Skip to content

About

Offline transcription for macOS. Speaker verification picks one voice out of a noisy classroom. Nothing leaves the machine.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

VoiceNotes

A voice to text note taking application for macOS that runs entirely offline. Speech recognition, grammar correction, summarisation and speaker filtering all happen on the machine. No recording ever leaves the device.

The problem it was built for

Recording a lecture in a classroom gives you a file in which the teacher is one voice among twenty. Generic transcription turns that into an unusable wall of overlapping speech.

VoiceNotes solves it with speaker verification. Record thirty seconds of the person you actually want, and the filter transcribes only them. Everyone else in the room becomes silence.

What it does

Transcription. Live recording with real time transcription in German and English through Whisper, with automatic language detection. Existing files work too: MP3, MP4, WAV, M4A, FLAC, OGG and WebM, and video files can produce SRT subtitles.

Speaker filtering. A voice profile built with TitaNet, NVIDIA's speaker verification model, chosen because it holds up in noisy environments where simpler embeddings do not.

Custom vocabulary. Whisper reliably mangles domain terms, names and abbreviations. You teach it: enter how a word sounds and how it should be written, and the correction is applied to every future transcription. "crisp er" becomes CRISPR.

Post processing on device. Grammar correction through MLX on Apple Silicon with LanguageTool as the fallback, summarisation and topic detection across sixteen subject categories through Qwen2.5 running locally, and automatic tagging from content.

Organisation. Nested folders, tags, full text search, editable transcripts, and export to Word, Markdown or plain text, either per note or a whole folder at once.

Adaptive learning. Corrections feed back in, so the same mistake gets made less often.

Requirements

macOS, ideally Apple Silicon, Python 3.9 or newer, and ffmpeg.

brew install ffmpeg
./setup_offline.sh    # first run only
./start.sh            # http://localhost:5050

First setup downloads roughly 2 GB of models: Whisper, TitaNet, LanguageTool language data and the MLX model. That happens once, and afterwards the application never needs a network connection again.

Stack

Component Used for
Whisper Speech recognition
TitaNet Speaker verification and voice filtering
MLX with Qwen2.5 Summarisation, topic detection, grammar, on Apple Silicon
LanguageTool Offline grammar fallback
Flask and SocketIO Real time web interface
SQLite Notes, tags, vocabulary

Privacy

Everything is local. Recordings, transcripts and voice profiles live in ~/Documents/voice-notes/, the database is a plain SQLite file, and there is no server to send anything to.

Detailed usage instructions and troubleshooting are in USAGE.md.

License

MIT

About

Offline transcription for macOS. Speaker verification picks one voice out of a noisy classroom. Nothing leaves the machine.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages