Expected amount of data needed in hours to get good model on a new language❓

[ErenBalatkan]

❓ Questions and Help

How much labeled data in terms of speech hours would you say we need to train a good model on a new language?

[snakers4]

For minor languages (non English) with limited dialects about 200-300h per domain, at least 4-5 domains

[matanox]

What about number of different speakers?

[snakers4]

As for speakers - I am not sure, since we did not really track it when building our models, but usually with in the wild data - data quality is the issue, not the speaker diversity
As a rough estimate I would suppose hundreds / thousands

[ErenBalatkan]

Thanks, closing