orthography: lowercase
a-zand spaces. Native text in any of the languages is converted to this one orthography withafrig2p, so one character tokenizer (tokens.txt, ZipVoice'ssimpletokenizer) covers every language. There are no language tags or speaker IDs. - Training data: ghanaopenai/african-speech-universal-tts: 258,205 clips (542.1 h) from every AfriSpeech-Selector subset, plus Ghanaian English podcast speech. The text is AfriSpeech/africa-whisper-ct2's universal-orthography transcript of each clip.
- Training: about 6 epochs at a fixed learning rate (1e-4, ScaledAdam, bf16, one H200). It was early-stopped on dev loss, and this checkpoint averages epochs 3 to 5.
| Dev loss (1,406 held-out clips) | |
|---|---|
| k2-fsa ZipVoice, before fine-tuning | 0.0711 |
| after epoch 1 | 0.0632 |
| after epoch 3 | 0.0620 |
| after epoch 5 (best) | 0.0617 |
Files
model.pt: the model weights ({"model": state_dict}, fp32)model.json: model configurationtokens.txt: the character vocabulary
Fine-tuning
Pass this checkpoint as --checkpoint, and tokens.txt as --token-file, to ZipVoice's
train_zipvoice.py --finetune 1 --tokenizer simple. The training text must be in the
universal orthography. Recipes for per-language and single-voice (no-prompt) fine-tuning are
in GhanaOpenAI/ghana-fast-voice.
Zero-shot inference
python -m zipvoice.bin.infer_zipvoice --model-name zipvoice \
--model-dir ghana-fast-voice-base --checkpoint-name model.pt --tokenizer simple \
--prompt-wav prompt.wav --prompt-text "universal-orthography transcript of the prompt" \
--text "universal-orthography text to say" --res-wav-path out.wav
License
CC BY-NC 4.0 (non-commercial), following the training data. ZipVoice's code and its original pre-trained weights are Apache-2.0.