VibeVoice 卡顿

我运行在mac mini 4 和jetson orin agx 64g 在做这个realtime voice 生成上。生成的声音有点重复。然后说话声音一卡一卡的

[vibevoice_realtime_audio_2025-12-08T03-49-

vibevoice_realtime_audio_2025-12-08T03-49-30-420Z.wav

30-420Z.wav](https://github.com/user-attachments/files/24023298/vibevoice_realtime_audio_2025-12-08T03-49-30-420Z.wav)

Dec 08 '25 03:12 zujian-y

The generated WAV file is normal, which suggests that the model output is correct. However, your hardware is likely not powerful enough to support real-time inference. When generation cannot keep up with playback, it sounds like stuttered.

Dec 08 '25 04:12 YaoyaoChang

谢谢哈。明白了看来还是得用好一点的gpu

Dec 08 '25 05:12 zujian-y

但是我用mac studio ultra m3 256g mps 推理也是一样的。 inference step 5 cfg scale 1.5

Dec 08 '25 06:12 zujian-y

是不是得用mlx 去支持一下这个模型

Dec 08 '25 06:12 zujian-y

mpx is already supported. You can try to add --device mpx when launching the demo.

python demo/vibevoice_realtime_demo.py --model_path microsoft/VibeVoice-Realtime-0.5B --device mpx

Dec 08 '25 07:12 YaoyaoChang

你看这个是mps 是pytorch 版本的但是用mlx 是apple 自家的ai加速库这样兼容一下是不是会更快一点呢

Dec 08 '25 07:12 zujian-y

Does it resolve the issue of stuttering? I’m not very familiar with Apple devices. If you have any technical suggestions, feel free to open a PR.

Dec 08 '25 07:12 YaoyaoChang

没有。感觉苹果的电脑运行还是会卡顿。可能支持不够。算力不够？无法流畅的做。我们在4070ti super 和 agx spark 上做都可以做到流畅。但是apple 的m3 studio ultra 上始终会卡顿。

Dec 08 '25 07:12 zujian-y

M4 Pro works well.

Dec 08 '25 07:12 YaoyaoChang