[BUG] <title> Minicpm-o-2.6 Audio Tends to Produce Gibberish during Speech-to-Speech Conversations in Audio-Roleplay Mode

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

When using the audio_roleplay mode, MiniCPM-o-2.6 tends to produce gibberish during speech-to-speech conversations. This issue does not occur when using the audio_assistant mode. Is there any way to resolve this problem while remaining in audio_roleplay mode?

期望行为 | Expected Behavior

No response

复现方法 | Steps To Reproduce

No response

运行环境 | Environment

- OS: 
- Python:3.10
- Transformers:4.44.2
- PyTorch:2.5
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

input1: Thanks for coming—I've poured myself a drink and could really use someone to listen; can I tell you what's been going on?
(audio_roleplay) output1: I'm afrai  neither of those to ay，sai  Holmes.
(audio assistant) output2: Of course, I'm here to listen. Whatever's on your mind, you can share it with me. It’s important to talk things through sometimes.


input2: I keep picking my nose and got in trouble in class — my teacher made me stop and wash my hands; can you help me figure out how to break this habit and handle it better?
(audio_roleplay) output1: I'm all ears，sai  Holmes.
(audio assistant) output2: It's important to understand why you're picking your nose. It might be a way of dealing with stress or boredom, but it can also spread germs and cause infections. Try taking deep breaths when you feel the urge, and find other ways to relax like stretching or going for a short walk. If this is hard, talking to someone about how you’re feeling could help too.

Start by reproducing the issue in audio_roleplay mode with the provided inputs and compare the output with audio_assistant mode under Python 3.10, Transformers 4.44.2, and PyTorch 2.5. Investigate the audio_roleplay entry point; done means speech-to-speech responses no longer produce gibberish while audio_assistant behavior remains unaffected.