Transcriber

Free, fully local, dual-stream live transcription for meetings and calls — no cloud, no API key, nothing uploaded anywhere. Speech-to-text runs entirely on your CPU with faster-whisper (Whisper), so your audio never leaves the machine.

It captures two kinds of audio at once and labels them separately:

• your microphone(s) (including Bluetooth headset mics), and
• the system / meeting audio — whatever is actually playing out of your speakers or headset (the other people in a Zoom/Teams/Meet/Preply call, a video, etc.).

Both streams are transcribed in real time into one timestamped, colour-coded transcript, so you can see who-said-what across the mic and the call.

Features

• Dual-stream capture — tick any combination of mics and speakers/outputs; each source gets its own label and colour in the transcript. "🔴 Track All" ticks everything and starts in one click.
• Timestamped notes — type a note during the session and it's inserted into the timeline with a timestamp, marked as higher-signal than the raw ASR text.
• Drag-and-drop attachments — drop images, PDFs, or any file onto the window (or click Add File…). They're copied into the session's assets/ folder, and you can reference one from a note by typing / and picking it from the autocomplete (/screenshot.png).
• Auto-exported markdown — every session writes a sessions/session_<timestamp>/transcript.md (plus an assets/ folder). The file leads with a short explainer aimed at AI/note-building tools: the timeline is raw, unedited ASR meant as source material for building a clean notes document — Note lines are flagged as higher-signal, and attachments are linked inline where they're relevant. Feed the whole folder to an AI to turn a messy live session into tidy notes.
• Model picker — choose Whisper size (tiny … large-v3) for a speed/accuracy trade-off. The first use of a size downloads it once (with a live progress bar) and caches it locally for later.

Requirements

• Python 3.12
• A few Python packages (see requirements.txt): faster-whisper, numpy, ttkbootstrap, tkinterdnd2, and — on Windows only — sounddevice.

Linux (validated)

Linux captures audio through PipeWire/PulseAudio using the pactl/parec CLI tools, which is why it sees every real source: physical mics, Bluetooth mics, and the monitor of any output (speakers/headset) for system audio.

# system packages (once)
sudo apt install python3.12-venv python3-tk pulseaudio-utils -y

# project setup
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

python transcribe.py

Windows

Windows capture uses sounddevice (PortAudio, WASAPI). System audio is captured by opening an output device in loopback mode; microphones are captured directly. See Known limitations below.

Option A — prebuilt executable: grab Transcriber.exe from the latest GitHub Actions "Build Windows executable" run (Artifacts section) and run it. No Python install needed.

Option B — from source:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python transcribe.py

Usage

1. Launch the app. It lists your microphones and your speakers/outputs.
2. Tick the sources you want (or hit 🔴 Track All), pick a Whisper model, and press ▶ Start. The first use of a given model size downloads it once.
3. Talk / play the call. Lines appear in the transcript, labelled and coloured per source.
4. Type notes in the Notes box as you go; reference a dropped-in file with /.
5. Press ■ Stop (or 📄 Export Notes) — a transcript.md and assets/ folder are written under sessions/. Hand that folder to an AI to build clean notes.

Privacy

Whisper runs 100% locally on your CPU. No audio, transcript, or note is ever uploaded — there's no API key and no network transcription. The only network access is a one-time download of the chosen Whisper model from Hugging Face, which is then cached locally.

Your sessions live in sessions/, which is gitignored and never leaves the machine.

Known limitations

• Windows audio backend is unverified on real hardware. The Windows capture path is implemented against the documented sounddevice/PortAudio (WASAPI) API and the CI build confirms it imports and bundles cleanly, but it has not yet been run on a real Windows machine to confirm mic and system-audio (loopback) capture actually work end to end. If something's off on Windows, please open an issue. The Linux path is validated.
• No audio resampling: if a device rejects 16 kHz, the app falls back to the device's native sample rate, which can affect transcription quality in that (rare) case.

License

MIT — see LICENSE.