Local speech-to-text with speaker diarization. No cloud, no API keys, 3x faster than Whisper.
You want to transcribe meetings, podcasts, or voice notes — but:
- OpenAI Whisper API sends audio to the cloud (privacy risk, costs money)
- Whisper.cpp is fast but requires complex setup and manual compilation
- Proprietary tools lock you into subscriptions
WhisperBridge gives you enterprise-grade local transcription with a single pip install and one command.
- 🎙️ Transcribe any audio format — mp3, wav, m4a, ogg, webm, flac
- 👥 Speaker diarization — knows who's talking (turns, segments)
- 🌐 Multilingual — 99 languages with auto-detection
- ⚡ 3x faster than OpenAI Whisper — CTranslate2 optimized, runs on CPU or GPU
- 📊 Word-level timestamps — precise text-audio alignment
- 🔧 Hotword boosting — improve accuracy on technical terms, names, product names
- 📝 Multiple output formats — SRT subtitles, VTT, JSON, plain text
pip install whisperbridgeOr from source:
git clone https://github.com/AmSach/WhisperBridge.git
cd WhisperBridge
pip install -e .Requirements:
- Python 3.8+
- For GPU acceleration: CUDA 11.8+ (auto-detected)
# Transcribe a single file
whisperbridge transcribe meeting.mp3
# Speaker diarization (who said what)
whisperbridge transcribe interview.wav --diarize
# Specific language, faster model
whisperbridge transcribe lecture.m4a --lang en --model small
# Custom hotwords for better accuracy
whisperbridge transcribe tech-talk.mp3 --hotwords "Zo Computer,Nexus,GhostPilot"
# Output as subtitles
whisperbridge transcribe video.webm --format srt --output ./subs/whisperbridge transcribe <file> [options]
Options:
--model [tiny|small|medium|large] Whisper model size (default: medium)
--lang <code> Language code, e.g. en, fr, de (auto-detect if omitted)
--diarize Enable speaker diarization
--hotwords <words> Comma-separated words to boost
--format [txt|srt|vtt|json] Output format (default: txt)
--output <path> Output file/directory
--device [cpu|cuda] Compute device (auto-detect)
whisperbridge batch <folder> [options]
Transcribe all audio files in a folder with the same settings.
whisperbridge serve --port 8080
Start a local HTTP API server for transcription.
curl -X POST -F "audio=@recording.mp3" http://localhost:8080/transcribe
| Model | Speed (GPU) | Speed (CPU) | Accuracy |
|---|---|---|---|
| tiny | 30x real-time | 4x real-time | 85% |
| small | 15x real-time | 2x real-time | 92% |
| medium | 8x real-time | 0.8x real-time | 95% |
| large | 4x real-time | 0.4x real-time | 97% |
Tested on: NVIDIA RTX 3090, AMD Ryzen 5950X, Intel i9-13900K
whisperbridge/
├── whisperbridge/ # Main package
│ ├── __init__.py
│ ├── cli.py # Click-based CLI
│ ├── transcriber.py # Core transcription engine
│ ├── diarizer.py # Speaker diarization (pyannote)
│ └── formats.py # Output format writers
├── tests/
│ ├── test_transcriber.py
│ └── test_cli.py
├── requirements.txt
├── setup.py
└── README.md
Store API keys or model paths in ~/.whisperbridge.yaml:
model: small
device: cuda
output_format: txt
hotwords: []| Feature | WhisperBridge | Whisper.cpp | OpenAI API |
|---|---|---|---|
| Local processing | ✅ | ✅ | ❌ |
| Speaker diarization | ✅ | ❌ | ❌ |
| Hotword boosting | ✅ | ❌ | ❌ |
| Python integration | ✅ | ❌ (C++) | ✅ |
| No API key needed | ✅ | ✅ | ❌ |
| Subtitle formats | ✅ | ❌ | ❌ |
| One-line install | ✅ | ❌ | ✅ |
MIT License — free for personal and commercial use.