Recommended audio input duration for optimal performance?

[thienphucope]

Hi team!

What's the recommended (or best) audio input length for SAM-Audio to achieve optimal performance and avoid memory issues?
Regarding the model's architecture, are there plans to support streaming inference in the near future?

Thanks for your hard work!

[xiejesses]

+1

[rushplayer]

lookslike 30s and 2s overlap

[chevalierNoir]

@thienphucope I'd say 10s is a preferred length as many of our training audios are around 10s. In practice, I also recommend to tune the duration a bit as we haven't fully established its effect on performance.

For streaming inference, we don't have an immediate plan yet. But you can refer to the segment-level autoregressive generation of moviegen paper for a potential streaming method, which can be easily adapted for SAM audio.