Mubbles

An Egui wrapper around HuggingFace's Candle Whisper model, which implements Whisper, the OpenAI speech-to-text model.

Can record from microphone and speakers.

Usage

Run the app.

Cuda builds require Nvidia / Cuda drivers.

Huggingface models will be downloaded automatically when transcription starts.

Some things are still a bit broken because I don't use them (this is really a personal project).

PRs welcome but I may not get to them for a few weeks.

Regular usage

Install the app into the default cargo bin directory (probably ~/.cargo/bin):

cargo install --path . --features cuda

Depending on the accelerator you have available, choose which feature -- mkl for intel CPUs or cuda for Nvidia cards with CudaNN installed in the system.

Screenshot

Build on ubuntu

apt install build-essential libssl-dev pkg-config libasound2-dev

Also, if you don't have cuda but do have an intel cpu, use --features mkl.

https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html?operatingsystem=linux&linux-install=apt

Features

• Real-time transcription from microphone or speaker loopback
• Echo cancellation (optional): mute mic while speaker output is active
• Per-device configurable trigger level and audio level monitoring (logarithmic dB scale)
• Switch audio devices without stopping transcription
• File transcription (wav, mp3, flac, ogg, m4a, aac, wma) with silence-based splitting
• AI summarization via Ollama, OpenAI, or any OpenAI-compatible API
  • Streaming output, incremental chunking, configurable prompts
  • Auto-detects local Ollama models and context length
• Statistical summary (most unusual words per section)
• Autotype: types transcript into the foreground application
• Word-level timestamps and alignment
• Model selection (tiny, base, small, medium, large, distil variants)
• CUDA and Intel MKL acceleration
• Transcription history with per-word accuracy
• Monthly log files

Roadmap

1. UI is a bit crowded, simplify
2. implement snippet saving, where you can click on a fragment and hear it / save the audio file.
3. figure out why the accuracy for the first and last word is so bad