I’m using CosyVoice 3 after an LLM to provide streaming speech output. Since the LLM generates text in a streaming manner, the text sent to CosyVoice 3 also arrives chunk by chunk. However, the speech generated from different text chunks sometimes varies significantly in speaking rate, pitch, and intonation, which results in an inconsistent and unnatural voice interaction experience.
Are there any recommended solutions or best practices to improve the consistency of speech across different chunks? Thank you.
I’m using CosyVoice 3 after an LLM to provide streaming speech output. Since the LLM generates text in a streaming manner, the text sent to CosyVoice 3 also arrives chunk by chunk. However, the speech generated from different text chunks sometimes varies significantly in speaking rate, pitch, and intonation, which results in an inconsistent and unnatural voice interaction experience.
Are there any recommended solutions or best practices to improve the consistency of speech across different chunks? Thank you.