Deepspeech 0 9 2 released

[JRMeyer]

>>> othiele
[December 3, 2020, 7:23pm]

Yippie, we got a new release. Check the release
notes.

Notable points for 0.9.x, please add more feedback

[This is an archived TTS discussion thread from discourse.mozilla.org/t/deepspeech-0-9-2-released]

[JRMeyer]

>>> othiele
[November 4, 2020, 8:44am]

Deprecated: Docs are up and running.

The new docs aren't up yet, use the 0.8.2 docs in the
meantime.

[Archived Post]

[JRMeyer]

>>> bfss
[November 4, 2020, 8:34am]

Wow! Chinese model! You are the best! Thanks very much for your work! slash
Nice slash ~

[Archived Post]

[JRMeyer]

>>> lissyx
[November 4, 2020, 8:43am]

In fact they are: https://deepspeech.readthedocs.io/en/r0.9/ we had a
bug on readthedocs, and v0.9.0 tag cannot be generated. But we could fix
it on the r0.9 branch, so I've made sure the redirection is there. When
we do a dot release, it should fix as well on RTD.

[Archived Post]

[JRMeyer]

>>> dan.bmh
[November 4, 2020, 11:07am]

The newly used noise augmentation does improve my accuracy in noisy
environments greatly: slash


pre-transcription
processing](https://discourse.mozilla.org/t/the-noise-reduction-componet-in-the-pre-transcription-processing/69264/14)
drop-close='true'
.box}

> Thanks for sharing your approaches. It did motivate me to run a
> benchmark for comparison:
> In my experiments frequency filtering only had a very small impact.
> The noise reduction (with rnnoise) did help much, but also can lower
> the accuracy in more silent environments. Note that the benchmark did
> not test transcription accuracy directly, because I'm doing an
> additional step afterwards (Speech - slash > Text - slash > Intent+Slots). The
> benchmark code can be found
> here
> Update: Th...

[Archived Post]

[JRMeyer]

>>> othiele
[November 4, 2020, 12:49pm]

Great to hear that. There are some reports on worsening
results for clean
speech with the newly released English model with noise augmentation.
Maybe this English model is better for noisy envoirnments (which was
requested a lot) than for clean speech.

[Archived Post]

[JRMeyer]

>>> dan.bmh
[November 4, 2020, 4:01pm]

> is better for noisy envoirnments than for clean speech.

In the benchmark I used above, the clean speech result did get better
too (90% - slash > 94%). I didn't expect this, because LibriSpeech WER got
worse, so I think this depends very much on the dataset.

[Archived Post]

[JRMeyer]

>>> alansiu
[November 5, 2020, 4:39am]

I am glad to see to new version released which support the Chinese
language. slash
And you also provide the checkpoint, pbmm and scorer files.

May I ask you to provide the lm_filtered.arpa and alphabet.txt also? slash
Because I also keep training a model with my city accent. It is use for
the reference only. Thank you.

[Archived Post]

[JRMeyer]

>>> dkreutz
[November 5, 2020, 6:06pm]

Do I get it right, that DeepSpeech 0.9.1 fixes the documentation issue
and there was no change to native client/binaries compared to v0.9.0?

[Archived Post]

[JRMeyer]

>>> othiele
[November 5, 2020, 6:38pm]

, you absolutely right, there
was some problem with readthedocs and the only solution was to push it
to 0.9.1
Thanks for the reminder

[Archived Post]

[JRMeyer]

>>> lissyx
[November 6, 2020, 8:05am]

And we got a contribution for Python 3.9 support, so we might do a 0.9.2
soonish, and how soon will depend on whether we get other bugs reported
or not

[Archived Post]

[JRMeyer]

>>> othiele
[December 3, 2020, 7:24pm]

And 0.9.2 was just released ...

[Archived Post]