### 现象 使用多模态模式,供应商为阿里百炼,可稳定复现错误模型:qwen-audio-3.1-realtime-plus <img width="547" height="170" alt="Image" src="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/user-attachments/assets/2bca8181-05a7-4f42-8045-4147872a3d28" /> 错误日志中无相关错误: 2026-09-25T05:51:28.255101Z [INFO] [coord] hotkey pressed (mode=Toggle, phase=Listening) 2026-09-25T05:51:28.312854Z [INFO] [recorder] cpal Stream dropped (mic released) 2026-09-25T05:51:28.313819Z [INFO] [coord] multimodal dictation dispatch session_id=ee86b470-cdb5-4d21-a49c-3f8214192764 mode=Structured translation=false voice_agent=false prompt_chars=3776 audio_ms=3786 2026-09-25T05:51:28.314235Z [INFO] [omni] POST https://llm-bwtn5j8mebhd3l4q.cn-beijing.maas.aliyuncs.com/compatible-mode/v1/chat/completions provider=dashscope-omni model=qwen-audio-3.1-realtime-plus audio=true 2026-09-25T05:51:28.789763Z [INFO] [omni] HTTP 200 body={"status_message":"Success.","status_name":"Success"} 2026-09-25T05:51:28.792119Z [INFO] [coord] hotkey released (mode=Toggle, phase=Idle) 语音识别完成结束后ui显示错误: <img width="363" height="49" alt="Image" src="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/user-attachments/assets/8bf9e68c-2cad-469a-9819-b65567989f13" /> ### 影响 多模态识别供应商为阿里百炼,模型:qwen-audio-3.1-realtime-plus,无法正常使用 ### 建议接受标准 多模态识别多模态识别供应商为阿里百炼,模型:qwen-audio-3.1-realtime-plus,可正常使用 ### TODO / 不确定项 _No response_
Activity
appergb commentedon Sep 26, 2026
已在 2.0.0-Beta.3 修复:百炼/DashScope 多模态音频改为携带 WAV data-URL 前缀,仅对解析后的真实主机名(dashscope/aliyuncs)生效,其他 OpenAI 兼容端点保持裸 Base64(提交 6ebf1f1,随 v2.0.0-Beta.3-tauri 发布)。如升级后仍复现请重新打开。