MDizzleZA/transcriber

Local, offline speech-to-text for meetings (English + Afrikaans) via Whisper/faster-whisper. No audio leaves your machine. # Meeting Transcriber Local, offline speech-to-text for English and Afrikaans meetings. Uses OpenAI Whisper via faster-whisper. No data leaves your machine. ## Quick Start 1. Run `setup.bat` once to install dependencies (~300 MB) 2. Run `transcriber.bat` to open the GUI 3. Select your audio/video file and click **Transcribe** ## Supported Formats mp3, wav, m4a, ogg, flac, mp4, mkv, webm, aac, wma ## Model Guide | Model | Quality | Speed | RAM needed | |----------|-----------|--------|------------| | turbo | Very good | Fast | ~2 GB | | large-v3 | Best | Slow | ~4 GB | | medium | Good | Medium | ~2 GB | | small | OK | Fast | ~1 GB | | base | Basic | Fast | ~500 MB | **Recommendation:** Start with `turbo`. Use `large-v3` for critical meetings. ## Language Tips - Set language to **Afrikaans** or **English** explicitly for best accuracy - **Auto-detect** works well when the language is clear - Code-switching (mixing Afrikaans/English) works better with `large-v3` ## First Run Note The first transcription downloads the model from Hugging Face (~1-3 GB depending on model). Subsequent runs use the cached model and are instant to load. ## GPU Acceleration If you have an NVIDIA GPU, select **GPU (CUDA)** in the app and set precision to `float16` for 3-5x faster transcription.