Other testing setup?

[danpovey]

I have been doing most of the acoustic model tuning on the librispeech setup, but the WERs don't seem to move below around 6.8% whatever I do.

I had a look at the detailed sentences the model is outputting and I think the issue is that the acoustic model is about as good as it can be already. Almost all the errors that I can see seem to be either extremely rare words (mostly names) or OOVs or language modeling issues (e,g. "in" vs. "an").

I think for purposes of model tuning it would be better if we prepared the test_other data because it's more acoustically challenging.

Eventually we can investigate modeling techniques that use neural LMs; these may be necessary to get really low WERs on this type of data.

[pzelasko]

Cool! I'm adding dev/test other in #134

[rickychanhoyin]

are these results from train-clean100 or the fullset of librispeech ?

[pzelasko]

It’s the 100h subset

Wiadomość napisana przez rickychanhoyin @.***> w dniu 3/29/21, o godz. 06:53:

 are these results from train-clean100 or the fullset of librispeech ?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.