CosyVoice2-0.5B — Asante Twi Bible Fine-tune

Fine-tuned version of CosyVoice2-0.5B for Akan (Asante Twi) text-to-speech synthesis.

Trained on the ghananlpcommunity/asante-twi-bible-speech-text dataset: 34.2k utterances, 10 speakers, Bible-domain Asante Twi.

Fine-tuning approach

Based on community guidance from FunAudioLLM/CosyVoice Issue #830:

• LLM module fine-tuned (learns Akan phoneme→speech token mapping)
• Flow module fine-tuned (improves prosody for new language)
• Vocoder, speech tokenizer, and speaker encoder remain frozen

Usage

from huggingface_hub import snapshot_download
from cosyvoice.cli.cosyvoice import CosyVoice2
import torchaudio

model_dir = snapshot_download('prince4332/CosyVoice2-0.5B-Akan-Bible')
model = CosyVoice2(model_dir, load_jit=False, load_trt=False, fp16=True)

# Zero-shot voice cloning
prompt_wav, sr = torchaudio.load('reference_akan.wav')
if sr != 16000:
    prompt_wav = torchaudio.functional.resample(prompt_wav, sr, 16000)

for chunk in model.inference_zero_shot(
    tts_text='Meda wo ase.',
    prompt_text='Akwaaba!',
    prompt_speech_16k=prompt_wav,
    stream=False,
):
    torchaudio.save('output.wav', chunk['tts_speech'], 22050)

Training Details

• Base model: FunAudioLLM/CosyVoice2-0.5B
• Dataset: ghananlpcommunity/asante-twi-bible-speech-text (34.2k utterances)
• Epochs: 30
• Hardware: A100 GPU