LiveKit Agents v1.8.10 records calls at a constant bitrate and moves Mistral to Voxtral 3

LiveKit Agents v1.8.10 lets RecorderIO save calls at a constant bitrate. Players like QuickLook stop misjudging a recording's length, and you can match Egress at 128 kilobits a second.

Slide text

LiveKit Agents Added

LiveKit Agents v1.8.10 records your calls at a constant bitrate

RecorderIO(agent_session=session, bit_rate=128_000, constant_bit_rate=True)

Defaults are unchanged

GitHub release · livekit-agents@1.8.10 · Recording bitrate control

Reson8 arrives as a speech-to-text plugin that detects turns on its own server. Mistral adds Voxtral 3 for transcription and speech, with 36 voices, and Inworld adds its TTS 2 Flash model.

Slide text

LiveKit Agents Added

Reson8 and Mistral Voxtral 3 bring new speech models

Reson8 STT plugin · batch and realtime · server-side turn detection

Mistral Voxtral 3 STT and TTS · 36 voice presets

Inworld inworld-tts-2-flash · plugin and LiveKit Inference

GitHub release · livekit-agents@1.8.10 · New providers and models

If your agent leaves the voice unset, AssemblyAI now speaks as Audrey. Inworld moves its default model from TTS 1.5 to TTS 2, so the same agent can sound different after the upgrade.

Slide text

LiveKit Agents Changed

AssemblyAI and Inworld agents may sound different after the upgrade

AssemblyAI TTS voice defaults to Audrey

Inworld TTS defaults to inworld-tts-2, off TTS 1.5

Only agents that leave the voice or model unset

GitHub release · livekit-agents@1.8.10 · Behavior changes

A streaming speech-to-text connection that stalls is now caught and its retries capped, so your fallback provider takes over. Worker health reports 503 until the worker registers, and Gemini thinking tokens now count toward usage.

Slide text

LiveKit Agents Changed

A stalled speech-to-text stream no longer holds up your fallback

Streaming STT Stalled connections detected · retries capped

Worker health 503 before registration and after a 30-second grace

Gemini Thinking tokens count in completion_tokens

GitHub release · livekit-agents@1.8.10 · Behavior changes · #7891

The Mistral plugin now uses only the Conversations API, defaults to Voxtral 3, and drops two older Voxtral models. If you run Mistral, I would test it on a branch first, since defaults and model names both move.

Slide text

LiveKit Agents Breaking

Mistral moves to Voxtral 3 and the Conversations API

LLM api_mode is ignored · Conversations only

STT voxtral-mini-latest and voxtral-mini-2507 are gone

Defaults Voxtral 3 · mistral-small-latest · no automatic Silero

Requires mistralai 3.5 or later · steps in the LiveKit Agents v1.8.10 release notes

GitHub release · livekit-agents@1.8.10 · Breaking changes