CTC layer missing
Hi, the saved checkpoint doesn't seem to have a CTC layer, is it normal?
Hi, i believe that it shouldn't as ctc isn't used at decoding time.
I wanted to use the model for CTC segmentation which requires the CTC layer :)
(https://speechbrain.readthedocs.io/en/latest/API/speechbrain.alignment.ctc_segmentation.html#speechbrain.alignment.ctc_segmentation.CTCSegmentation)
Indeed, then you will need to start from the gdrive folder instead, it contains all the checkpoints. In HF, we only have cleaned, ready to use models.
Thanks for the reply. I've been trying this link (from the model's README) but it seems broken:
https://drive.google.com/drive/folders/1SAndjcThdkO-YQF8kvwPOXlQ6LMT71vt?usp=sharing
If you just need an English CTC model for segmentation, you could use this one: https://huggingface.co/speechbrain/asr-wav2vec2-commonvoice-en
Thanks, I actually need a fairly low WER because I'm not using the ground truth text during segmentation but directly the decoded text (in a slight departure from the ctc segmentation code in speechbrain). I've tested the asr-wav2vec2-commonvoice-en model already but its WER is too high for my application.
So far I've been using https://huggingface.co/facebook/wav2vec2-base-960h with good success but I wanted to compare it to the speechbrain asr-crdnn-rnnlm-librispeech model as it looks promising.