How to train deepspeech that something isnt speech

[JRMeyer]

>>> dabinat
[July 19, 2019, 5:46pm]

I've noticed that DeepSpeech will sometimes interpret instrumental
music, sound effects and animal noises as speech. How I can I tell it
that something isn't speech - would I just provide a blank transcript
with the clip?

Or is it not necessary - will DeepSpeech get better at interpreting this
as it gets more samples of what speech actually is?

[This is an archived TTS discussion thread from discourse.mozilla.org/t/how-to-train-deepspeech-that-something-isnt-speech]

[JRMeyer]

>>> kdavis
[July 19, 2019, 5:53pm]

We have, without the express purpose of what your are suggesting, done a
bit of this.

One of our data sets, Fisher, has lots of 'ummms', 'ahhhs', and 'hmmms'
in it and we didn't give deep speech the transcript for such
disfluencies. But it did have the transcript of the surrounding fluent
speech. So it tends not to transcribe such disfluencies.

So you could do the same by adding background noise, such as music or
animal noises, to your standard training data then train or fine tune on
that data set.

[Archived Post]

[JRMeyer]

>>> dabinat
[July 19, 2019, 10:14pm]

Thanks, I'll just extend the clip so it contains actual speech and
transcribe that.

[Archived Post]