Show HN: Caramel: a speech-to-speech conversation model trained only on audio

The author trained caramel-1, an open-source speech-to-speech conversation model with 413 million parameters that was trained exclusively on audio without ever seeing text. Operating at speeds fast enough for live conversation, the model learns turn-taking timing, overlaps, and backchannels directly from a curated dataset of public speech rather than via rule-based detectors. Open article →