bbb vs galene

Posted Sep 23, 2022 3:16 UTC (Fri) by pabs (subscriber, #43278)
In reply to: bbb vs galene by WolfWings
Parent article: Two visions for the future of sourceware.org

> automated subtitling of each speaker

This new "open" multi-lingual TTS system would be useful for that. Apparently it is better quality than other options that exist such as Kaldi/DeepSpeech/coqui/etc.

https://openai.com/blog/whisper/
https://news.ycombinator.com/item?id=32927360

Please note that according to the HN comments, while the model/code is freely licensed, the audio/text data used for training/evaluating the model is not public and not freely licensed.

bbb vs galene

Posted Sep 23, 2022 3:16 UTC (Fri) by pabs (subscriber, #43278) [Link]

(In addition, it is a very large amount of data ~600K audio hours and presumably took a long time to train, so retraining is not feasible for most people)