01 A voice assistant that stays close
Say it.
Think locally.
Hear it back.
MurmurX turns your voice into text, asks a local Qwen model for a concise answer, then speaks the answer back—without building a cloud account around your conversations.
TRY IT NOWOpen the voice lab↘
Your typed conversation stays in this tab.
VOICE / LOCAL AI
01LISTEN
02TRANSCRIBE
03REASON
04SPEAK
02 Interactive browser lab
Talk to the
pipeline.
TRY IT NOW
Load Qwen Lite for real on-device model replies, or open the instant guided preview to explore the interface first.
SESSION / EPHEMERALCHOOSE ENGINE
01Listen
02Transcribe
03Reason
04Speak
Try asking
Microphone transcription uses your browser's speech service and may not be fully local. Typed messages and Qwen inference remain in this tab.
03 The real desktop pipeline
Four small steps.
One private loop.
Listen until silence.
PyAudio records 16 kHz mono sound. WebRTC VAD notices when you stop speaking, so no button press is needed.
WHISPER
Turn sound into words.
Whisper.cpp runs on the computer and writes the transcript. Audio does not need to be uploaded to the MurmurX app.
Q2.53B
Ask a local model.
Qwen 2.5 3B receives a short system instruction plus the latest conversation turns, then returns a concise answer.
Hear the answer.
The operating system reads the answer aloud. Then the loop is ready for the next question.
04 Demo versus desktop
No hidden
pretending.
Easy to try
Browser microphone transcription, Qwen 2.5 0.5B through WebLLM, and browser speech synthesis. The guided preview is clearly labelled and is not an AI model.
The full local build
WebRTC voice activity detection, Whisper.cpp transcription, Qwen 2.5 3B through Ollama, short conversation memory, and macOS spoken output.