LiveKit
Agents (Python)
The project is consolidating its LLM and STT/TTS ecosystem around reliability and seamless provider switching. v1.5.0 shipped adaptive interruption handling—an ML model that distinguishes genuine user interruptions from false positives like backchannels with 86% precision and 100% recall, automatically resuming playback when misdetected. v1.5.1 and 1.5.2 extended provider support (Mistral SDK v2, translation for Gladia and Soniox, WebSocket streaming for Baseten TTS) while fixing Azure OpenAI realtime compatibility and tool schema validation for function calling. The 1.4.x series tackled infrastructure stability—v1.4.2 eliminated multiple memory leaks in process pools and IPC pipelines, improved STT/TTS fallback robustness, and introduced a new RecordingOptions API—alongside incremental provider fixes (Telnyx, SambaNova, Sarvam codec formats, Google denoiser) and tighter schema handling for nullable enums and tool parameters.
March shipped a major interruption-handling redesign centered on an ML model that distinguishes genuine user interruptions from incidental sounds like backchannels and coughs — hitting 86% precision with 100% recall and rejecting 51% of traditional VAD false positives. Alongside this, the month added translation support to Gladia and Soniox STT, introduced the Telnyx and SambaNova plugins, upgraded Cartesia TTS to Sonic 3, and fixed Azure OpenAI realtime support and Google realtime's deprecated mediaChunks handling.