LiveKit Integration

• lokutor.STT: Conv 2.0 speech to text, streaming over /ws/stt.
• lokutor.TTS: Versa 2.0 text to speech, streaming over /ws/tts.

​

Installation

pip install "livekit-agents[lokutor]~=1.8"

GIT_LFS_SKIP_SMUDGE=1 pip install "livekit-plugins-lokutor @ git+https://github.com/livekit/agents@main#subdirectory=livekit-plugins/livekit-plugins-lokutor"

​

Authentication

​

A voice agent

from dotenv import load_dotenv

from livekit.agents import Agent, AgentServer, AgentSession, JobContext, cli
from livekit.plugins import lokutor, openai, silero

load_dotenv()

server = AgentServer()


@server.rtc_session()
async def entrypoint(ctx: JobContext) -> None:
    session = AgentSession(
        stt=lokutor.STT(language="es"),
        llm=openai.LLM(),
        tts=lokutor.TTS(voice="F1", language="es"),
        vad=silero.VAD.load(),
    )
    await session.start(
        agent=Agent(instructions="Eres una recepcionista amable. Responde en frases cortas, sin listas ni emojis."),
        room=ctx.room,
    )
    await session.generate_reply(instructions="Saluda y pregunta en qué puedes ayudar.")


if __name__ == "__main__":
    cli.run_app(server)

​

Speech to text

stt = lokutor.STT(language="en")

• Streaming: interim transcripts while the user speaks, then a final transcript when the utterance ends, after about 0.7 s without speech (how utterances end). The text comes back written out, with punctuation, casing and digits.
• Audio: frames at any sample rate; the plugin resamples to the 16 kHz the API takes.
• One clip: await stt.recognize(frames) transcribes a list of rtc.AudioFrame with one POST /stt/transcribe request.
• Not exposed: custom vocabulary (keyterms) and word timings.

​

Text to speech

tts = lokutor.TTS(voice="F1", language="en")

• Audio: 16-bit PCM, 44.1 kHz, mono. LiveKit resamples it if the room needs another rate.
• Sentences: the first sentence is sent as soon as the tokenizer completes it, so the agent starts speaking before the LLM has finished the reply.
• Pronunciation: your account’s default pronunciation dictionary applies to every sentence.
• Server defaults: temperature (0.3) and the other /ws/tts fields the plugin does not send keep their defaults.

​

Outside an agent

import asyncio

import aiohttp
from livekit.plugins import lokutor


async def main() -> None:
    async with aiohttp.ClientSession() as http:
        tts = lokutor.TTS(voice="F1", language="es", http_session=http)
        stt = lokutor.STT(language="es", http_session=http)

        stream = tts.stream()
        stream.push_text("Hola, soy tu asistente. ")
        stream.push_text("¿En qué puedo ayudarte hoy?")
        stream.end_input()
        frames = [ev.frame async for ev in stream]

        event = await stt.recognize(frames)
        print(event.alternatives[0].text)  # Hola, soy tu asistente. ¿En qué puedo ayudarte hoy?

        await tts.aclose()
        await stt.aclose()


asyncio.run(main())

​

Concurrency and billing

• Two sessions per agent. While an AgentSession runs it holds one /ws/stt stream and one /ws/tts socket, and each counts as one of your account’s concurrent sessions (Limits): 3 before the first top-up, 10 after. An account carries one LiveKit agent at a time before its first top-up and five after; past that, the next connection is refused (429 / auth.rate_limited). Higher limits on request.
• Billing. The same prices as calling the API directly, from your prepaid balance: speech to text per minute of audio, text to speech per million characters (rate card). The LLM and LiveKit itself are billed by their providers.

​

When to use Lokutor’s own agent instead

​

Resources

• Source: https://github.com/livekit/agents/tree/main/livekit-plugins/livekit-plugins-lokutor
• Pull request: https://github.com/livekit/agents/pull/7256
• PyPI: https://pypi.org/project/livekit-plugins-lokutor/ (from the next LiveKit Agents release)
• LiveKit Agents docs: https://docs.livekit.io/agents/