Caramel: a speech-to-speech conversation model trained only on audio
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Caramel-1 is an open-source 413M-parameter speech-to-speech conversation model trained only on about 11,500 hours of public audio, with no text at all. Its training data was filtered to match the conversation statistics of Dessa's examples, and it generates speech in real time, including on a MacBook Pro.