LiveKit

LiveKit Agents builds realtime voice agents around an AgentSession. Bodhi provides the speech-to-text stage as a standard LiveKit stt.STT provider, so transcription behaves like any other LiveKit STT plugin.

Install

pip install "livekit-plugins-bodhi"

Needs livekit-agents 1.0 or newer. Verified against 1.8.6. For a full voice agent you will also want a VAD and an LLM plugin:

pip install "livekit-plugins-bodhi" "livekit-agents[silero]" "livekit-plugins-openai"

Add it to a session

import os

from livekit.agents import Agent, AgentSession, JobContext, WorkerOptions, cli
from livekit.plugins import bodhi, silero, openai


async def entrypoint(ctx: JobContext):
    session = AgentSession(
        stt=bodhi.STT(
            model="hi-general-v2-8khz",
            sample_rate=8000,
            hotwords=[bodhi.BodhiHotword("बजाज फिनसर्व", 2.0)],
        ),
        vad=silero.VAD.load(),
        llm=openai.LLM(model="gpt-4o-mini"),
        tts=bodhi.TTS(language="hi", voice="default_female"),
    )
    await session.start(
        agent=Agent(instructions="You are a helpful voice assistant."),
        room=ctx.room,
    )


if __name__ == "__main__":
    cli.run_app(WorkerOptions(entrypoint_fnc=entrypoint))

model is the only argument you usually set. The API key is read from $BODHI_API_KEY when api_key is omitted, and base_url defaults to wss://stt.navana.ai.

Arguments

Argument	Required	Description
`api_key`	No	Your platform API key, sent as the `x-api-key` header. Falls back to `$BODHI_API_KEY`.
`model`	No	Any transcription model. Defaults to `hi-general-v2-8khz`. See Models and languages.
`base_url`	No	Override the endpoint, usually to point at a different deployment. Defaults to `wss://stt.navana.ai`.
`sample_rate`	No	Rate audio is sent at, in Hz. The session’s audio is resampled to it automatically. Defaults to `16000`.
`language`	No	Language tag for transcripts. Defaults to the model’s prefix.
`interim_results`	No	Emit `INTERIM_TRANSCRIPT` events from partials. Defaults to on.
`parse_number`	No	Normalise spoken numbers, dates, and currency in the returned text.
`aux`	No	Attach timing and confidence metadata to each result. Required for `min_confidence`.
`min_confidence`	No	Drop finals below this utterance confidence. Defaults to `0.5`; pass `0` to keep everything.
`hotwords`	No	Phrase boosting, as `bodhi.BodhiHotword(phrase, score)`. `score` is optional, and roughly `1.0` to `3.0` is useful. See Advanced features.
`endpoint_silence_duration`	No	Trailing silence before an utterance is finalised, in seconds. Clamped to `0.44` to `1.2`.

How it maps onto LiveKit

Partial transcripts arrive as INTERIM_TRANSCRIPT speech events and endpointed finals as FINAL_TRANSCRIPT, which is the contract every LiveKit STT plugin follows. VAD, turn detection, and interruption handling are driven by the session, so swapping another provider for Bodhi does not touch the rest of the agent.

Python SDK

Using the SDK directly, without LiveKit.

Models and languages

Every model, and how to pick one.

Text-to-speech with LiveKit

The other half of the agent.