Moshi by Kyutai

Open-source speech-to-speech AI model from French research lab Kyutai. Features full-duplex conversation (can listen and speak simultaneously), emotional expression, and real-time interaction without intermediate text transcription.