Continuous training

[JRMeyer]

>>> reyxuan
[August 13, 2019, 7:05am]

Hi! slash
I'm preparing a Spanish model with Common Voice 40h, Caito
dataset and Open SLR
dataset. These are 200 hours, which are not
so many. The possitive fact is that I'll be receiving 1h of
audio+transcription every day. Would it be better to wait for say 1000h
or train the model with the audio I currently have while fine tuning it
every day with that hour of audios? Thanks!

[This is an archived TTS discussion thread from discourse.mozilla.org/t/continuous-training]

[JRMeyer]

>>> kdavis
[August 13, 2019, 7:31am]

That's a good question. I guess maybe my question to you is:

> Does the model created from 200 hours provide a
> 'MVP' for the
> use case you're working on?

[Archived Post]

[JRMeyer]

>>> reyxuan
[August 13, 2019, 9:01am]

I don't think so. But I won't use the model until it works well. The
idea of continuos training is saving time. Using just 1h of audio each
day will save resources. I guess...

EDIT: There is also the option of Data Augmentation. With 200h I could
get +500h with this technique. And the same with the daily hour of
audios.

[Archived Post]

[JRMeyer]

>>> caucheteux
[August 13, 2019, 11:34am]

,



> EDIT: There is also the option of Data Augmentation. With 200h I could
> get +500h with this technique. And the same with the daily hour of
> audios.

do you have the link for this method ?

[Archived Post]

[JRMeyer]

>>> reyxuan
[August 14, 2019, 6:38am]

You can check this
link. It is an
implementation from a user but also Mozilla developers explain their
techniques there. Hope you find it useful.

[Archived Post]

[JRMeyer]

>>> caucheteux
[August 14, 2019, 11:02am]

Very useful, it seems that they'll implemented it in Mozilla
DeepSpeech. slash
Did you use it already ?

[Archived Post]

[JRMeyer]

>>> reyxuan
[August 14, 2019, 11:28am]

Not yet. At this moment I'm training the model with Common Voice data on
Google Colab, while waiting for a GPU.

[Archived Post]