Hello! I would like to discuss what steps are needed to implement Fast Conformer in speechbrain? As I see main differences from vanilla Confomer implemented in speechbrain are:

• Modified ConvolutionFrontEnd with total stride 4 -> 8 + depthwise convolutions instead of base ones
• Other modifications that are needed for inference cause stride changes 40 ms -> 80 ms, like DynChunkTrainConfig for streaming training / inference - this ones are harder to track down.

Could someone please help guiding me with inference part, if my vanilla Conformer has doubled stride (4 -> 8 in mel features or 40 ms -> 80 ms) what should be changed on inference (streaming one)?