Mirco Ravanelli

Thank you for the updates! It look like now, the performance with ctc +wfst is better than the one with ctc only. I'm wondering if we have now to close...

Hi, it might be helpful to take a look at the official pre-trained models + inference functions that you can find on HuggingFace: - https://huggingface.co/speechbrain/asr-crdnn-rnnlm-librispeech (CRDNN model, test acc=3.09%) -...

Weird, batch decoding is supported. For instance, when we do validation and testing (see e.g., https://github.com/speechbrain/speechbrain/blob/develop/recipes/LibriSpeech/ASR/seq2seq/hparams/train_BPE_1000.yaml) we use batches with more than one sentence (and achieving the expected performance). Are...

Dear @Kuray107, I did some experiments and I'm not able to replicate your problem. To help you find the issue, I created a Colab notebook where you can see how...

Interesting, I will take a deep look as soon as I can. Meanwhile, I would encourage @30stomercury and @TParcollet to take a look as well. It looks to me that...

I have created this [PR](https://github.com/speechbrain/speechbrain/pull/933) with the goal of fixing step by step all the batch-dependent computations. Any help is welcome!

Sure, we can conclude this PR and merge it as soon as I will be back in Montreal. On Tue, Oct 12, 2021, 7:39 PM Gaétan Lepage ***@***.***> wrote: >...

The batch-independent PR (#933) is still a work in progress. Everything works fine there (feel free to use it), but we have some jitability issue with the new MaskCapableSequential class...

@aheba any update on this PR?

@aheba, any update of the GigaSpeech recipe?