Automatically setting `multispeaker` and `multilingual` should look at all datasets not just whether a specific dataset is multilingual or multispeaker

EveryVoiceTTS / EveryVoice

The EveryVoice TTS Toolkit - Text To Speech for your language

https://docs.everyvoice.ca

Other

13 stars 1 forks source link

Automatically setting `multispeaker` and `multilingual` should look at all datasets not just whether a specific dataset is multilingual or multispeaker #470

Open roedoejet opened 1 month ago

roedoejet commented 1 month ago

We might use the wizard to combine multiple single-speaker, single-language datasets. In doing so, multilingual and multispeaker never get set to True by default, but they should be.

So, we need:

meta logic at the end of the wizard process to scan all the data for languages and speakers
as soon as there is more than one language, multilingual should be set to true
if the speaker is set to default in two different languages - the speaker name should actually be set to something different. i.e. the speaker default in a single-speaker English dataset is probably a different speaker than default in a single-speaker Sinhala dataset.
if we find more than one speaker across dataset, multispeaker should be set to true

joanise commented 1 month ago

The default speaker could systematically be set to "default"+lang_code, even with a single-speaker corpus? That would simplify the logic in the case of multiple single-speaker corpora.

roedoejet commented 1 month ago

The default speaker could systematically be set to "default"+lang_code, even with a single-speaker corpus? That would simplify the logic in the case of multiple single-speaker corpora.

true - even that might not be enough though. I mean, you could have two single speaker corpora from the same language. maybe default_<dataset_index>?

wiitt commented 1 month ago

Although, I find it very unlikely but there could be two or more data sets of different languages but with the same speaker Dealing with this case, I guess, will require additional question in the wizard

joanise commented 1 month ago

Dealing with this case, I guess, will require additional question in the wizard

Or we might say that such cases need to be manually configured by editing the config files afterwards.

wiitt commented 1 month ago

Dealing with this case, I guess, will require additional question in the wizard

Or we might say that such cases need to be manually configured by editing the config files afterwards.

Config files do not include information with lists of speakers and languages. So, manual processing will be required for rows in text files which could be large.

In case when there are data sets of just one speaker, multispeaker option could easily be switched to false for sure. But if there are, for instance, three sets:

lang_1, speaker_1
lang_2, speaker_1
lang_3, speaker_2

it could be problematic.

joanise commented 1 month ago

So maybe the wizard could add for a speaker name when there is no speaker column in the data, and offer "default"+lang by default but let the user change it as they wish?

This would parallel what we do for lang, where we demand a language code when there isn't a language column.

wiitt commented 1 month ago

I think it would work. That's what I was thinking about, adding a question in the wizard about speakers.