Batch inference from a diverse list of audio sources for better GPU utilization

[MajaSoure]

I'm really surprised that the "transcribe" method doesn't support running with a list of audio files/tensors. I haven't found a ready-made implementation of this yet. I'm really curious whether I simply missed it or if it really doesn't exist. How can I maximize the GPU to process a list of diverse audio files: from short to long, with different and previously unknown languages? Is this possible? Has anyone already tried implementing this algorithm? I see that I can work with segments of a single audio file. But I don't see an effective implementation with a list without hacks.

[MahmoudAshraf97]

you can do this but as a workaround, using the batched transcribe method, stitch all audio files into a single array, and pass clip_timestamps along with it to trim it back to segments, if there is a file longer than 30s, you need to trim it using VAD first