You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is granted manually. Please describe your intended use.
Log in or Sign Up to review the conditions and access this model content.
CosyVoice3 Central-Asian (Tajik + Uzbek + Kazakh)
A single zero-shot text-to-speech / voice-cloning model covering three Central-Asian languages โ Tajik (ัะพาทะธะบำฃ), Uzbek (oสปzbekcha), Kazakh (าะฐะทะฐาัะฐ) โ fine-tuned from FunAudioLLM CosyVoice3-0.5B on a combined ~1.78M-utterance corpus (Tajik 642k + Uzbek 645k + Kazakh 500k).
Results โ round-trip WER (FLEURS, synth โ fine-tuned CA ASR โ WER, n=20 per language)
| Language | median WER | collapse |
|---|---|---|
| Tajik | 17.6% | 0/20 |
| Uzbek | 16.7% | 1/20 |
| Kazakh | 12.5% | 0/20 |
One model, all three languages, stable synthesis. (Single-language specialists Tohirju/cosyvoice3-tajik-642k / -uzbek-full645k score a bit lower per language, but this is one combined model.)
Usage โ correct prompt format
CosyVoice3 zero-shot needs the instruct prefix on the reference transcript (in the target language), not the target text:
import sys; sys.path.append('third_party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import CosyVoice3
cosy = CosyVoice3('cosyvoice3-central-asian', load_trt=False, load_vllm=False, fp16=False)
for out in cosy.inference_zero_shot(
TARGET_TEXT, # target text (PLAIN), in tg/uz/kk
'You are a helpful assistant.<|endofprompt|>' + REF_TRANSCRIPT, # instruct + reference transcript (same language)
REF_WAV_PATH, stream=False):
audio = out['tts_speech'] # 24 kHz
The language is inferred from the reference + target text (Tajik/Kazakh are Cyrillic, Uzbek is Latin).
License
Apache-2.0 (inherits from CosyVoice3-0.5B). Access is gated (manual approval).
- Downloads last month
- -