Trying to create own czech model

[JRMeyer]

>>> MrDiscordMan
[February 17, 2021, 7:21am]

Hello, slash
I try to train my own TTS model, but I am already in 95800 step and I
can't understand a single word from the output. I am lost, and I don't
know what I can do anymore. Any ideas what I did wrong?

Dataset:
https://drive.google.com/drive/folders/13CSvJhH68C7BqepyOdRts9w0IY5hqILw?usp=sharing slash
Model:
https://drive.google.com/drive/folders/1V9ILmK31SV8vnN-Et2y1BW6lb8njlvNS?usp=sharing

Any help is appreciated, thanks.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/trying-to-create-own-czech-model]

[JRMeyer]

>>> erogol
[February 17, 2021, 1:27pm]

Can you share a Colab to try the model?

Do you plan to release the model open source?

Is this an open source dataset?

Maybe we can work together on that to make Czech available on the TTS.

[Archived Post]

[JRMeyer]

>>> MrDiscordMan
[February 17, 2021, 1:53pm]

I don't plan on making this one open source because I'm doing it for
friend, but I would definitely collaborate on an open source czech
model.

[Archived Post]

[JRMeyer]

>>> erogol
[February 17, 2021, 2:10pm]

do you know any open dataset ?

[Archived Post]

[JRMeyer]

>>> MrDiscordMan
[February 17, 2021, 2:12pm]

Yes, but it is under CC-0 slash

commonvoice.mozilla.org

### Common Voice by Mozilla

Common Voice is a project to help make voice recognition open to
everyone. Now you can donate your voice to help us build an open-source
voice database that anyone can use to make innovative apps for devices
and the web.

[Archived Post]

[JRMeyer]

>>> erogol
[February 17, 2021, 2:13pm]

but does it have a enough size single speaker subset?

[Archived Post]

[JRMeyer]

>>> MrDiscordMan
[February 17, 2021, 2:15pm]

how much do you need?

[Archived Post]

[JRMeyer]

>>> mrthorstenm
[February 18, 2021, 11:20am]

It's hard to tell a number, because it primarily depends on a good
phoneme coverage. But most single speaker datasets i know provide a
minimum of 16 hours (or more) of voice recordings.

Additionally this might help you: slash

{.site-icon
GitHub

### mozilla/TTS

forum: https://discourse.mozilla.org/c/tts) - mozilla/TTS

[Archived Post]

[JRMeyer]

>>> erogol
[February 18, 2021, 4:22pm]

I feel like bargaining
but at least 5 hours is like a good value to fine-tune a pre-trained
model.

[Archived Post]

[JRMeyer]

>>> MrDiscordMan
[February 18, 2021, 5:02pm]

I have about 5 hours of audio of the same voice.

Sample:
https://drive.google.com/file/d/1FeVXnKWIFHDepuXo44b1V1Lv-q8Vtb9J/view?usp=sharing

[Archived Post]

[JRMeyer]

>>> erogol
[February 18, 2021, 5:25pm]

let me check and thx for sharing.

[Archived Post]