Benchmark results for Stream-VAD and AED models?

[LattifaiHQ]

Hi, great work on FireRedVAD! 🎉

The README provides comprehensive benchmark results for the VAD (non-streaming) model on FLEURS-VAD-102, showing impressive SOTA performance (F1: 97.57%, AUC-ROC: 99.60%).

However, I noticed that the benchmark results for the other two models are not included:

1. Stream-VAD — Are there comparable benchmark results on FLEURS-VAD-102 or other datasets? Specifically, how does the streaming model compare to the non-streaming VAD in terms of F1, AUC-ROC, and latency?

2. AED — Since it detects three event types (speech, singing, music), are there benchmark results on audio event detection datasets? What metrics were used to evaluate the multi-class detection performance?

It would be very helpful for users to understand the trade-offs between the three models when choosing which one to use.

Thanks!

[kaituoxu]

Got this! I'll test these two models