About a sft bug

[CamellIyquitous]

hi, thanks for you excellent work!

i noticed a bug fixing last week in official repo: QwenLM/Qwen3-TTS@022e286

in sft_12hz.py, talker_hidden_states = hidden_states[codec_mask[:, 1:]] -> talker_hidden_states = hidden_states[codec_mask[:, :-1]]

may i ask this fixing is correct or not? i tried both ways but seems there's no different for my both sft results.

[CamellIyquitous]

Hi, I have a question regarding multi-speaker SFT.

I tried training with multi-speaker SFT, but during inference I encountered an issue: even when I explicitly specify a fixed speaker, the synthesized voice does not resemble that speaker. Moreover, the voice timbre varies across different samples during inference, instead of staying consistent.

For inference, I followed the Qwen3-TTS official repo and implemented it based on infer_ft.py.

I’m wondering:

• Is this likely an issue with my inference implementation?
• Or could there be something wrong with the multi-speaker SFT setup/training itself?

Any guidance would be greatly appreciated. Thanks!