Mirco Ravanelli

Hi @30stomercury, thank you very much for this PR! I like the idea of making beamsearch more modular by separating the "search" part from the "rescoring" one. Let me share...

Let me also put on the table one idea for the `ScorerBuilder`. At this moment it is a bit hardcoded and not that scalable to other scorers possibly defined by...

As for point 4, this PR already requires us to change all the ASR recipes. So we can really take this opportunity to do the change that adds the NeuralLMScorer...

Hi @30stomercury. I like method2 more because looks more modular and users can modify the scoring pipeline directly in yaml file. To make it even more modular, I would suggest...

Makes sense. What is your preference for the yaml part? On Fri, 25 Jun 2021 at 10:01, Sung-Lin Yeh ***@***.***> wrote: > Hi @mravanelli , the partial scorers > score...

I'm fine with both solutions. Let's hear a bit the others (e.g., @Gastron, @TParcollet ,..)

Hi @30stomercury, I'm fine with that. This is very transparent because the beasearcher always returns the topk hypothesis and then the user can select the best one. As for the...

Thank you Aku, the refactor is already backward incompatible because we have to change all the yaml files using beamsearch. This is an opportunity to change everything once and then...

Dear @jamfly, thank you for this PR! How the numbers you got are compared to the state-of-the-art results on this dataset? I found some numbers [here](https://paperswithcode.com/sota/speech-to-text-translation-on-must-c-en-de), but I'm not sure...

Hi @jamfly, yes it could be interesting to see the numbers with wav2vec and eventually extend it to all the languages. As for the performance, my understanding is that we...