Comment by conjecTech
9 months ago
If you are hosting whisper yourself, you can do something slightly more elegant, but with the same effect. You can downsample/pool the context 2:1 (or potentially more) a few layers into the encoder. That allows you to do the equivalent of speeding up audio without worry about potential spectral losses. For whisper large v3, that gets you nearly double throughput in exchange for a relative ~4% WER increase.
Do you have more details or examples on how to downsample the context in the encoder? I treat the encoder as an opaque block, so I have no idea where to start.
It's a very simple change in a vanilla python implementation. The encoder is a set of attention blocks, and the length of the attention can be changed without changing the calculation at all.
Here(https://github.com/openai/whisper/blob/main/whisper/model.py...) is the relevant code in the whisper repo. You'd just need to change the for loop to an enumerate and subsample the context along its length at the point you want. I believe it would be:
for i, block in enumerate(self.blocks): x = block(x) if i==4: x = x[,,::2]