SummaryMixing icon indicating copy to clipboard operation
SummaryMixing copied to clipboard

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 t...

Results 9 SummaryMixing issues
Sort by recently updated
recently updated
newest added

Hi, This PR fix an issue with the provided yaml. Indeed, the `SpeedPerturb` is missing an input arg in the constructor which is `orig_freq`.

Hi, This PR fix an issue related to SummaryMixing when using it as an encoder only. For more context, please see: #5 :)

Hi, Thanks for this repo. I have been playing a bit with your creation SummaryMixing and tried to plug it in a CTC-only recipe (https://github.com/speechbrain/speechbrain/blob/develop/recipes/LibriSpeech/ASR/CTC/train.py) which has been implemented by...

Hi author, I'm getting the following when training branchformer using summary_mixing [autodl-container-4d6411b93c-8a044365] 2024-04-10 17:11:12,899 (ctc:67) WARNING: 13/34 samples got nan grad. These were ignored for CTC loss. [autodl-container-4d6411b93c-8a044365] 2024-04-10 17:11:13,133...

This PR adds the summarymixing for streaming conformer. I think this PR should be merged before the SSL one #9. Once both PR have been merged, we should create a...

enhancement

This PR provides necessary code and recipes for reproduce the results of the SummaryMixing SSL paper.

Dear Team, I want to compare the ASR results we have reached based on wav2vec2 & whisper architectures, with your SummaryMixing one. We are performing a custom ASR training, our...

1. Added the layernorm flag for SummaryMixing 2. Added the code and yamls for SSL 3. Added the yaml for SB 1.0 Conformer 4. Updated readme

This PR includes the yaml and the layernom flag to produce the Conformer results with SpeechBrain 1.0