livekit-plugins-reson8

Reson8 STT plugin for LiveKit Agents.

A single reson8.STT class that adapts to how LiveKit uses it:

• Streaming (stream(), used by voice agents) connects to the turn-aware endpoint. Reson8 detects conversational turn boundaries server-side: it emits a preflight transcript (an eager guess that the turn is over) that your agent can start responding to, then confirms it as a final transcript — or cancels it if the speaker keeps talking. Great for low-latency voice agents.
• Batch (recognize()) transcribes pre-recorded audio and returns the full transcript.

Languages

Leave language unset to auto-detect the spoken language, or pin recognition to one or more supported codes. You can pass a single code, a comma-string, or a list — a list is normalized to Reson8's comma-joined form and any unsupported code raises ValueError locally.

reson8.STT()                        # auto-detects the spoken language
reson8.STT(language="en")           # English only
reson8.STT(language="nl,de")        # Dutch or German
reson8.STT(language=["nl", "de"])   # same, as a list

Supported languages (reson8.SupportedLanguages):

Installation

pip install livekit-plugins-reson8

Usage

from livekit.plugins import reson8

stt = reson8.STT(
    api_key="your-api-key",   # or set RESON8_API_KEY
    # language is auto-detected when omitted; pass one or more supported codes to pin it
)

With a Voice Agent

from livekit.agents import AgentSession
from livekit.plugins import openai, reson8

session = AgentSession(
    stt=reson8.STT(),          # streaming + turn detection, language auto-detected
    llm=openai.LLM(),
    tts=openai.TTS(),
    # "stt" hands turn detection to Reson8 and lets the agent start
    # generating on our preflight transcript instead of the confirmation.
    turn_handling={
        "turn_detection": "stt",
        "preemptive_generation": {"enabled": True},
    },
)

By default LiveKit runs its own turn detector. Set turn_handling as above to hand turn-taking to Reson8 instead.

Transcribing a file

event = await reson8.STT().recognize(audio_buffer)
print(event.alternatives[0].text)

Configuration

STT.update_options(...) changes settings at runtime; active streaming sessions reconnect automatically to apply them.

Turn detection

Reson8 decides turn boundaries by confidence: it emits the preflight transcript at eager_turn_probability and commits the turn at final_turn_probability. Lower final_turn_probability to commit sooner, at the risk of cutting off longer utterances.

flush() commits the current turn immediately, keeping final_turn_probability intact. LiveKit never calls it for you.

See Turns for how turn events work server-side.

Running the example

cp .env.example .env
# Fill in your keys
python examples/voice_agent.py dev

Development

We use uv to manage the environment and dev dependencies, so you'll need to have it installed.

uv sync sets everything up — it reads the pinned Python version from .python-version, creates a .venv/, and installs the dev tools (Ruff and mypy).

Run the tooling through uv run, which uses the project environment without you having to activate anything:

uv run ruff check .    # lint
uv run ruff format .   # format
uv run mypy            # type-check