Open-source real-time speech-to-speech AI from French lab Kyutai — full-duplex conversations with sub-200ms latency, native voice in and voice out. Apache 2.0 license.