Qwen Scribe: Local Transcription and System-Wide Dictation for Apple Silicon

Overview

Qwen Scribe is a private, local transcription and system-wide dictation tool specifically designed for Apple Silicon Macs. It leverages the Qwen3-ASR model running through mlx-qwen3-asr on the Mac's Metal GPU, ensuring that all audio processing and transcript text remain on the local device without the need for accounts, API keys, or cloud services.

Core Features

Qwen Scribe provides two primary modes of operation: file-based transcription and real-time system-wide dictation.

Local File Transcription

Users can transcribe audio and video files via a focused local web interface. Key capabilities include:

  • Model Selection: Choice between Qwen3-ASR 1.7B for higher accuracy or 0.6B for increased speed.
  • Transcription Tools: Automatic language detection, optional forced language settings, vocabulary hints, and word-level timestamps.
  • Export Options: Support for SRT export and an automatic local transcript history with controls to reopen or delete entries.
  • Data Privacy: All processing happens on-device; the local API includes host and browser-origin checks for security.

System-Wide Dictation

Qwen Scribe enables dictation into any active text field on macOS. By holding the right Command key, speaking, and then releasing the key, the tool transcribes the audio locally and pastes the resulting text at the cursor position.

This workflow is supported by a non-focus-stealing HUD that indicates the current state: Listening, Transcribing, success, or failure.

Technical Architecture and Requirements

System Requirements

To run Qwen Scribe, the following environment is required:

  • Hardware: Apple Silicon Mac.
  • OS: macOS 14 or newer (required by MLX releases).
  • Software: Native Python 3.12 or newer and ffmpeg (for non-WAV audio and video).
  • Memory: The 1.7B model requires approximately 3.4 GB of unified memory, while the 0.6B model requires approximately 1.2 GB.

Architecture

Qwen Scribe utilizes a FastAPI job queue to serialize GPU jobs, preventing memory contention. The data flow for dictation is as follows:

  1. Trigger: Right Command key press detected by a native macOS helper.
  2. Processing: Audio is captured as a temporary WAV file and sent to the FastAPI queue.
  3. Inference: MLX / Qwen3-ASR processes the audio on the Metal GPU.
  4. Insertion: The helper restores the clipboard and uses Command-V to paste the transcript into the focused application.

Installation and Build Process

Qwen Scribe is currently in beta (v0.1.0-beta.1). Because binaries are not yet Developer ID signed or notarized, users must build the app from source:

make app
open "dist/Qwen Scribe.app"

Alternatively, the server can be run via the terminal using ./run.sh and accessed at http://127.0.0.1:8990.

Permissions

For system-wide dictation to function, the following permissions must be granted in System Settings $\rightarrow$ Privacy & Security:

  • Microphone: To record audio while the modifier key is held.
  • Input Monitoring: To detect the right Command key while other apps are active.
  • Accessibility: To paste the finished text into the focused field.

Development and Optimization

Developers can optimize decoding latency by performing a local 8-bit conversion of the model weights using the provided quantize_8bit.py script. The project is licensed under the Apache-2.0 license and is an independent community project not affiliated with Alibaba Cloud, the Qwen team, or Apple Inc.

Community Insights

While Qwen Scribe offers a high degree of privacy, community discussions on Hacker News highlight a few considerations regarding performance and alternatives:

  • Performance Comparisons: Some users noted that performance may vary compared to other models; one user claimed that while Qwen Scribe might run at 2x real-time on an M1, other solutions like Parakeet on fluidaudio can reach 60x.
  • Privacy Use Cases: The local nature of the tool is highlighted as a significant advantage for sensitive professions, such as medical practitioners, who need to avoid sending patient data to cloud-based AI scribes.
  • Alternative Tools: Users pointed to other Mac-native alternatives such as Voice Baseline as options for those seeking different implementations.

Sources