CosyVoice2-0.5B — Tajik (тоҷикӣ) TTS

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

CosyVoice2-0.5B fine-tuned for Tajik text-to-speech.

• Base: FunAudioLLM/CosyVoice2-0.5B
• Fine-tuned component: LLM (text -> speech-token) on a 68k-clip Tajik read-speech dataset (audio + transcripts).
• Quality: ~13.8% round-trip WER on FLEURS-tg (synthesize -> ASR -> compare). Base CosyVoice2 cannot produce Tajik (hallucinates); this fine-tune produces natural, correctly-timed Tajik.
• Flow / HiFiGAN: base CosyVoice2 (render Tajik acoustics well).

Usage

from cosyvoice.cli.cosyvoice import CosyVoice2
cosy = CosyVoice2('THIS_MODEL_DIR', load_jit=False, load_trt=False, fp16=False)
for out in cosy.inference_zero_shot(tajik_text, ref_text, ref_wav_path, stream=False):
    audio = out['tts_speech']  # 24kHz

Zero-shot: provide a reference wav + its transcript to set the voice.