Uses some sample videos from YouTube of varying duration to benchmark pyannote.audio speaker diarization on an NVIDIA GPU.
More YouTube video id's can be added in video_ids.txt.
Have pyannote.audio installed, of course, as well as yt-dlp and ffmpeg.
Also obtain a Hugging Face token for pyannote.audio speaker diarization and put it in the script (line 103).
Then simply run the script, passing in the GPU number you'd like to benchmark (obtain from nvidia-smi).
python3 benchmark.py <GPU_NUMBER>
I've run this script on some GPUs. If you've run it on a GPU not in this list, feel free to submit a PR and we can get it added.
Note: the benchmark I ran for each of these GPUs was with many more videos than the ones in video_ids.txt. I included videos ranging from a few minutes long to over 5 hours long. I have reduced the number of videos in video_ids.txt for this repo so that people can get a quick benchmark result.
| GPU | Diarization Processing Secs / 1 Audio Sec |
|---|---|
| H100 PCIe | 0.01150315 |
| 4090 | 0.014636619 |
| A100 | 0.015494026 |
| 3090 | 0.017570489 |
| A10G | 0.023336284 |
| 1080 Ti | 0.028720307 |
| 3060 Ti | 0.029580563 |
| 980 Ti | 0.038951152 |
| RTX A4000 | 0.039963005 |
| 1070 Ti | 0.040482773 |
| Apple M1 | 0.147083095 |
For Apple M1 --> pipeline.to(torch.device("mps")) (I used a seperate script)