This repository implements SummaryMixing, a simpler, faster and much cheaper replacement to self-attention for automatic speech recognition (see: https://arxiv.org/abs/2307.07421). The code is ready to be used with the SpeechBrain toolkit).
This commit contains the initial code for running SummaryMixing recipes as in the paper. It does not contain the LICENCE. If the corresponding files are put in the corresponding folders in speechbrain (the folder structure is preserved) then it should just work as intended.
[x] Test that LibriSpeech runs.
[x] Unit tests
[x] CV
[x] Aishell
[x] Start re-training on librispeech to reproduce results (performance is almost identical, 0.08 worst, certainly due to grad accum)
TParcollet
commented
7 months ago
This commit contains the initial code for running SummaryMixing recipes as in the paper. It does not contain the LICENCE. If the corresponding files are put in the corresponding folders in speechbrain (the folder structure is preserved) then it should just work as intended.
The LICENCE will be added in a subsequent PR.