Three Ways to Build With Liforma: Full Stack, Bring Your Own LLM, or Add Avatars to Your Voice Agent

There are three fundamentally different ways to build an interactive avatar with Liforma. You can use the complete Liforma conversational stack, keep Liforma's speech and animation layers but bring your own intelligence, or keep your existing realtime voice agent and use Liforma only to give it a visual character.

These are not merely different API options. They represent three different boundaries between what Liforma owns and what your application owns.

The three Liforma integration layers

Mode	Liforma provides	You provide	Best when
Liforma Live	STT + intelligence + TTS + speech-to-animation	Your experience configuration	You want the simplest complete solution
Liforma Relay	STT + TTS + speech-to-animation	Your LLM / agent / dialogue logic	You want complete control over what the character says
Liforma Motion	Speech-to-animation	Your complete speech-to-speech stack	You already have a voice agent and just need the visual layer

Liforma also offers Liforma Speak for text-to-speech plus animation when your application already knows exactly what the character should say. But for open-ended conversational applications, Live, Relay and Motion are the three most important architectural choices.

1. Liforma Live: use the complete stack

Liforma Live is the default and simplest way to build.

The flow is:

User speech → Liforma STT → Liforma intelligence → Liforma TTS → Liforma animation

Your application does not have to select a speech recogniser, call an LLM, stream text into a voice provider, synchronize generated speech with an avatar or manage the conversational pipeline between those systems.

You create an Experience, define the characters, instructions, world, state and tools, and Liforma handles the live conversational path.

When should you use Liforma Live?

Use Live when you want:

• the fastest route from an idea to a working interactive character;
• one vendor for STT, intelligence, voice and animation;
• no additional AI-provider integration or billing;
• a predictable low operating cost;
• the Liforma authoring model for characters, scenes, state and tools; and
• good conversational quality without optimising every model individually.

Liforma's current published price for Live is $0.010 per generated speech minute. That includes STT, intelligence, TTS and animation.

There is no extra LLM bill or TTS-provider bill for the developer. That simplicity is a major part of the product.

The main advantage: fewer moving parts

Voice agents can become surprisingly complicated when each stage comes from a different provider.

A typical assembled stack might require:

• microphone capture and turn detection;
• speech-to-text;
• conversation history;
• LLM inference;
• tools and state updates;
• text-to-speech;
• streaming audio;
• interruption handling;
• speech-to-animation;
• avatar playback; and
• error handling across all of those boundaries.

With Live, that complexity stays behind the Experience interface. Liforma's Quick Start describes the complete managed path: session creation, microphone capture, speech recognition, AI response, TTS and avatar rendering are handled by the runtime.

Is the included LLM and TTS always the absolute best available?

Not necessarily — and that is intentional.

Liforma optimises the managed stack for a combination of quality, latency and very low cost. For many training, education, game and website-assistant experiences, the included models are more than good enough and the simplicity is worth far more than a marginal improvement from a more expensive model.

But there are applications where a developer wants a particular frontier model, a particular voice provider or much richer emotional speech. That is why the other two integration layers exist.

2. Liforma Relay: keep the speech stack, bring your own intelligence

Liforma Relay moves one boundary.

The flow becomes:

User speech → Liforma STT → your intelligence → Liforma TTS → Liforma animation

Liforma still handles the microphone-to-text side of the conversation and the text-to-spoken-avatar side. Your application decides exactly what the character should say next.

Why bring your own LLM?

For many developers, “the LLM” is not just a model choice. It is the application.

Your intelligence layer may contain:

• a specific frontier or specialist LLM;
• your own system prompts;
• proprietary business rules;
• RAG over your own data;
• your own tool-calling framework;
• a game dialogue system;
• a tutoring engine;
• a deterministic rules engine;
• multiple model calls or agent steps; or
• logic that is not based on an LLM at all.

Relay lets you keep all of that while avoiding the need to separately implement STT, TTS and avatar animation.

Example: a regulated customer-service agent

Imagine a financial-services company with an existing response engine.

It may have carefully controlled prompts, policy retrieval, compliance checks, tool calls and a final response validator. Replacing that logic merely to add a visual character would make little sense.

With Relay, Liforma can transcribe the customer, pass the text to the company's existing intelligence layer, then speak and animate the approved response.

Example: a game with its own dialogue engine

A game developer may already have explicit world state, quests, inventory, faction relationships and NPC logic.

The game can decide what the character says while Liforma handles listening, voice generation and animation. The character therefore fits into the game's existing architecture instead of forcing the game to fit into a generic agent architecture.

How is Relay implemented?

Liforma exposes a custom conversation processor for this pattern. The SDK can continue to own turn orchestration while your function owns the response.

For example:

const experience = await Experience.startSession({
  experienceId: 'exp_YOUR_EXPERIENCE_ID',
  mode: 'conversation',
  speechInputMode: 'auto',
  conversationProcessor: async ({ text, conversation, signal }) => {
    const response = await yourAgent.respond({
      text,
      conversation,
      signal
    });

    return response.text;
  }
});

The processor can also stream text incrementally rather than waiting for one complete response. See the Custom Conversation Processor guide.

What does Relay cost?

Liforma currently publishes Relay at $0.009 per generated speech minute.

But unlike Live, that is not your complete AI cost. You also pay whatever your intelligence layer costs.

That might be nearly zero for a small local model, or materially more if every turn uses a large frontier model with extensive context and tool calls.

The benefit is control, not necessarily lower total cost.

3. Liforma Motion: keep your complete voice agent and add the visual layer

Liforma Motion moves the boundary again.

The flow is:

Your speech-to-speech stack → generated audio → Liforma speech-to-animation

Your existing voice system owns listening, turn detection, reasoning and speech generation. Liforma receives the resulting audio and turns it into timed avatar speech and facial animation.

This is the right model when your voice agent is already working and the missing piece is simply: give it a face.

Modern speech-to-speech systems are increasingly complete products

By 2026, several providers offer sophisticated realtime voice-agent stacks of their own.

Examples include:

• ElevenLabs ElevenAgents — a realtime voice-agent platform with highly expressive voice technology, tool use and configurable LLMs.
• Google Gemini Live API — realtime bidirectional interaction with Gemini using audio, video and text input and native audio output.
• OpenAI Realtime API — native speech-to-speech conversation with realtime models, tool calling and low-latency audio input/output.
• Deepgram Voice Agent and other voice-agent systems that can produce streaming audio responses.

If you are already using one of those platforms successfully, replacing it with another STT → LLM → TTS pipeline solely to add an avatar is unnecessary.

Liforma sits after the voice agent

In Motion, Liforma does not need to understand how your speech was generated.

The audio may come from ElevenLabs, OpenAI, Gemini, Deepgram, LiveKit or your own system. Liforma accepts streamed PCM or other supported audio, handles playback timing and drives the avatar mouth and facial animation.

The Bring Your Own Voice guide documents the generic path, with dedicated integrations for the major realtime providers.

ElevenLabs ElevenAgents + Liforma

ElevenLabs is a particularly natural example because many teams already use ElevenAgents for voice conversation and specifically chose it for speech quality.

You can keep that entire stack and use Liforma only for the visual character.

Liforma provides a dedicated connectElevenLabsAgent() helper that receives ElevenLabs agent audio, forwards it to the Liforma character and handles the audio/transcript path needed for lip-sync. See the ElevenLabs → Liforma integration guide.

Conceptually:

ElevenAgents listens → ElevenAgents reasons → ElevenLabs speaks → Liforma animates

Gemini Live API + Liforma

Google's current product name is the Gemini Live API. It supports realtime bidirectional interaction with audio, video and text, with native spoken audio responses.

Google currently positions Gemini 3.8 Live as its default low-latency model for many realtime voice agent experiences.

Liforma's Google bridge keeps Gemini Live as the speech-to-speech brain and streams its generated audio into the avatar. See the Gemini Live → Liforma guide.

Conceptually:

Gemini Live listens → Gemini reasons → Gemini speaks → Liforma animates

OpenAI Realtime API + Liforma

OpenAI's product is the Realtime API. It provides native speech-to-speech conversation rather than requiring developers to separately chain speech recognition, a text model and TTS.

Liforma provides helpers for both WebSocket and WebRTC Realtime integrations. The OpenAI session owns the realtime conversation and generated speech; Liforma receives that speech and turns it into the visual character. See the OpenAI Realtime → Liforma guide.

Conceptually:

OpenAI Realtime listens → OpenAI reasons → OpenAI speaks → Liforma animates

Why choose Motion instead of Liforma's included voice?

There are several strong reasons.

1. You already have a working voice agent

This is the most obvious case.

If your application already has prompts, tools, RAG, analytics, turn-taking and a production voice agent, the cost of replacing all of that may be much higher than the cost of adding animation.

Motion lets the avatar become another presentation layer over the system you already trust.

2. You want frontier-quality speech

The highest-end realtime voice models can produce richer prosody, emotion, pacing and conversational nuance than a TTS system optimised primarily around low cost.

ElevenLabs, for example, now offers an Expressive mode in ElevenAgents built around Eleven v3 Conversational, which adapts tone and emotional delivery to conversational context.

OpenAI's Realtime models and Gemini Live similarly generate native spoken responses rather than reducing every turn to text before speech synthesis.

For a dramatic character, premium concierge, companion or emotionally rich experience, that extra expressiveness may be worth paying for.

3. You want frontier intelligence and voice in one realtime system

Speech-to-speech models can preserve information that a traditional STT → text LLM → TTS chain may lose: tone of voice, hesitation, emotion, emphasis and other acoustic cues.

That can matter when the quality of the spoken interaction itself is central to the product.

4. You already have startup or cloud credits

This can change the economics dramatically.

A startup may have substantial promotional credits with an AI or cloud provider. If those credits cover a realtime voice stack, your effective STT/LLM/TTS cost may temporarily be close to zero.

In that situation, there is little reason to pay Liforma to reproduce capabilities that another provider is already giving you under your credit allowance. Use the credits for the expensive frontier intelligence and speech, then use Liforma Motion for the visual layer.

The important caveat is that those economics change once the credits expire, so the production cost should still be modelled at normal provider pricing.

5. You need a provider-specific capability

Your existing voice stack may provide something Liforma's managed pipeline is not intended to replace: a particular cloned voice, provider-specific compliance setup, sophisticated telephony, multimodal video input or another specialised feature.

Motion allows you to keep that capability and add animation independently.

Is frontier-quality TTS always necessary?

No.

There is a danger in assuming the most expressive available voice is automatically the right voice for every application.

For many experiences, Liforma's included TTS is already entirely adequate:

• workplace training;
• product assistants;
• educational tutors;
• information guides;
• many game NPCs;
• role-play scenarios; and
• high-volume applications where cost matters.

In these cases the quality of the scenario, character behaviour, latency and response content can be more important than obtaining the last increment of vocal expressiveness.

But there are also applications where voice performance is a central part of the character. Motion means you do not have to choose between Liforma's avatar system and a premium external voice stack.

Motion also avoids forcing speech-to-speech systems back into text

This is an important technical point.

If your existing system is genuinely native speech-to-speech, inserting a separate text-only intelligence layer just to drive an avatar can throw away some of the reason you chose the system in the first place.

With Motion, the generated audio can remain the authoritative output. Liforma consumes the audio stream directly for playback and animation.

When a transcript is available, it can also be supplied to improve alignment, but the voice provider still owns the spoken performance.

What does Motion cost?

Liforma currently publishes Motion at $0.008 per generated speech minute.

That buys the avatar animation layer, not the external voice-agent service.

Your true total cost is therefore:

your voice-agent cost + $0.008/min for Liforma Motion

That can be more expensive than Liforma Live at normal provider prices. Motion is primarily about control, reuse of an existing stack and access to capabilities or quality levels that matter to your application.

Live vs Relay vs Motion: who owns each layer?

Layer	Live	Relay	Motion
Microphone / speech input	Liforma	Liforma	Your voice stack
Speech-to-text	Liforma	Liforma	Your voice stack
Conversation intelligence	Liforma	You	You / your provider
Text-to-speech / native audio output	Liforma	Liforma	Your voice stack
Speech playback	Liforma	Liforma	Liforma receives your audio
Speech-to-animation	Liforma	Liforma	Liforma
Avatar / scene rendering	Liforma	Liforma	Liforma

Which mode should you choose?

Choose Live if...

• you want the simplest integration;
• you do not already have a conversational stack;
• you want predictable end-to-end pricing;
• Liforma's included LLM and voice quality meet your needs; and
• you would rather spend engineering time on the experience than on AI infrastructure.

Choose Relay if...

• you want your own prompts, LLMs, tools or business logic;
• you already have an agent or dialogue engine that consumes and returns text;
• you want Liforma to continue handling STT and voice generation;
• you need deterministic or specialist processing between transcription and speech; or
• your intelligence layer is a core piece of proprietary technology.

Choose Motion if...

• you already use ElevenAgents, Gemini Live, OpenAI Realtime or another voice-agent stack;
• you want premium or especially expressive external speech;
• you want native speech-to-speech rather than a chained text pipeline;
• startup credits make a frontier voice stack economically attractive;
• you need provider-specific voice, telephony or multimodal capabilities; or
• you simply want to add an animated character to a system that already works.

The modes are a migration path, not three separate products

An important consequence of this architecture is that your decision does not have to be permanent.

A team can begin with Live to launch quickly. Later, it may move the intelligence layer into its own backend and use Relay. If it eventually adopts a specialist realtime speech provider, it can use Motion while retaining the Liforma characters and experience layer.

Or the direction can go the other way: a prototype built around an expensive third-party voice stack can move to Live when credits expire or when unit economics become more important.

The avatar should not force the rest of your AI architecture.

The visual layer should be composable

This is the larger design principle.

There is no single “best” LLM, speech recogniser or voice model for every application. The frontier will keep moving, pricing will keep changing and developers will have different provider relationships.

A visual character platform should therefore work at more than one level of the stack.

If you want everything managed, use everything. If your intelligence is special, keep it. If your voice stack is special, keep that too.

Liforma's job at the narrowest layer is simply to take the speech your AI already produces and turn it into a character that can exist inside an interactive visual experience.