Repolode
Submit
fikrikarim

fikrikarim/parlor

Fully on-device real-time voice and camera AI assistant built on Gemma 4 and llama.cpp, running locally on Apple Silicon or Linux.

Open-source alternative to:ChatGPT Advanced Voice

View on GitHub
Stars
2,112
Forks
281
Open issues
1
Last commit
3 months ago

What is it?

Parlor is a self-hosted, real-time multimodal assistant that listens through the microphone and sees through the camera, then answers out loud. The whole pipeline runs on the user's machine: Gemma 4 E4B served by llama.cpp handles speech and vision, Kokoro handles text-to-speech, and Pipecat's smart-turn-v3 decides when the user has finished speaking. The browser front end captures audio and JPEG frames over WebSocket, and a FastAPI server runs the conversation loop.

The project focuses on low latency and natural conversation. Replies start with a transcript and are spoken sentence by sentence while still generating, and speaking over the assistant interrupts it. Timers, mode switches such as live translation and silent note-taking, and an optional background research step that delegates web questions to an OpenAI-compatible model are handled by a separate action request. Research is off unless an API key is configured.

The author describes it as a research preview built as a local alternative to GPT-Live-style voice assistants, and notes that the code was developed with substantial AI assistance.

Key features

  • 01Runs speech recognition, vision and reply generation on-device with Gemma 4 via llama.cpp.
  • 02Hands-free turn-taking with Silero VAD and smart-turn-v3 end-of-turn detection.
  • 03Streams transcripts and spoken replies sentence by sentence, with barge-in support.
  • 04Supports timers with a countdown and cancel control.
  • 05Offers live translation and just-listen modes.
  • 06Optional background research through any OpenAI-compatible endpoint.

Use cases

  • Practicing spoken English with a private, free voice tutor on a home server or laptop.
  • Thinking out loud while the assistant transcribes notes without speaking back.
  • Consecutive interpretation of conversations in any language Gemma understands.
  • Experimenting with local multimodal assistants on Apple Silicon Macs.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/fikrikarim/parlor.git
  2. 2

    Enter the project directory

    cd parlor
  3. 3

    Install llama.cpp (macOS)

    brew install llama.cpp
  4. 4

    Install dependencies and start the server, then open http://localhost:8000

    uv sync && uv run parlor

Things to know

  • Research preview with rough edges and bugs.
  • Requires Apple Silicon macOS or a supported Linux GPU, and needs llama.cpp build b9503 or newer.
  • Needs about 6 GB free RAM for the default model and a ~5.7 GB model download on first run.

In collections

Run LLMs locallyOpen-source Mac appsAI video & audio tools

Topics

#on-device-ai#local-llm#voice-assistant#multimodal#speech-recognition#text-to-speech#gemma#llama-cpp#apple-silicon#kokoro

FAQ

What is parlor?
Fully on-device real-time voice and camera AI assistant built on Gemma 4 and llama.cpp, running locally on Apple Silicon or Linux. Parlor is a self-hosted, real-time multimodal assistant that listens through the microphone and sees through the camera, then answers out loud. The whole pipeline runs on the user's machine: Gemma 4 E4B served by llama.cpp handles speech and vision, Kokoro handles text-to-speech, and Pipecat's smart-turn-v3 decides when the user has finished speaking. The browser front end captures audio and JPEG frames over WebSocket, and a FastAPI server runs the conversation loop.
Is parlor free to use? What license does it have?
parlor is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is parlor still maintained?
Updated within 6 months. The last commit to parlor was on Aug 3, 2026.
What can I use parlor for?
Practicing spoken English with a private, free voice tutor on a home server or laptop. Thinking out loud while the assistant transcribes notes without speaking back. Consecutive interpretation of conversations in any language Gemma understands. Experimenting with local multimodal assistants on Apple Silicon Macs.
How do I install parlor?
Clone the repository: git clone https://github.com/fikrikarim/parlor.git Enter the project directory: cd parlor Install llama.cpp (macOS): brew install llama.cpp Install dependencies and start the server, then open http://localhost:8000: uv sync && uv run parlor
Where can I get parlor?
The source code and installation instructions are on GitHub: https://github.com/fikrikarim/parlor

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects