Verba

Cross-platform transcription tool: audio files and entire folders are transcribed with faster-whisper, optionally cleaned up and translated by an LLM, refined in an editor with an audio timeline, exported as PDF, and made searchable through a local semantic search. Plus an OpenAI-compatible transcription API for external programs. Everything runs locally — no cloud required.

Feature set

• Transcription (faster-whisper, GPU with CPU fallback), live progress, fair queue for multiple users
• Editor with waveform, segment editing, speakers, range re-transcription and audio cutting
• AI processing (cleanup, translation into nearly any language) via a local LLM (managed llama.cpp server) or an OpenAI-compatible endpoint
• Transcript types with their own prompts (interview, song, role play, …) and type-aware PDF export (single or compiled)
• Semantic search across all transcripts (hybrid: full text + embeddings, timestamp jump marks, AI answers with sources)
• Public API: POST /v1/audio/transcriptions in the OpenAI wire format, API key management in the settings
• PWA (mobile first, de/en/ru, offline-capable interface), Material 3 design

Screenshots

Transcript project	Semantic search

Installation

Ready-made packages are on the releases page:

Package	For	Installation
Verba-Setup-<version>.exe	Windows desktop	double-click → wizard → start menu entry
Verba-<version>-x86_64.AppImage	Linux desktop	make executable, double-click
verba-server-<version>.zip	Linux server (headless)	unpack, sudo ./deploy/install.sh

On first start, the in-app first-run setup checks the system and installs missing components (ffmpeg, Whisper, PDF export, search) automatically with live progress. An existing ffmpeg installation is used; otherwise a static build is set up.

From source

Prerequisite: Python 3.11+ (python.org or your package manager).

Windows: double-click start.bat — Linux/macOS:

./start.sh

The first start automatically creates a virtual environment and installs the core dependencies; the application then opens in the browser (http://127.0.0.1:8710).

Server Operation

./start.sh --server --port 8710

Binds to 0.0.0.0, opens no browser, and is reachable via IP, domain or reverse proxy. Health check: GET /health. All URLs are relative — running behind nginx/Caddy/Traefik works without special configuration; the WebSocket (/ws) must be forwarded.

Ready-made templates live under deploy/: systemd unit, nginx and Caddy example configuration (incl. WebSocket and upload limits) and install.sh for the complete server setup. For running the public API on a network, create a key in the settings under "API".

More options: python run.py --help (--host, --port, --no-browser, --data-dir).

Directories

Path	Contents
data/	settings (settings.json), database, logs (with rotation), models, tools
workspaces/	one folder per transcript project (audio, transcripts, exports)

Both paths are configurable. Installed desktop builds store data per user (%LOCALAPPDATA%\Verba or ~/.local/share/verba).

Development

python -m pip install -r requirements/core.txt -r requirements/dev.txt
python -m pytest tests/ -q
python -m ruff check backend/ tests/ run.py

Details on architecture, packaging and the release pipeline: DEVELOPER_GUIDE.md. Guidelines for code agents: CLAUDE.md and AGENTS.md. The user guide (de/en/ru) is available in the app under Settings → Documentation and lives under docs/user/.