martin

Privacy-first meeting transcriber and dictation tool for Linux. Records microphone and system audio, transcribes locally with Whisper, stores results in SQLite. No cloud, no internet — your audio never leaves your machine.

Two modes: Record meetings with dual audio capture (mic + system), or Dictate with real-time speech-to-text that shows words as you speak.

Screenshots

Features

Recording mode

• Dual audio capture — records microphone + system audio (browser, Zoom, Meet) via PipeWire
• Pending recordings — recordings are tracked in the database, survive app restarts, and can be transcribed or deleted from a list
• Non-blocking stop — stopping long recordings runs mixing in the background, UI stays responsive

Dictation mode

• Real-time transcription — speaks and sees text appear live as you talk
• Mic-only capture — uses microphone directly, no system audio needed
• Three-state indicator — Listening / Processing / Paused, so you know what the app is doing
• VU meter — visual confirmation that the microphone is picking up your voice
• Provisional vs stable text — recently transcribed text shown in gray italic (may still be refined), confirmed text in normal style
• Automatic paragraphs — long pauses (>2.5s) become paragraph breaks in the output
• Voice formatting commands — say "novo parágrafo", "nova linha", "ponto final", "vírgula", "ponto de interrogação", "ponto de exclamação", "abre aspas", "fecha aspas" to insert formatting
• Smart capitalization — sentences are capitalized automatically and punctuation spacing is normalized
• Silence-aware — skips whisper passes during silence to save CPU
• Continuous auto-save — partial transcript is persisted every 5 seconds; nothing is lost if the app crashes
• Audio preserved — raw mic audio is saved as a WAV alongside the transcript so you can reprocess later if needed

General

• Auto-download model — Whisper model downloads automatically on first use, with progress bar
• Local transcription — Whisper runs on your machine, no internet needed
• Bilingual UI — Portuguese and English, follows system locale (also used for transcription language)
• Transcription history — browse, view, copy, and delete past transcriptions
• AI summary — summarize transcriptions with key points via Claude Code CLI, with copy support
• Privacy first — audio files deleted after transcription, data stays in local SQLite

Requirements

• Linux with PipeWire (Ubuntu 22.10+, Fedora 36+, etc.)
• pw-record and wpctl (included with PipeWire)
• Rust 1.70+
• Node.js 18+

System dependencies

# Ubuntu/Debian
sudo apt install -y libwebkit2gtk-4.1-dev libgtk-3-dev libasound2-dev libpulse-dev \
  build-essential libssl-dev libayatana-appindicator3-dev librsvg2-dev pipewire

Install

From .deb (recommended)

Download the latest .deb from GitHub Releases and install:

sudo apt install ./martin_0.1.0_amd64.deb

On first use, the app automatically downloads the Whisper model (~466MB). Internet required only for this one-time download.

From source

git clone https://github.com/vagnerzampieri/martin.git
cd martin
npm install
cargo install tauri-cli --version "^2"
cargo tauri build

The binary will be in src-tauri/target/release/martin.

Usage

Recording a meeting

1. Open martin and select the Record tab
2. Click Start Recording / Iniciar Gravação
3. Have your meeting — both your mic and system audio are captured
4. Click Stop Recording / Parar Gravação — recording appears in the pending list
5. Click Transcribe / Transcrever on the pending item — wait for local transcription
6. Done. Text is saved, audio is deleted.

You can record multiple times before transcribing — each recording is tracked separately. Close and reopen the app; your pending recordings are still there.

Dictating text

1. Select the Dictation / Ditado tab
2. Click Start Dictation / Iniciar Ditado
3. Speak — text appears in real time as you talk
4. Click Stop Dictation / Parar Ditado — the full text is saved to history

Voice commands (Portuguese)

Speak these phrases mid-dictation to insert punctuation or formatting. Matching is case-insensitive and accent-insensitive — novo paragrafo works the same as novo parágrafo, and Whisper dropping a diacritic does not break the command.

The dictation also reacts to silence:

• A pause of ~2 seconds flips the UI state to Pausa detectada (no other effect — speak again and it resumes)
• A pause of ~5 seconds commits the current text and starts a new paragraph automatically — useful when you stop to think between thoughts

How audio capture works

Recording mode

Records two audio sources simultaneously:

• Microphone — captured via cpal (ALSA backend)
• System audio — captured via pw-record targeting the default PipeWire sink

When you stop recording, both WAV files are mixed into a single file. If PipeWire is not available or the system audio is corrupt, martin falls back to microphone-only recording. The mixed file is saved as a pending recording and can be transcribed later.

Dictation mode

Captures microphone audio via cpal and transcribes in real time:

• Audio streams into a shared buffer at the device's native sample rate
• A second thread emits the audio level (dictation://level) and session state (dictation://state, one of listening/processing/paused) for UI feedback
• Every ~500ms, the transcription loop drains new samples, measures RMS, and skips the whisper pass if the chunk is silent
• When there is enough new audio, the full accumulated buffer is converted to mono 16kHz and sent to Whisper for re-transcription (this is what keeps the output accurate — Whisper self-corrects with more context)
• The output passes through a text-normalization pipeline (voice command substitution → punctuation spacing → whitespace collapse → sentence capitalization) before being emitted
• A pause longer than ~2.5s commits the current text as a segment and inserts a paragraph break (\n\n) before the next segment
• When the buffer exceeds 120 seconds, the current segment is committed (reusing the last transcription) and a new buffer starts
• The mic audio is written to a WAV file in real time, so the audio is available for reprocessing after the session ends
• The partial transcription row is created on Start Dictation and updated every ~5s during the session; on Stop Dictation the same row is finalized

Whisper Models

Download with: ./scripts/download-model.sh <model>

Development

cargo tauri dev          # Full app dev mode with hot reload
cargo test               # Run Rust tests
npm run check            # Svelte/TypeScript type checking
cargo fmt                # Format Rust code
cargo clippy             # Lint Rust code

Architecture

src-tauri/src/
├── lib.rs              # Tauri commands, app state
├── audio/
│   ├── capture.rs      # Mic (cpal) + system audio (pw-record)
│   ├── mix.rs          # WAV mixing (mic + system → single file)
│   └── wav_writer.rs   # Thread-safe WAV writer (used by recorder and dictation)
├── db/
│   └── store.rs        # SQLite CRUD for transcriptions + pending recordings
├── dictation.rs        # Real-time mic capture + transcription loop + level/state emitter
├── model.rs            # Auto-download Whisper model with progress events
├── postprocess.rs      # Pure text normalization: voice commands, spacing, capitalization
├── summarize.rs        # Claude CLI integration for AI summaries
├── vad.rs              # Pure RMS-based silence detection helpers
└── transcribe/
    ├── whisper.rs      # Whisper transcription + WAV loading + resampling
    └── job.rs          # Finalize worker + cancel/progress orchestration

src/
├── lib/
│   ├── i18n.js         # Locale detection + translations (pt/en)
│   ├── format.js       # Shared date/duration formatting
│   ├── appBusy.js      # Cross-tab "app busy" store
│   ├── Recorder.svelte # Recording controls + pending recordings list
│   ├── Dictation.svelte # Real-time dictation with state/level/provisional UI
│   ├── VuMeter.svelte  # Audio-level meter component
│   ├── FinalizingProgress.svelte # Progress overlay for finalize phase
│   ├── ModelDownload.svelte # Model download progress overlay
│   ├── History.svelte  # Transcription list
│   └── TranscriptionView.svelte  # View + copy + summarize
└── routes/
    └── +page.svelte    # Main page (three tabs: Record / Dictation / History)

License

GPLv3