Tracet

A local desktop app for transcribing audio and video, with speaker identification, word-level confidence scoring, and an editable transcript. Everything runs on your machine: no cloud APIs.

Features

• High-quality transcription with whisperX (faster-whisper + phoneme alignment + integrated speaker diarization)
• Speaker identification via pyannote: give each detected voice a custom name once and it propagates through the transcript and exports
• Word-level confidence highlighting: three-tier visual scale (yellow → red dashed underline) makes likely errors easy to spot
• Per-segment audio playback: click ▶ on any line to hear the original audio for that exact segment
• Inline editing: click any text to fix it, edits persist and appear in exports
• Hardware-aware model selection: settings detect your chip and RAM, then filter Whisper and Ollama model options to what your machine can actually run
• Optional LLM accuracy review via Ollama (any local model)
• .tracet project bundles: save a zip containing the converted audio
  • transcript so you can transcribe on a powerful machine and edit elsewhere
• Three export formats:
  • Markdown (.md): prose for humans
  • Plain text (.txt): minimal
  • AI-friendly Markdown (.ai.md): YAML frontmatter with metadata, numbered segments, structured-but-readable; designed to hand off to any LLM without committing to a specific downstream task

Architecture

flowchart TB
    FE["<b>React + TypeScript + Tailwind CSS frontend</b><br/>Vite, served by Tauri webview"]
    BE["<b>Rust backend (Tauri v2)</b><br/>Pipeline orchestrator<br/>FFmpeg child process for media conversion<br/>Python sidecar manager<br/>Ollama HTTP client (LLM review)<br/>.tracet zip pack/unpack"]
    SC["<b>Python sidecar (sidecar/.venv)</b><br/>whisperX (faster-whisper backend)<br/>wav2vec2 phoneme alignment<br/>pyannote speaker diarization"]

    FE -- "Tauri IPC" --> BE
    BE -- "stdin/stdout JSON" --> SC

Loading

Requirements

• macOS on Apple Silicon (M1/M2/M3 or later)
• Python 3.12 (3.14 is too new for the ML stack at the moment)
• FFmpeg (auto-installed on first run if Homebrew is available)
• Optional: Ollama (https://ollama.ai) for LLM-based accuracy review
• Optional: HuggingFace account and access token for speaker diarization. You will also need to accept the pyannote/speaker-diarization-3.1 license. Tokens are issued at https://huggingface.co/settings/tokens.

Installation

From a release DMG

Download the latest .dmg from GitHub Releases, drag the app to Applications, and launch. On first run a setup screen offers to install the remaining dependencies (FFmpeg, the Python venv, whisperX).

From source

git clone [email protected]:theagitist/tracet.git
cd tracet
npm install
cargo tauri build

The packaged app and .dmg end up in src-tauri/target/release/bundle/.

Development

./dev.sh

dev.sh finds the latest installed Node (working around the broken nvm lazy-load shim some setups have), installs npm deps if missing, then runs cargo tauri dev with hot reload.

Settings

A HuggingFace token is required for speaker diarization. Without it, transcription still works but every utterance is attributed to a single speaker. Set the token via Settings → HuggingFace Token; it persists in localStorage.

For LLM-based review, install Ollama and pull a model:

brew install ollama
ollama pull llama3.1:8b

Then enable LLM review in Settings.

Privacy

Tracet is privacy-first by design. Audio files, transcripts, and saved .tracet projects never leave your machine for normal use. There is no telemetry, no analytics, no crash reporting, and no auto-updater. The app does not phone home.

There are exactly three places where the app makes network requests:

1. Model downloads on first run. whisperX and pyannote download their model weights from HuggingFace and the Torch Hub the first time you transcribe. These are one-way downloads (no audio or transcript data is sent), and once cached at ~/.cache/huggingface/hub/ and ~/.cache/torch/hub/checkpoints/, transcription works fully offline.
2. HuggingFace token use. If you set a HuggingFace token to enable speaker diarization, that token is sent to huggingface.co solely to authenticate the pyannote model download. It is not transmitted to any other endpoint. The token is stored as plain text in your browser's localStorage (under the key tracet:settings) and is readable by anything with access to your user account on this machine. Treat it like a low-stakes credential and do not share the tracet:settings value.
3. Ollama LLM review (optional, local by default). When LLM review is enabled, transcript text is sent to whatever URL the Ollama URL setting points at. The default is http://localhost:11434, so the data stays on this machine. If you change that URL to a non-local host (anything other than localhost, 127.0.0.1, ::1, or 0.0.0.0), every transcript segment Tracet reviews will be sent in plaintext to that host. Settings detects this case, shows a red warning, and requires you to type I understand to apply the change. Keep the URL on localhost to preserve the privacy guarantee.

Things that explicitly do NOT happen:

• No transcript, audio, or project data is ever sent to a server Tracet controls. There is no Tracet backend.
• No analytics SDKs (Sentry, PostHog, Mixpanel, Amplitude, Google Analytics, Segment, etc.) are bundled.
• The Tauri auto-updater is not registered. You update by downloading a new release manually.
• The webview only loads bundled local assets; no third-party scripts are fetched at runtime.

If you want to verify any of this, the only outbound HTTP code paths are src-tauri/src/commands/setup.rs (Ollama health check), and src-tauri/src/commands/review.rs (Ollama generate call); plus the HuggingFace fetches inside the Python sidecar (sidecar/diarize.py via the whisperX and pyannote libraries).

Project layout

.
├── dev.sh                       # Dev launcher
├── index.html                   # Vite entry HTML
├── package.json                 # Frontend deps + scripts
├── src/                         # React + TypeScript frontend
│   ├── App.tsx
│   ├── components/              # Toolbar, TranscriptViewer, etc.
│   ├── hooks/                   # useTranscription, useSegmentPlayer, ...
│   ├── stores/                  # Zustand transcript store
│   ├── lib/                     # Settings persistence
│   └── types/                   # TypeScript mirrors of Rust structs
├── src-tauri/                   # Rust backend
│   ├── Cargo.toml
│   ├── tauri.conf.json
│   ├── capabilities/            # Tauri ACL
│   └── src/
│       ├── lib.rs               # Command registration
│       ├── commands/            # Tauri commands (one file per concern)
│       ├── pipeline/            # End-to-end orchestrator
│       ├── models/              # Transcript / Speaker / config structs
│       └── sidecar/             # Python venv + FFmpeg helpers
├── sidecar/                     # Python sidecar
│   ├── diarize.py               # whisperX entry point
│   └── requirements.txt
└── test-audio/                  # Generated test files for QA
    ├── short-single.wav         # 4.7s, single voice
    ├── short-multi.wav          # 12.7s, two voices alternating
    └── troublesome.wav          # 16.8s, designed to stress whisperX

Acknowledgments

Tracet wires together excellent open-source work:

• whisperX: transcription + alignment
• pyannote-audio: speaker diarization
• faster-whisper: CTranslate2 Whisper backend
• Ollama: local LLM runtime
• Tauri: desktop shell
• FFmpeg

License

MIT. See LICENSE.