Can Zipformer's 130ms of padding be reduced further?

[uni-rini-sharon]

The 13 frame padding for convolution and right context in zipformer is in effect adding 130ms of latency to the overall streaming ASR processing. Is there any way we can reduce this ? Has anyone tried training with causality or other methods to bring this down?

We see that the 13 frames constitute 7 frames for convolutional kernel functioning and 6 for zipformer's internal right context.

@csukuangfj @pkufool