Skip to content

Best AI Avatar APIs for Bring Your Own LLM & TTS (2026)

If you already have a prompt, a knowledge base and a voice you like, you want an avatar API that takes your LLM and TTS instead of replacing them. This guide compares five APIs on exactly that: how each one accepts your own stack, where it draws the avatar, and what it charges at 1,000 and 3,000 minutes a month.

Why Bring Your Own

Keeping the LLM and TTS outside the avatar service lets you choose models, prompts and voices freely and swap them without touching the avatar layer. The cost is ownership: your team connects the services and handles streaming, audio formats, session state and recovery. Compare the three common setups below.

Interactive

Three ways to run the same conversation

Pick an architecture. Each box shows who runs it; each arrow shows what passes along it.

For example: Anam audio passthrough, D-ID Echo, LemonSlice

YouLLMWrites the reply
Reply text
YouTTSTurns the reply into speech
Your app sends the audio
Avatar vendorAvatar rendererDraws every video frame in the vendor’s cloud
Video stream
User’s deviceUser’s browser or appPlays the video
Trade-off

You choose the model and voice and own streaming, formats and recovery; the output is still a cloud video stream.

How We Compared

  • Your own stack: can you bring your own LLM, your own TTS audio, or both?
  • Where the avatar is drawn: does the API stream rendered video, or does an SDK draw the avatar on the device?
  • Price at the same usage: the cheapest public plan for the same monthly minutes, with at least 3 concurrent sessions.

Latency and bandwidth depend on configuration, workload and network, so this guide does not rank vendors on performance figures it could not verify.

Cost at Two Usage Levels

PlatformHow you bring your own stack1,000 minutes / month3,000 minutes / month
Monthly (USD / month)Annual (USD / year)Monthly (USD / month)Annual (USD / year)
SpatiusYour ASR, LLM and TTS; avatar drawn on the deviceStarter $19$189Builder $49$469
AnamYour TTS audio (passthrough) or its managed pipeline; cloud videoExplorer $154*No public annual priceGrowth $419*No public annual price
LemonSliceYour LLM or voice model; cloud videoProfessional $174.10*$1,889.20*Scale $532.03*$5,904.36*
TavusYour LLM inside its managed pipeline; cloud videoBuilder $347.75*20% off the plan fee (per its FAQ); not listedGrowth $924*20% off the plan fee (per its FAQ); not listed
D-IDYour LLM endpoint, or your own audio (Echo); cloud videoEnterprise only: self-serve API plans include a few hundred streaming minutes a month at most

Prices were checked on each linked pricing page on November 12, 2026. They cover the avatar service only and exclude your LLM, TTS and taxes. * includes minutes beyond the plan allowance at the plan’s published rate. Spatius Starter includes 2,000 minutes and Builder 5,500, so neither needs extra minutes here.

A few rules change the comparison:

  • Tavus includes its own LLM, TTS and WebRTC in the price, and rounds each conversation to 6 seconds with a 30-second minimum.
  • LemonSlice bills idle time like talking time, and on self-serve plans up to Scale a call above the concurrency limit cannot start (only Ultra and Enterprise allow bursts, at 2×).
  • D-ID self-serve API plans include a few hundred streaming minutes a month at most, so 1,000 minutes needs Enterprise pricing.

Platform Notes

Spatius

Spatius never runs your LLM or TTS. Your stack produces the reply audio; the Spatius Motion Server turns it into motion data, and the client SDK draws the avatar on the device GPU. Sending motion data instead of video is specific to Spatius: the stream runs at about 10–20 KB/s of motion data, plus audio. SDKs cover Web, iOS, Android and Flutter, and the public paths are Direct Mode, LiveKit Agents and Agora Conversational AI.

Trade-off: you integrate the conversation services yourself and test rendering on your target devices. The AvatarKit Scenes Demo has Web, iOS and Android examples with a shared backend.

Anam

Anam’s Python SDK documents audio passthrough: you send TTS audio in chunks (24 kHz mono is recommended) and receive synchronized audio and video frames. Anam also offers a managed pipeline with its own speech recognition, LLM and voice.

Trade-off: you keep control of the voice, but the output is still a cloud video stream.

LemonSlice

LemonSlice lists “bring your own LLM or voice model” and API access on its paid self-serve plans. A free chat preview exists, but API access requires a paid plan. Hosted avatars with speech, LLM and voice included cost an extra $0.08–$0.09 a minute.

Trade-off: concurrency rises with the plan (3 on Starter to 10 on Scale), so check the plan that fits your peak sessions.

Tavus

Tavus’s Conversational Video Interface accepts your own OpenAI-compatible LLM. Voices can come from your Cartesia, ElevenLabs or Azure account, but the TTS still runs inside the Tavus pipeline.

Trade-off: fewer services to orchestrate, less control over the audio path.

D-ID

D-ID agents can call your own OpenAI-compatible LLM endpoint. Echo sessions take a different route: your backend streams its own audio (WAV, MP3 or PCM16) over LiveKit and the avatar lip-syncs it, skipping D-ID’s speech recognition, LLM and TTS.

Trade-off: these are two separate modes, so pick the one that matches your voice and session flow.

The Pipeline in Four Steps

  1. Transcribe: the client sends the user’s speech to your ASR, which returns text and an end-of-turn signal.
  2. Reply: your backend sends the transcript and context to the LLM and owns prompts, retrieval, tools and session state.
  3. Speak: your TTS turns the reply into audio. Stream chunks only if both the TTS and the avatar input support streaming.
  4. Animate: the avatar service takes the audio. Cloud APIs return rendered video; Spatius returns motion data for its SDK to draw on the device.

Before building, confirm the audio format, chunking, session lifecycle, interruption and reconnect behavior, and which side owns playback.

How to Choose

What matters most for your project?

Spatius

Why it fits
Your stack keeps the LLM and TTS; Spatius sends motion data (about 10–20 KB/s) and draws the avatar on the device, with Web, iOS, Android and Flutter SDKs.
Confirm first
GPU load and rendering on the lowest-end devices you support.
1,000 / 3,000 min
$19 (Starter) / $49 (Builder) a month

FAQ

What is the difference between BYO LLM and BYO TTS?

BYO LLM means your own model writes the replies, with your prompts, history and retrieval. BYO TTS means your own voice engine produces the speech audio, which you send to the avatar API to drive lip sync. Some APIs accept only one of the two.

Why does it matter where the avatar is drawn?

Cloud APIs draw every frame on their servers and stream video, so the experience depends on a continuous video stream. Spatius sends motion data, about 10–20 KB/s plus audio, and draws the avatar on the device, so device performance becomes the thing to test.

Can I use an open-source LLM such as Llama or DeepSeek?

Yes, with any API that takes your own audio or an OpenAI-compatible LLM endpoint. Run the model on your own server, send its reply to your TTS, and pass the audio to the avatar layer.

What audio format do avatar APIs expect?

It varies. Anam recommends 24 kHz mono for audio passthrough, and D-ID Echo accepts WAV, MP3 or PCM16. Check the provider’s current docs for sample rate, chunk size and transport.

Bring your own LLM and TTS, and let Spatius draw the avatar on the device. Open the Playground, or See pricing.

Give your agent a face that responds.

Start building