Vibe Documentation

Vibe is a privacy-focused, offline transcription tool for audio and video that runs locally on your device. It supports multiple AI models and provides advanced features for transcription and analysis.

Key Capabilities:

• Transcription: Supports almost every language, batch processing, and real-time previews.
• Input Sources: Local audio/video files, system audio, microphone input, and URLs from popular websites (YouTube, Vimeo, Facebook, Twitter, etc.).
• Supported Models: Whisper, Nemotron 3.5, and Parakeet TDT v3.
• Output Formats: SRT, VTT, TXT, HTML, PDF, JSON, and DOCX.
• AI Analysis: Summarize transcripts using the Claude API or local AI via Ollama support.
• Advanced Features: Speaker diarization, stable timestamps mode (VAD-backed), translation to English, and GPU optimization (Vulkan/CoreML) for macOS, Windows, and Linux.

Vibe is a desktop transcription application composed of three primary components:

1. Desktop App (desktop/): A Tauri-based application using a TypeScript/React frontend and a Rust backend. It manages the UI, file management, settings, and analytics. It communicates with the Sona runner via local HTTP.
2. Sona Runner (sona/): A standalone Rust process (using whisper.cpp bindings) responsible for audio transcription, model loading, streaming, and diarization. It is bundled as a Tauri sidecar.
3. FFmpeg Helper: Bundled with macOS and Windows builds to assist with media processing. The path to this binary is passed to Sona via the SONA_FFMPEG_PATH environment variable.

On Windows (and Unix), you can speed up development builds by using the cranelift codegen backend with a nightly Rust toolchain.

1. Install Nightly and Cranelift:

   rustup nightly install
   rustup component add rustc-codegen-cranelift-preview --toolchain nightly

2. Build with Cranelift: Unix:

   CARGO_PROFILE_DEV_CODEGEN_BACKEND="cranelift" cargo +nightly build -Zcodegen-backend

   PowerShell:

   $env:CARGO_PROFILE_DEV_CODEGEN_BACKEND = "cranelift" ; cargo +nightly build -Zcodegen-backend

# Unix
CARGO_PROFILE_DEV_CODEGEN_BACKEND="cranelift" cargo +nightly build -Zcodegen-backend

# Powershell
$env:CARGO_PROFILE_DEV_CODEGEN_BACKEND = "cranelift" ; cargo +nightly build -Zcodegen-backend

Ensure your environment meets the following requirements:

• Windows: Version 8 or greater.
• macOS: Version 13.3 or greater.
• Linux: Tested on ubuntu-22.04+. Note that listening for audio files is currently not supported on Linux.
• Hardware: No special requirements; resource usage can be customized via advanced settings in the main window.

Linux Environment Variable: You may need to set the following environment variable before starting the application:

export WEBKIT_DISABLE_COMPOSITING_MODE=1

Follow these steps to verify the integration is working correctly:

1. Download Binaries: Run uv run scripts/pre_build.py to ensure both sona and sona-diarize are downloaded to desktop/src-tauri/binaries/.
2. Run App: Use pnpm exec tauri dev to build and launch the application.
3. Test UI: Navigate to More Options $\rightarrow$ Diarization and enable the toggle. You should be prompted to download the model (~25MB).
4. Test Transcription: Transcribe audio with diarization enabled; segments should display speaker labels.
5. Test Fallback: Disable diarization; transcription should proceed normally without speaker labels. If the sona-diarize binary is missing, diarization should be silently skipped without breaking the transcription process.

To sign Windows binaries, follow these steps:

1. Install Windows SDK:

   winget install -e --id Microsoft.WindowsSDK.10.0.26100

2. Generate Certificate: Generate a self-signed certificate and export it to a .pfx file:

   openssl req -x509 -newkey rsa:2048 -keyout key.pem -out cert.pem -days 3650 -nodes -subj "/CN=Vibe" -addext "extendedKeyUsage=codeSigning"
   openssl pkcs12 -export -out cert.pfx -inkey key.pem -in cert.pem -password pass:YOUR_PASSWORD

3. Setup GitHub Secrets: Convert the .pfx to base64 and add the output to GitHub Secrets:

   • WINDOWS_CERTIFICATE: The base64 output of the .pfx file.
   • WINDOWS_CERTIFICATE_PASSWORD: The password used during the openssl pkcs12 step.

4. Cleanup:

   rm key.pem cert.pem cert.pfx

base64 -i cert.pfx

If you need to regenerate the DMG background PNG from a source SVG (design/dmg_background.svg), use rsvg-convert to set the dimensions to 1440x652 and sips to set the DPI to 144. This ensures the image is a 2x retina asset suitable for a 720x340 window.

rsvg-convert -w 1440 -h 652 design/dmg_background.svg -o design/dmg_background.png
sips -s dpiWidth 144 -s dpiHeight 144 design/dmg_background.png

Replace react-hot-toast with shadcn's Sonner integration.

1. Install:

   pnpm remove react-hot-toast
   pnpm dlx shadcn@latest add sonner

2. Setup: Add <Toaster /> from sonner to your App.tsx.
3. Usage: Replace toast() calls with the toast() function from sonner. The API for toast.promise() is compatible. For progress-based toasts, use toast.loading() with description updates.

AI-generated contributions are welcome but must meet the same standards as manual contributions:

• Focus: Submit one clear change per PR. Avoid including drive-by refactors, formatting noise, or your own local build scripts/helper tools.
• Context: Ensure your AI agent reads the project documentation and follows existing conventions before generating code.
• Responsibility: You are responsible for reviewing and verifying all AI-generated code.
• Alignment: For large or wide-reaching changes, open an issue first to align with maintainers before investing time.
• Quality: AI-generated code must be clean, tested, and minimal.