wav2letter

Mozilla DeepSpeech has had a release [1] that comes with a pre-trained model achieving 11% WER on clean audio in the LibriSpeech test corpus. That's close to the WER you're getting with Google, but I guess your audio quality isn't as good, so DeepSpeech would perform worse. Mozilla Common Voice [2] is the project to collect more training data so that DeepSpeech (and other projects) can achieve the accuracy that is known to be possible with the same architecture trained on larger private datasets. Then there's Facebook's newly released wav2letter++ [3], which claims to achieve better accuracy with the same training data. However, some people have been unable to exactly reproduce those results, getting "only" 5.15% WER [4]. Still better than what Mozilla DeepSpeech can deliver, though. [1] https://github.com/mozilla/DeepSpeech/releases/late

I'm not sure if those models would work with the version in the current master (wav2letter++). The old version of the master branch (wav2letter, written in lua/torch: https://github.com/facebookresearch/wav2letter/tree/wav2lett...) contains links to the models you listed, so I guess they belong to that version.

https://github.com/facebookresearch/wav2letter/issues/88 mentions a "pre-trained model named librispeech-glu-highdropout.bin", but I couldn't find it anywhere. https://github.com/facebookresearch/wav2letter/issues/93 also mentions a pre-trained model but without any reference which one or where to find it. Googling "librispeech-glu-highdropout.bin" still shows the text "luajit ~/wav2letter/test.lua ~/librispeech-glu-highdropout.bin -progress -show -test dev-clean -save -datadir ~/librispeech-proc/ -dictdir ~/librispeech-proc/ -gfsai ..." for https://github.com/facebookresearch/wav2letter/blob/master/R..., but clicking it, it's gone. But the Google Cache still has the result, including 3 pre-trained models: wget https://s3.amazonaws.com/wav2letter/models/librispeech-glu-highdropout.bin wget https://s3.amazonaws.com/wav2letter/model

It'll be interesting to see if anyone can reproduce their results, thus far its been troublesome: https://github.com/facebookresearch/wav2letter/issues/88

> a huge pain to set up. This project doesn't look like it's particularly easy to build: https://github.com/facebookresearch/wav2letter/blob/master/d...

Link to the GitHub repo, for the lazy: https://github.com/facebookresearch/wav2letter/