Actually there's a whole new subfield called textless NLP doing just that: Learning language models from raw audio. https://ai.facebook.com/blog/textless-nlp-ge

Actually there's a whole new subfield called textless NLP doing just that: Learning language models from raw audio. https://ai.facebook.com/blog/textless-nlp-generating-expressive-speech-from-raw-audio/ https://ai.facebook.com/blog/textless-nlp-generating-express...