background music/noise in the synthesized samples

[taalua]

Hi,

Thank you for this work.
I tried to synthesize Japanese text using CosyVoice2-0.5B model, and after trying many times with various clean prompt_speech audios, there are some quite loud music/noise in the background. Why this is happening, and how to fix?

for i, j in enumerate(cosyvoice.inference_cross_lingual('この新しいシステムのパフォーマンスには感心しています。楽
しい会話が行われているのが聞こえます。', prompt_speech_16k, stream=False, text_frontend=False)):
    torchaudio.save('JA_SP.wav', j['tts_speech'], cosyvoice.sample_rate)

JA_SP_sample.zip

Thank you.

[aluminumbox]

改成<|jp|>この新しいシステ

[taalua]

Thank you for your prompt answer.
May I ask, why this is the case?

[aluminumbox]

Thank you for your prompt answer. May I ask, why this is the case?

during training, for some language like chinese/english/japanese/korean, we add langauge tag at sentence start. for some language like cantonese, we use instruct like '请用粤语讲这句话'. please check out demo page for details