Some updates regarding this PR:

I think we can now extend the beamsearcher to all the other recipes using CTC (e.g., whisper could be the first one).

I changed the following recipes: CommonVoice, Aishell, LibriSpeech and Switchboard. I can't change the others recipe due to the argument space_token. On Media for instance, I don't know what could be this value and I prefer to not change anything in case it will harm the results.

One note: Aishell with Beam Search is not leading to any improvement. I tried the TorchAudio/CTCBeamSearch/CTCPrefixBeamsearch with and without kenLM and it doest yield to any gains.

We can add some full-inference recipe tests for the CTC beamsearchers. As mentioned privately, we can take this opportunity to make the inference test fasters by downloading a version of LibriSpeech test-clean with only a few sentences.

Done.

I would retrain wav2vec2 ctc, due to the dimensionality issues that we have with the previous checkpoints.

Done.

I am currently running the recipe tests. I will report shortly if everything went fine.

I am currently running the recipe tests. I will report shortly if everything went fine.

Everything went fine!

I ran again recipe tests and all seem to work. The only tests that fail are the direct recipes in timers and such, SLURM, fluent. This is expected because they use the pattern:

asr_model: !apply:speechbrain.pretrained.EncoderDecoderASR.from_hparams
    source: speechbrain/asr-crdnn-rnnlm-librispeech
    run_opts: {"device":"cuda:0"}

which is calls code not compatible with this version. When we will release the new version, we have to modify the HF repo accordingly and the issues will be fixed.

I think we can finally merge this PR. This is an amazing work @Adel-Moumen l and a fundamental step toward speechbrain 1.0!