在每次合成的时候，输入不变如何能得到相同的输出//使用Qwen3-TTS-12Hz-1.7B-CustomVoice/）

[keeper0steam-rgb]

Description

Qwen3-TTS-12Hz-1.7B-CustomVoice在输入相同的时候，每次合成的音频都不同，音色or情感都有差异，，，调整了temperature无效，有什么好办法能得到相同输出吗

Reproduction

/

Logs

Environment Information

linux cuda12.8

Known Issue

• The issue hasn't been already addressed in Documentation, Issues, and Discussions.

[keeper0steam-rgb]

对了，使用的vllm

[BeeegZee]

Basically the only stable way I found for us - use the same seed for sample generation.
vllm omni exposes that paramteter, or you could patch the repo yourself or with AI agent.

[hfl112]

Basically the only stable way I found for us - use the same seed for sample generation.
vllm omni exposes that paramteter, or you could patch the repo yourself or with AI agent.

I tried to use the seed, but still got some voice changes, any details for the voice seed setting ?

[BeeegZee]

I mean, you need to define "voice changes" or the "voice stability".

The fixed seed does not guarantee byte identical output, because it is applied only to one part of the model, but the fixed seed voice will "behave" the same - same emotional and vocal range, same artifacts.

The last part means that there will be seeds that produce more varying output, close to or beyond "unrecognizable", but more expressive. And there will more stable seeds, without much artifacts. You just need to find one.

PS: This is based on my experience with proper SFT custom voice, not zero-shot voice cloning.

[hfl112]

Thanks for sharing. I tried to use custom design based on the same reference, for now, sounds better, at least have the similar "voice", but still got some "clone voice difference" for different chunks

[lcedaw]

same problem

[lcedaw]

使用CustomVoice模型生成语音后，使用Base模型克隆语音形成固定音色，这样每次输出的音色都很相同