Huge ! Is this comparable to the SOTA around?

Hi @TParcollet , I think it's good for a system pure-CTC/greedy/without LM.
Hybrid models from espnet got better CER:

I see, not bad, but we use extra pre-training while they don't, correct ?

Yes exact. Fair enough, but their branchformer is quite something according to the results.

Hi @anautsch thanks for the review, the fix is done. And yes that's why I want to reduce pandas, it is only used to generate csv for all the recipes.

Hi @TParcollet @anautsch @Adel-Moumen,

Thank you all for the reviews and tests! The HF link is added, here's a brief summary of the PR:

1. add a CTC recipe
2. fix naming problems
3. fix dynamic batching conflicts for seq2seq & transformer recipes

lgtm.

Tested recipes in --debug mode & the wav2vec2 with ddp.

Side note: we have an internal issue with --debug and eval checkpointing - this comes clear when running this transformer wav2vec2 recipe - here's the relevant log

   asr_brain.evaluate(
  File "speechbrain/core.py", line 1260, in evaluate
    self.on_evaluate_start(max_key=max_key, min_key=min_key)
  File "train_with_wav2vect.py", line 272, in on_evaluate_start
    ckpt = sb.utils.checkpoints.average_checkpoints(
  File "speechbrain/utils/checkpoints.py", line 1174, in average_checkpoints
    return averager(parameter_iterator)
  File "speechbrain/utils/checkpoints.py", line 1080, in average_state_dicts
    raise ValueError("No state dicts to average.")
ValueError: No state dicts to average.