Add a face to your Baseten speech models

LemonSlice puts a lifelike, interactive avatar on top of the speech models you serve on Baseten. Whether you run Whisper-class STT, open TTS, or fine-tuned checkpoints, the interactive avatar turns that audio into natural, face-to-face conversation.

LemonSlice

At a glance

Categories
Voice (TTS), Transcription (STT + VAD)
Status
Compatible (works with any)
Type
Avatar API
Languages
Python, Node.js
Setup time
~5 minutes

How the integration works

Baseten keeps doing what it does today. LemonSlice turns the audio into a face.

  1. Keep your Baseten deployment

    Your open speech models keep running on Baseten exactly as deployed. Model weights, endpoints, and scaling configuration stay untouched.

  2. Stream the audio to LemonSlice

    Point the audio output from your Baseten-served TTS model at the LemonSlice avatar API. The avatar renders in real time with accurate lip sync and expression.

  3. The avatar video streams to your users

    LemonSlice delivers the avatar video over WebRTC, natively via LiveKit, Daily (Pipecat), or Agora. Your users see a responsive face in about the blink of an eye.

Add your real-time avatar with a few lines of code

A Baseten session with a LemonSlice avatar.

Choose a code example
from livekit import agents
from livekit.agents import AgentSession, RoomOutputOptions
from livekit.plugins import lemonslice

async def entrypoint(ctx: agents.JobContext):
    await ctx.connect()

    session = AgentSession(
        # Add Baseten for TTS and STT, plus your LLM, here
    )

    avatar = lemonslice.AvatarSession(
        # Publicly accessible image URL for the avatar
        agent_image_url="...",
        # Prompt to guide the avatar's demeanor while speaking
        agent_prompt="a person talking",
        # Recommended. Adapts video quality to each viewer's bandwidth
        simulcast=True,
    )

    # Start the avatar and wait for it to join
    session_id = await avatar.start(session, room=ctx.room)

    # Start your agent session with the user
    await session.start(
        # ... room, agent, room_options, etc....
    )

An avatar that listens, reacts, and feels human

Control actions and emotions

The Action Engine moves face, hands, shoulders, and posture together. You trigger actions and emotions from the API, down to a prompt like "talk with hand gestures."

Works with any character, in any style

LemonSlice is the only avatar model that brings non-humans to life, a turtle or a smiling teacup, with the same ease as cartoons and photorealistic people.

Ultra-low latency

LemonSlice leads every major avatar provider on response time, so replies feel instant.

Create an avatar from one image

Turn any image into an interactive avatar. No training or fine-tuning needed. Make unlimited custom characters and change outfits or settings in real time.

Built for enterprise scale

Avatar sessions alongside Baseten run at production scale.

Thousands of concurrent calls

Scale avatar sessions into the thousands, ready for high-volume production workloads.

Multi-region servers

Place avatar sessions in the region nearest your users so latency stays minimal.

Zero data retention mode

Store no conversation data during avatar sessions, made for privacy-sensitive enterprise workflows.

What builders ship with Baseten + LemonSlice

Create avatars for everything from language practice to customer support, live streaming, and airport kiosks.

Language practice partner

Practice real conversations in another language.

Customer support agent

Answer, triage, and resolve customer questions face to face.

Streaming host

A live host that reads and responds to chat.

Frequently asked questions

Stream the audio output from your Baseten-served TTS model to the LemonSlice avatar API. The interactive avatar lip-syncs in real time and the video streams back to your users over WebRTC.
No. Baseten serves your speech models and handles inference, while LemonSlice handles the visual layer, a real-time avatar rendered from that audio.
Yes. LemonSlice accepts streamed audio, so any open TTS model you deploy on Baseten works, including custom and fine-tuned checkpoints.
Yes. Your Baseten-hosted transcription handles the listening side of the pipeline as usual, and LemonSlice animates the interactive avatar from whatever audio your response stage produces.
No. Model weights, endpoints, and autoscaling configuration stay exactly as deployed. LemonSlice only consumes the audio your endpoint already streams out.
LemonSlice renders the avatar in real time as audio arrives, so conversations keep feeling instant. Multi-region servers run each session close to your users to keep latency as low as possible.

Give your Baseten speech stack a face

Stream your model's audio into a live avatar with a 5-minute API setup.