The speech file generated by cosyvoice.inference_zero_shot does not match the order of the words. The generated audio file is basically wrong.

[cckamiya]

The speech file generated by cosyvoice.inference_zero_shot does not match the order of the words. The generated audio file is basically wrong. I used fastapi to encapsulate the cosyvoice.inference_zero_shot interface, but the order of words in the audio file generated by the model is wrong and words are missing. code： async def inference_zero_shot( tts_text: str = Form(...), instruct_text: str = Form(...), prompt_audio: UploadFile = File(...) ): try: file_content = await prompt_audio.read() check_audio_duration(file_content) with open('temp_prompt.wav', 'wb') as f: f.write(file_content) prompt_speech_16k = load_wav('temp_prompt.wav', 16000)

    for i, j in enumerate(cosyvoice.inference_zero_shot(tts_text, instruct_text, prompt_speech_16k, stream=False)):
        output_path = 'zero_shot_{}.wav'.format(i)
        torchaudio.save(output_path, j['tts_speech'], 24000)
        return FileResponse(output_path)
except Exception as e:
    raise HTTPException(status_code=500, detail=str(e))

[FAFUuser]

24000 chang to cosyvoice.sample_rate, add format="wav"

[aluminumbox]

update code and model, follow readme instruction to see if it works first

[hayeong0]

@cckamiya I encountered a similar issue. When I trained a model initialized with Qwen-2.5 and the training did not converge well, the word order became misaligned and some words were skipped. Once the training proceeded successfully, this failure case was greatly reduced.