Fully on-device, real-time multimodal AI, with features similar to GPT-Live.
parlor_v2_compressed.mp4
Research preview. This is an early experiment. Expect rough edges and bugs.
I'm self-hosting a totally free voice AI on my home server to help people learn speaking English. It has hundreds of monthly active users, and I spent weeks improving it until it's almost on par with ChatGPT Advanced Voice.
Then, my heart broke when OpenAI released GPT-Live, which is miles ahead of ChatGPT Advanced Voice. GPT-Live is so good that I use it almost every day. It made my app feel obsolete. I either need to redirect my users to GPT-Live, or update the app to match it. I chose to try, with the additional constraint that it had to run 100% locally on my MacBook M3 Pro.
My first attempt was to fine-tune Gemma 4 12B to behave like a full-duplex model. Something like grafting a decision tick + speech head to the model. It failed after multiple trials. For now, I think a classic cascade system is still better. We just need to wait until a benevolent frontier AI company releases a full-duplex model that's on par with GPT-Live.
This software is developed with strong assistance from Claude and with humans leading the ideas, testing, and debugging. We say this openly because it shaped how the project was built. If you are not happy with AI-developed code, this software is not for you. This disclosure is inspired by ds4.
Everything above this section is human-written, everything from here down is AI-generated.
Browser (mic + camera)
│
│ WebSocket (audio PCM + JPEG frames)
▼
FastAPI server
├── smart-turn-v3 (~20ms) → did you finish your thought?
├── Gemma 4 E4B via llama.cpp (QAT q4_0) → hears + sees, streams the reply
│ └── action head (same model, JSON) → timers · modes · research
├── Kokoro TTS (MLX on Mac, ONNX on Linux) → speaks sentence-by-sentence
└── background reasoner (optional) → frontier model for web research
│
│ WebSocket (transcript + streamed audio chunks)
▼
Browser (playback + transcript)
- Background research (optional). "Find the best pizza in Rome right now" hands the task to a frontier model on any OpenAI-compatible endpoint (OpenRouter + web search by default) while the conversation continues; the answer is spoken when it arrives. Off unless
REASONER_API_KEYis set — without it Parlor stays fully on-device. - Hands-free turn-taking. Browser-side Silero VAD with a ~200ms silence cutoff — no push-to-talk. Pipecat's smart-turn-v3 then judges whether you actually finished your thought: mid-thought pauses are held silently until you continue, or answered once you stay quiet.
- Everything streams. The reply opens with a transcript of what was heard (on screen immediately — committing to it first measurably improves accuracy), then is spoken sentence-by-sentence while still generating. Your camera frame and speech are pushed through llama.cpp's prompt cache while you're still talking, so even long questions start answering almost instantly.
- Barge-in. Speak over the AI to interrupt it; generation is aborted server-side. Echo is handled without the browser's echo canceller — or just wear headphones.
- Actions never ride the speech. Timers, mode switches, and research requests are decided by a separate grammar-forced JSON request over the same prompt cache, hidden under TTS playback. The spoken reply stays pure speech, so control markup can never leak into TTS and a promise made aloud is always kept (recall 1.0 vs 0.955 for in-band tags —
benchmarks/archbench.py). - Timers. "Set a timer for three minutes for the pasta" — the server owns the clock (a turn-based model can't ring into silence —
benchmarks/timerprobe.py), the model announces the ring out loud in any mode, and a countdown chip with a cancel button tracks it. - Live translation mode. "Translate everything I say into English" turns Parlor into a consecutive interpreter: each utterance rendered after a short silence, no conversational replies, in any language Gemma understands — until you say "stop translating" or hit the stop chip.
- Just-listen mode. "Just listen for a while, I want to think out loud" makes it a silent scribe: every utterance transcribed on screen, nothing spoken back, until you address it again.
- A sense of time. The model is told how much quiet preceded a turn, how long research took, and when the session started — so "how long was I gone?" gets a real answer.
- Python 3.12+
- llama.cpp (
brew install llama.cppon macOS; other platforms: install guide). Needs build b9503 (June 2026) or newer, b9512 forMODEL=12b— older builds lack Gemma 4 audio or crash loading its mmproj - macOS with Apple Silicon, or Linux with a supported GPU
- ~6 GB free RAM for the default E4B model (
MODEL=e2bfits in ~4 GB)
git clone https://github.com/fikrikarim/parlor.git
cd parlor
# Install uv and llama.cpp if you don't have them
curl -LsSf https://astral.sh/uv/install.sh | sh
brew install llama.cpp
uv sync
uv run parlorOpen http://localhost:8000, grant camera and microphone access, and start talking.
Models are downloaded automatically on first run (~5.7 GB for Gemma 4 E4B QAT + its multimodal projector, plus TTS models).
Set these in your shell or a .env at the repo root. The common ones:
| Variable | Default | Description |
|---|---|---|
MODEL |
e4b |
Gemma 4 size: e2b (fastest), e4b (better answers, ~1.8x e2b latency), 12b (needs ~8GB) |
PORT |
8000 |
Server port |
REASONER_API_KEY |
(unset — research off) | API key enabling background research |
REASONER_BASE_URL |
https://openrouter.ai/api/v1 |
Any OpenAI-compatible chat endpoint |
REASONER_MODEL |
openai/gpt-5.6-luna |
Model the endpoint should run |
The full list — local model paths, llama.cpp tuning, TTS backend, test hooks — is in docs/configuration.md.
Measured with MODEL=e2b from end of utterance to first audio heard (add ~200ms of VAD silence detection on top; the default E4B is roughly 1.8x these numbers, with noticeably better answers). The camera frame and the speech itself are prefilled while you're still speaking, and the reply opens with the transcript line (its decode time is included below — the price of accurate transcripts):
| Turn | First audio | Turn complete |
|---|---|---|
| Short question (~2s speech) | ~0.7s | ~1.3s |
| Short question + camera | ~0.8s | ~1.3s |
| Long question (~9s speech), streamed | ~1.3-1.4s | ~2.1-2.3s |
| Long question + camera | ~1.5s | ~2.3s |
Reproduce with the end-to-end benchmark (real spoken audio, synthesized locally). Run it before and after a change to see the impact:
uv run parlor # terminal 1
uv run python benchmarks/bench.py --label before --out benchmarks/results/before.json # terminal 2
# ...make changes, restart the server...
uv run python benchmarks/bench.py --label after --out benchmarks/results/after.json
uv run python benchmarks/compare.py benchmarks/results/before.json benchmarks/results/after.jsonAn end-to-end suite spawns the real server (llama.cpp, TTS, turn detector) and drives it over WebSocket with synthesized speech — including degraded audio (clipped word endings, noise, other voices) that reproduces live-mic failure modes:
uv run pytest # ~1 minute + model loadSet PARLOR_TEST_URL=ws://localhost:8000/ws to run it against an
already-running server. Browser-only behavior (echo at speaker volume, VAD
feel, multilingual speech) still needs a live mic and ears.
src/parlor/
├── server.py # FastAPI app + per-connection conversation loop
├── llama.py # llama-server lifecycle + chat API client
├── pipeline.py # Streaming turn pipeline (decode → sentences → TTS)
├── actions.py # Action decider (grammar-forced JSON: timers, modes, research)
├── reasoner.py # Background research delegation (OpenAI-compatible)
├── modes.py # Session modes (conversation, translate, listen)
├── turn_detector.py # smart-turn-v3 end-of-turn classifier
├── tts.py # Platform-aware TTS (MLX on Mac, ONNX on Linux)
└── web/ # Frontend (markup, styles, app logic: VAD, camera, playback)
tests/ # End-to-end test suite (uv run pytest)
benchmarks/
├── bench.py # End-to-end latency benchmark
├── fixtures.py # Spoken-audio fixtures (synthesized locally)
├── compare.py # Diff two benchmark result files
├── turnbench.py # Turn-detection accuracy benchmark
├── archbench.py # In-band control tags vs decoupled action head
├── camerabench.py # Attach-every-turn vs camera-as-tool-call
└── timerprobe.py # Why the server owns the timer clock
- Gemma 4 by Google DeepMind
- llama.cpp by Georgi Gerganov and contributors
- Kokoro TTS by Hexgrad
- Silero VAD for browser voice activity detection
- smart-turn-v3 end-of-turn detection by Pipecat
