CosyVoice3-Tajik (642k)

How to get access: please fill in the short form at saidzoda.com/access. We approve requests after we receive the form.

Zero-shot Tajik text-to-speech and voice cloning, fine-tuned from FunAudioLLM CosyVoice3-0.5B on a 642,000-utterance Tajik speech corpus.

Results — round-trip WER (synthesize FLEURS-tg text → ASR → WER), n=60, zero-shot voice cloning

Usage — IMPORTANT: correct prompt format

CosyVoice3 zero-shot needs the instruct prefix on the reference transcript, NOT the target text:

import sys; sys.path.append('third_party/Matcha-TTS')
from cosyvoice.cli.cosyvoice import CosyVoice3
cosy = CosyVoice3('cosyvoice3-tajik-642k', load_trt=False, load_vllm=False, fp16=False)
for out in cosy.inference_zero_shot(
        'Матни тоҷикӣ барои гуфтан.',                                   # target text (PLAIN)
        'You are a helpful assistant.<|endofprompt|>' + REF_TRANSCRIPT, # instruct + reference transcript
        REF_WAV_PATH, stream=False):
    audio = out['tts_speech']   # 24 kHz waveform

Putting <|endofprompt|> on the target text instead of the reference transcript breaks the model's stopping behavior and causes over-generation. Keep it on the reference transcript.

License

Apache-2.0 (inherits from the CosyVoice3-0.5B base). Access is gated (manual approval).