Mulive

Build local-first, interactive voice and multimodal applications in Python.

Mulive is a lightweight real-time substrate that connects microphone, camera, models, application state, and web UI. Build a tiny voice interaction first, then grow it into a full product without replacing the runtime underneath.

Voice is an interface to your application—not the application itself.

Why Mulive

• Build applications, not just chatbots. Voice can operate a game, tutor, workflow, dashboard, or custom web UI.
• Keep product logic explicit. Voice, video, UI actions, and model responses can be handled with deterministic state transitions.
• Run close to the product. Use direct WebRTC and local-capable STT/TTS; choose cloud models only where they add value.
• Make interactions testable. Test a single utterance or a complete interaction, including interruption, cancellation, and application outcomes.

Choose your path

• Browser quickstart: Run a complete WebRTC voice interaction in the bundled browser client. Start with Quickstart.
• Embedded FastAPI: Add Mulive's browser client and voice transport to an existing application. See Embed in FastAPI.
• Local microphone: Run speech recognition and speech output without a browser. See Local microphone.

Quickstart

A Mulive voice interaction has five parts:

microphone → utterance detection → speech recognition → response → speech output

The response can be a direct stream transformation, normal application code, or a model-backed ConversationHarness. Start with the direct path so each boundary is visible.

This guide uses five terms consistently. An utterance is one piece of user speech. A transcript is its recognized text. A response is the text the application returns. A client event is structured data sent to the browser. Speech output is the generated audio the user hears.

Echo recognized speech

pip install mulive

# Browser audio + local voice stack
python -m mulive.quickstart.web --http --tts-browser

Open http://localhost:9000 and start streaming. localhost is trusted by modern browsers, so microphone access works without a certificate. The browser sends real-time media to Python. It receives transcripts and speech output. Without GROQ_API_KEY, the quickstart echoes each recognized utterance. The core flow is:

from mulive.core.stt import STTConfig

audio = Stream.source(session.audio_input, name="audio")
turn = audio | turn_detector()
transcripts = turn.value | stt(STTConfig(provider="faster_whisper", variant="small"))
final_transcripts = transcripts | filter_items(lambda event: event.is_final)
user_text = final_transcripts | non_empty_text()

user_text | to_user(session=session, role="user", subs=subs)
user_text | to_user(
    session=session, role="assistant", subs=subs,
    tts=tts_config, tts_provider=tts_provider, interrupts=turn.started,
)

turn_detector() groups microphone audio into utterances. stt() converts each utterance into a transcript. The two to_user() connections show the transcript, then return the same text as the response through the display and speech output.

Let an LLM control responses

Use ConversationHarness when responses need conversation history and a language model. The audio, utterance detection, speech recognition, and speech output stay the same. Only the response step changes:

from mulive.apps.conversation_harness import ConversationHarness

async def lookup_order(order_id: str) -> dict:
    """Return order data without changing it."""
    return {"order_id": order_id, "status": "shipped"}

final_transcripts = transcripts | filter_items(lambda event: event.is_final)
user_text = final_transcripts | non_empty_text()

conversation = ConversationHarness(
    system_prompt="You are a helpful voice assistant.",
    llm_timeout_s=30,
    tools=(lookup_order,),
)
# Connect the transcript stream and turn started signal to the conversation harness
conversation.connect(final_transcripts, turn.started)

user_text | to_user(session=session, role="user", subs=subs)
conversation.assistant_text | to_user(
    session=session, role="assistant", subs=subs,
    tts=tts_config, tts_provider=tts_provider, interrupts=turn.started,
)
conversation.client_events.to(client_message_sink(session), subs=subs)
conversation.start()

Install Pydantic AI Slim with its Groq provider, then set its key before running the same browser quickstart:

pip install "mulive[groq]"
export GROQ_API_KEY="your-key"
python -m mulive.quickstart.web --http --tts-browser

ConversationHarness owns conversation history and runs one model call at a time. A new utterance stops current speech output but does not cancel model work. If the user asks another question while a call is running, the harness acknowledges it and waits for the call or its 30-second timeout before answering the newer utterance.

The harness delegates each request to a ModelRunner. The default PydanticAIModelRunner supports registered synchronous and asynchronous Python tool calls. Pass allowed functions through tools. The model can request a tool, receive its return value, and then produce the response. The tool function still defines what data it can read or change.

If that optional package is unavailable when LLM control is requested, Mulive emits a runtime warning and continues with transcript echo.

The web quickstart uses Faster-Whisper small for recognition and sends speech output to browser. Add --tts-local to send speech output to the server's speakers. Override the recognition model with --model-variant, use --stt-provider mlx --model-variant turbo on Apple Silicon, and change LLM model with --llm-model. Change the model timeout with --llm-timeout-s.

Embed in FastAPI

Mount voice into an application you already own:

pip install "mulive[fastapi]"

from fastapi import FastAPI
from mulive import mount_voice

app = FastAPI()
mount_voice(app)

Mulive serves its browser client and voice WebSocket transport from the same application.

Voice model installation options

Mulive includes portable local defaults: Faster-Whisper for speech recognition and Piper for speech output. They run on CPU and download their selected model weights on first use. pip install mulive is enough for the browser quickstart without Groq.

Option	Install	Use
MLX Whisper (Apple Silicon)	pip install "mulive[mlx]"	--stt-provider mlx
Groq responses	pip install "mulive[groq]"	Set GROQ_API_KEY for the web quickstart
In-process Kokoro ONNX	pip install "mulive[kokoro]"	Select kokoro_onnx in an app's TTS configuration; not a web quickstart flag
External Kokoro-FastAPI	pip install "mulive[openai]"	Run Kokoro-FastAPI separately at localhost:8880 (or set KOKORO_FASTAPI_URL)
Gemini models	pip install "mulive[gemini]"	Select Gemini in an app that supports it; the web quickstart uses Groq only

Extras can be combined, for example pip install "mulive[mlx,groq]". TTSConfig() defaults to kokoro_onnx, so install mulive[kokoro] when using that default. The browser quickstart explicitly selects Piper.

Local microphone

The local microphone example uses WebRTC acoustic echo cancellation (AEC) by default. With --tts --allow-interruptions, it uses the speech output as the echo reference. You can speak while the assistant is speaking. Start Kokoro-FastAPI first, then:

pip install "mulive[local-audio,openai]"
python -m mulive.quickstart.mic --stt-provider faster_whisper --model-variant small --tts --allow-interruptions

For MLX on Apple Silicon, add the mlx extra and use --stt-provider mlx. Use --no-aec only if you need to disable echo cancellation.

HTTPS for another device or deployment

Use HTTP only when the browser and Mulive run on the same machine. For a phone, LAN host, public hostname, or an HTTPS app embedding Mulive, serve it over HTTPS. For local device testing, create a trusted development certificate with mkcert:

mkcert -install
mkdir -p ~/.config/mulive/certs
mkcert -key-file ~/.config/mulive/certs/key.pem \
  -cert-file ~/.config/mulive/certs/cert.pem \
  localhost 127.0.0.1 ::1 <your-lan-hostname-or-ip>

python -m mulive.quickstart.web --tts-browser

For production, terminate TLS with the deployment platform or reverse proxy and set MULIVE_SSL_KEYFILE and MULIVE_SSL_CERTFILE when Mulive should serve TLS itself.

What you can build

Example	What it demonstrates	Status
Microphone loop	Local microphone, echo cancellation, utterance detection, speech recognition, and optional speech output	Available
Browser audio	WebRTC audio, transcript echo or Groq response, and browser speech output	Available
Math helper	Browser audio/video, latest-frame vision, and spoken responses	Experimental
Audio todo app	Voice-driven UI state and actions	Planned
Market voice dashboard	Voice control plus live visual data	Planned
Card-cancellation flow	Guarded voice workflow with explicit confirmation	Planned

The application model

Mulive separates the parts that benefit from models from the parts your product must control:

mic / camera / browser UI
            ↓
  transcripts + client events
            ↓
  model calls + explicit app state
            ↓
 speech, UI updates, and safe effects

Use models for perception and conversation. Keep scoring, workflow state, permissions, retries, and external effects in normal application code. That is especially useful for tutoring, games, customer workflows, and any interaction where an answer must be checked before the app moves on.

Model architectures

Today, Mulive's examples use a cascaded pipeline:

audio → utterance detection → speech recognition → model → speech output

The same application model is intended to support direct streaming speech-to-speech models as they are added. The product state and UI should not need to change just because the voice model does.

What is next

Capability	Direction
Embeddable web voice	A small browser client and stable server integration for existing web apps
Backend speech output	Stream generated speech from a Python service without the browser UI
Systematic evaluation	Utterance-level assertions and end-to-end interaction scenarios
Latency explorer	Per-utterance timing from speech end through playback
Streaming voice models	Direct speech-to-speech and hybrid model adapters

Project structure

mulive/      Runtime, browser assets, and runnable quickstarts
mulive/core/ Real-time transport, streams, and media helpers
mulive/apps/ Application helpers
mulive/client/ Browser WebRTC client and UI
mulive/quickstart/ Small runnable applications
tests/       Pipeline, controller, transport, and application tests
docs/        Architecture and design notes

Current status

Mulive is actively evolving. The available quickstarts are useful foundations; the embed API, latency explorer, streaming-model adapters, and polished product demos are intentionally marked as planned rather than presented as shipped.