This project provides a self-hosted, offline transcription and diarization service with LLM-based summarization. It targets users needing to process audio files, offering a web interface and SSH drop zones for easy integration into workflows. The service leverages Whisper for transcription/diarization, Ollama with Mistral for summarization, and Meilisearch for full-text search, aiming for a turnkey solution.

How It Works

The system utilizes Docker for deployment, bundling transcription, diarization, summarization, and search functionalities. Audio files can be uploaded via a web UI or dropped via SSH. Whisper-diarization handles speaker identification and transcription, while Ollama integrates with Mistral to generate summaries based on a customizable prompt. Meilisearch provides fast indexing and retrieval of transcribed text.

Quick Start & Requirements

• Install/Run: ./start-nobuild.sh (for Docker images) or ./install.sh followed by ./run.sh (for local build).
• Prerequisites: NVIDIA GPU (required for ts-gpu image, ~26GB).
• Resources: 12GB VRAM may be insufficient for both Whisper-diarization and Ollama Mistral simultaneously.
• Docs: install and ts-web walkthrough videos (linked in README).

Highlighted Details

• Turnkey, self-hosted, and offline operation.
• SSH and Web UI for file upload and management.
• LLM summarization via Ollama and Mistral with customizable prompts.
• Full-text search powered by Meilisearch.
• HTML5 web player with time-synced scrubbing and highlighting.

Maintenance & Community

• Developed by transcriptionstream, with contributions acknowledged from MahmoudAshraf97 and jmorganca.
• The README notes it is "example code for example purposes and should not be used in production environments without additional security measures."
• To-do list includes fixing UI errors and adding Meilisearch controls.

Licensing & Compatibility

• The README does not explicitly state a license. The project is presented as community edition.

Limitations & Caveats

• The project is explicitly stated as example code not suitable for production without security hardening.
• Potential CUDA memory issues exist when running diarization and Ollama Mistral concurrently on the same host due to VRAM limitations.
• The web interface has known console errors when summary files are missing.