beautiful-stt

Local, private, and free speech-to-text transcription powered by OpenAI's Whisper model.

Commercial transcription services (Google, AWS, Azure, etc.) require you to:

• Pay per minute of audio
• Upload your audio files to their servers

This is a problem when your audio contains:

• Confidential business information
• Government or legal information
• Medical records
• Personal conversations
• Sensitive client data

Beautiful STT runs 100% locally. Your audio never leaves your machine.

• Fully local processing, no internet required after model download
• Supports multiple audio formats (MP3, WAV, FLAC, OGG, AAC, M4A, MP4)
• Download audio directly from YouTube and Facebook URLs via yt-dlp
• AI-powered summarization via local LLM (llama-completion)
• Dark mode and accessibility support
• Automatic noise reduction
• GPU acceleration (Metal on macOS, CUDA on Windows/Linux)
• Uses from tiny to large-v3 Whisper models depending on available resources

Beautiful STT ships three CLI binaries alongside the app, no manual installation needed:

All processing happens on your machine. yt-dlp only contacts the internet to fetch the requested URL; whisper-cli and llama-completion run fully offline.

1. Download the .dmg file from Releases
2. Open the file and drag the app to Applications
3. Open Terminal and run:

   xattr -cr /Applications/beautiful\ speech\ to\ text.app

4. Now you can open the app normally and enjoy!

Why is this needed? macOS blocks apps that aren't signed with an Apple Developer certificate ($99/year). This is standard for open source software. The command above removes the quarantine flag that macOS adds to downloaded apps.

1. Download the .msi file from Releases
2. Run the installer and follow the prompts
3. if "Windows preogetió su PC" or similar message appears, click on "more information" and "Ejecutar de todas formas"
4. Run next, next, next and finish. Enjoy!

Why is this needed? Windows blocks apps that aren't signed with an Apple Developer certificate. This is standard for open source software. The step above removes the quarantine flag that Windows adds to downloaded apps.

Choose the package format for your distribution:

Debian / Ubuntu

1. Download the .deb file from Releases
2. Install it (replace <version> with the downloaded version):

   sudo dpkg -i Beautiful-STT_<version>_amd64.deb

Fedora / RHEL / openSUSE

1. Download the .rpm file from Releases
2. Install it (replace <version> with the downloaded version):

   sudo rpm -i Beautiful-STT_<version>-1.x86_64.rpm

Universal (AppImage)

1. Download the .AppImage file from Releases
2. Make it executable and run (replace <version> with the downloaded version):

   chmod +x Beautiful-STT_<version>_amd64.AppImage
   ./Beautiful-STT_<version>_amd64.AppImage

• macOS: Apple Silicon (M1/M2/M3) with Metal
• Windows/Linux: NVIDIA GPUs with CUDA

On first launch, the app will download the Whisper/Llama models (400MB ~3 GB depending on the model size). This only happens once for model.

Open source - MIT License

Contributions are welcome! Feel free to open issues or submit pull requests.

• Language selection Whisper supports many languages but the app is currently hardcoded to Spanish. Exposing this as a UI option unlocks the full model capability with no pipeline changes.
• Export results Add export to .txt or .md via tauri-plugin-fs (already a project dependency). Currently only copy-to-clipboard is available.
• Cancellation No way to stop a transcription in progress. A long audio file forces the user to close the app entirely.

• Model download progress Models range from 500 MB to 9 GB and download silently. The first use of any model looks like a crash. A real progress bar is needed.
• Extracted entities in the UI The backend already extracts structured JSON (people, dates, organizations, figures) via Gemma in detailed mode, but this data never reaches the frontend. Displaying it as a dedicated panel would be a meaningful differentiator.
• Session history Persist transcriptions and summaries in localStorage or SQLite (Tauri has a plugin) so work isn't lost on close.

• GPU/CPU fallback n_gpu_layers(99) assumes unlimited VRAM. On low-VRAM hardware the process crashes silently. This needs a configurable limit or at least a graceful CPU fallback.
• Batch processing The pipeline is already reusable. Extending the UI to accept a queue of files is the main work.
• Direct microphone recording The natural use case is recording a meeting or voice note directly in the app, not only loading pre-recorded files.

• Global shortcut / system plugin A system-wide hotkey to start recording from any app, similar to Whisper Flow or SuperWhisper. Tauri supports global shortcuts.
• Timestamps in the transcript Whisper produces per-segment timestamps internally but they aren't surfaced in the UI. Useful for navigating long audio (click text → jump to that minute).

set -gx CXXFLAGS "-mmacosx-version-min=11.0 -std=c++17"
set -gx CFLAGS "-mmacosx-version-min=11.0"
set -gx MACOSX_DEPLOYMENT_TARGET 11.0
set -gx CMAKE_OSX_DEPLOYMENT_TARGET 11.0
CMAKE_GENERATOR="Unix Makefiles" CMAKE_POLICY_VERSION_MINIMUM=3.5 pnpm tauri build

APPIMAGE_EXTRACT_AND_RUN=1 pnpm tauri build

On Debian/Ubuntu, the AppImage bundling step needs librsvg2-dev (provides librsvg-2.0.pc for the linuxdeploy GTK plugin). Install it first: sudo apt install librsvg2-dev.

pnpm tauri build

# first time (download whisper-cli zip)
.\scripts\build-windows-cuda.ps1

# second time  (without download zip)
.\scripts\build-windows-cuda.ps1 -SkipDownload

# diferent version of whisper.cpp
.\scripts\build-windows-cuda.ps1 -WhisperVersion v1.8.4 -SkipDownload

Models (Whisper, VAD and LLM) are not stored next to the binary, inside an AppImage that location is read-only and writing fails with Permission denied (os error 13). Instead they go to the user's data directory (dirs::data_dir() + beautiful-stt):

Layout inside that folder:

beautiful-stt/
├── ggml-<model>.bin          # Whisper models
├── ggml-silero-v5.1.2.bin    # VAD model
└── llm_models/
    └── <model>.gguf          # summarization LLMs

To free disk space or force a re-download, delete the relevant file(s); they are fetched again on next use.