Training custom voice doesnt train

[JRMeyer]

>>> dubreuia
[July 8, 2019, 11:52am]

Hello,

I'm trying to train a custom voice, with 24 hours of an audio book
recording and associated transcription in LJSpeech format for the
dataloader (I've posted a sample on my github
https://github.com/dubreuia/hosting/tree/master/mozilla-tts/custom-dataset-sample).
I've used the notebooks AnalyzeDataset and CheckSpectrograms to check my
data first, it looks good to me.

After 60K iter (see screenshots), I have nothing (I've trained to 100K
to no avail), the audio is blank, or faintly humming. I've trained on
LJSheech, I have good results at 100K as it should.

187

slash

78.8

It is probably similar to [Custom voice - TTS not
.

Does someone have an idea why my dataset wouldn't train?

Thanks, slash
Alex



[Custom voice - TTS not learning

[This is an archived TTS discussion thread from discourse.mozilla.org/t/training-custom-voice-doesnt-train]

[JRMeyer]

>>> erogol
[July 8, 2019, 1:34pm]

You are overfitting as the plots show. Take a good care of the
difference between eval and train stats.

In case: https://en.wikipedia.org/wiki/Overfitting

How large is your dataset? Do you have any samples to share?

[Archived Post]

[JRMeyer]

>>> dubreuia
[July 8, 2019, 1:52pm]

for this project and
for helping me out.

> You are overfitting as the plots show. Take a good care of the
> difference between eval and train stats.

Yeah I can see my validation error going up while training error going
down.

> How large is your dataset? Do you have any samples to share?

I've posted a small sample on my github
https://github.com/dubreuia/hosting/tree/master/mozilla-tts/custom-dataset-sample
but I can post more. I have 22 hours or recordings, split in 14995
chunks, lengths varying from 1 seconds to 10 seconds (with most recs at
4,5,6 seconds).

I'm thinking maybe the sample rate is wrong (because I split the audio
from an mp3 file, then converted it to wav), but I've used both
notebooks to tune my config so I'm pretty sure it is good.

[Archived Post]

[JRMeyer]

>>> erogol
[July 8, 2019, 2:06pm]

you can check SR with soxi command. Make sure it is the same with
config.json.

However, the specs sound good for your dataset.

You can also use dataset analysis folder and the notebooks there to see
more about your dataset.

I also see some pre-mature ends in some of the samples you share

It is, in general, the data quality that hinders TTS, Just iterate
through your dataset until it works.

[Archived Post]

[JRMeyer]

>>> dubreuia
[July 8, 2019, 5:07pm]

> I also see some pre-mature ends in some of the samples you share.

Yes, I'll try splitting the audio on pause instead of word count as I'm
doing right now. Do you think having no punctuation plays a role?

> It is, in general, the data quality that hinders TTS, Just iterate
> through your dataset until it works.

Do you think the audio quality is sufficient? My audio has a lot less
dynamics then the LJSpeech and the speaker is a lot more monotonic.

Thanks,

[Archived Post]

[JRMeyer]

>>> erogol
[July 9, 2019, 4:37pm]

{.avatar

> Yes, I'll try splitting the audio on pause instead of word count as
> I'm doing right now. Do you think having no punctuation plays a role?

yes it prevents attention to work and learn properly.

{.avatar

> Do you think the audio quality is sufficient? My audio has a lot less
> dynamics then the LJSpeech and the speaker is a lot more monotonic.

Sounds good but you need to check it quantify it as I told.

[Archived Post]