Jiawei Li

I tried this: https://github.com/pyannote/pyannote-audio/issues/1403#issuecomment-1594032763 This temp_hack_cache speeds up a bit, but still slow. I look at the CPU utilization, most of the time 200-300% on 32-threads Xeon(R) Silver 4314 CPU...

demo1.py里的vad配置完就会先用vad把音频分段。分完了再放到设备上转录，可避免oom问题。demo1.py里的model.generate(batch_size_s, batch_size, merge_vad, merge_length)这几个参数可以调整每个batch放到设备内存推理的时长(batch_size_s)和batch_size,我试过不会再oom.但demo2.py无此类示例，我还在研究。