大模型流式回答，然后文本流式输入到MOSS-TTS-Nano，保证音色一致性

[iamyishan]

使用场景：智能体回答时，实时语音播报，就像手机豆包一样，边回答文本，边语音播报。
尝试了：把大模型回答文本切片之后喂入语音合成模型，但是各个片段的音色会出现不一致，而且衔接处会音调会跳跃
如果等大模型回答完整之后，再喂入语音合成模型，这样会有延迟
请问大佬如何解决这个问题？

[alpacaking]

This scenario is worth trying, but please note that cross-chunk speaker consistency mainly depends on whether the prompt audio and the conditioning between chunks are stable.

First, we suggest choosing a clean, clear, and stable prompt audio. You can test the same prompt audio on different text chunks first, and check whether it can generate stable voice quality across chunks without obvious speaker changes.

Another approach you can try is a rolling prompt strategy: when generating chunk by chunk, take a short segment from the end of the previously generated audio and use it as the prompt audio for the next chunk. This lets the next chunk refer more directly to the voice and prosody of the previous chunk, which may help reduce speaker changes and unnatural transitions between chunks.