Sber Whisper Desktop

Cross-platform desktop voice-to-text app (Windows + macOS) with:

• Tray/top-bar background mode
• Global hold-to-talk hotkey (Ctrl+G on Windows, Cmd+G on macOS)
• Siri-style popup with streaming partials and final transcription
• Automatic clipboard copy on final result
• Local-only processing with GigaAM v3_e2e_rnnt
• Bundled Python sidecar in installer (target machine does not need Python)

Stack

• Desktop shell: Tauri v2 + React + TypeScript
• ASR sidecar: Python (python/asr_service.py)
• Model: gigaam.load_model("v3_e2e_rnnt")
• GigaAM source: GitHub salute-developers/GigaAM (PyPI 0.1.0 is too old for v3_e2e_rnnt)

Prerequisites

• Node.js 20+
• Rust toolchain (stable)
• Python 3.10+
• Windows only: Visual Studio 2022 with Desktop development with C++ and Windows 10/11 SDK
• On Windows with CUDA: NVIDIA drivers + CUDA-compatible PyTorch build

Only the build machine needs Python. End users install via .exe/.dmg and do not need Python.

Quick Start

make setup
make dev

Local Debug Without Installer (Windows)

If you want to debug the app locally without using installer and without localhost errors:

npm run run-local-win

This builds a local debug exe with embedded frontend and runs: src-tauri/target/debug/sber-whisper.exe

By default, this command rebuilds the sidecar every run, so dependency updates are applied.

For local GPU debug sidecar build:

set SIDECAR_VARIANT=gpu && npm run run-local-win

Build Artifacts

make release builds for the current host OS:

• Windows host -> NSIS .exe
• macOS host -> .dmg

Output artifacts are copied to artifacts/releases.

The release build first creates a standalone ASR sidecar binary (sber-whisper-sidecar) and embeds it into the installer resources, so you can share just the installer file.

Windows installer is CPU-first by default, so resulting setup.exe stays smaller and stable for sharing.

Manual targets:

make release-win
make release-mac

Optional Windows GPU Build

GPU variant is separate from installer build.

Build GPU portable package (no installer):

make release-win-gpu

Result:

• folder: artifacts/releases/sber-whisper-gpu-portable-win-x64
• zip: artifacts/releases/sber-whisper-gpu-portable-win-x64.zip

Run GPU portable build:

1. Unzip package.
2. Keep sber-whisper.exe and sber-whisper-sidecar/ in the same folder.
3. Start sber-whisper.exe.

Build only GPU sidecar (without packaging):

make gpu-sidecar-win

Build On macOS

Run these commands on a Mac host:

xcode-select --install
brew install node python@3.11 rustup-init
rustup-init -y

Then from the project directory:

make setup
make release-mac

Resulting .dmg is copied to artifacts/releases.

Runtime Behavior

• App starts hidden in tray/top-bar.
• Hold global hotkey to record; release to transcribe.
• If retriggered while busy: current job is cancelled, new one starts.
• Popup closes by timeout, close click, or any keypress while focused.
• Audio is written to temp file only during job and immediately deleted after transcription.

Settings

Settings window supports:

• Hotkey
• Popup auto-hide timeout
• Model keepalive timeout (minutes before ASR model unloads from RAM/VRAM when idle)
• Launch at login toggle

Settings are stored in app config directory as app_settings.json.

Logs

Local rotating logs (no telemetry):

• app.log (Rust app)
• asr.log (Python sidecar)

Both are stored under app config logs folder.

Commands

Python sidecar command IPC (stdin JSON lines):

• init
• start_recording
• stop_and_transcribe
• cancel_current
• set_config
• healthcheck
• shutdown

Python sidecar event IPC (stdout JSON lines):

• ready
• recording_started
• recording_stopped
• partial_transcript
• final_transcript
• job_cancelled
• error
• metrics

Tests

make test

Troubleshooting

If popup shows Model 'v3_e2e_rnnt' not found, your sidecar was built with old GigaAM. Rebuild sidecar and rerun debug/release build:

powershell -ExecutionPolicy Bypass -File scripts/build-sidecar.ps1 -Platform windows

If you want forced GPU sidecar build:

powershell -ExecutionPolicy Bypass -File scripts/build-sidecar.ps1 -Platform windows -Variant gpu

License

This project is licensed under the MIT License. See LICENSE.

Third-party notice:

• GigaAM is MIT-licensed: https://github.com/salute-developers/GigaAM/blob/main/LICENSE