Preproccesing pipelines

I want to start playing with SpeechBrain, I want to start with the pipelines/tasks with the heaviest pre/post preprocessing.
Can you recommend such pipelines?

(my motivation: we will be accelerating the DNN parts of the network, and the next thing is accelerating the pre/post-processing, which will become a bottleneck)

Hi, I am not sure to understand what you are really looking into.

Most of the time, SpeechBrain does not perform any pre-processing as it is done on the fly (features, augmentation etc are done on the fly).

By “preprocessing” I refer to all the computation before neural layers. (E.g., audio decoding, mel-spectograms, etc.)
Examples to post processing can be beam-search/CTC/language model.

The motivation is that neural layers are usually matmuls, and matmuls can be extremely accelerated by ASICs (e.g., TPU).


Then you can go for any recipe that uses beam searching through big LMs such as LibriSpeech transformers.