Difference between ctc_lin and seq_lin

Hello ! :grinning_face_with_smiling_eyes:

I see in the tutoriel sequence for fine-tuning that you used seq_lin in a way and you do not used it when you load the model from EncoderDecoderASR.from_hparams (iโ€™ve checked asr_model.modules.decoder and asr_model.transcribe_batch)

Is there something I am missing?

Thanks for help :slight_smile:

Hi,
ctc_lin is a linear transformation applied after he encoder (on the top of which CTC loss is applied). seq_lin, instead, it is applied after the decoder (on top of which NLL loss is applied).
both ctc_lin and seq_lin are used for inference as well, as you can see from the inference hyperparameters (hyperparams.yaml ยท speechbrain/asr-crdnn-transformerlm-librispeech at main).