Ollama for ASR. Run open-source speech recognition models on your machine or private cloud. Freedom from proprietary APIs, full privacy, no quality compromise.
Pluggable and extensible — Mix and match transcription, alignment, diarization, and PII detection models. Swap components without breaking your pipeline. Completely open source and free.
Drop-in integration — OpenAI and ElevenLabs compatible APIs mean you can point your existing code at Dalston and it just works. Need more power? The native Dalston API unlocks advanced functionality like multi-engine routing, pipeline customization, and detailed engine metadata.
Cheap to run — make dev is free. A 1-hour podcast on a spot GPU costs cents. A 24/7 ElevenLabs/OpenAI-compatible API on AWS runs around $87/month all-in. See the cost estimator.
Transcribe audio files or live streams with speaker diarization, word-level timestamps, and GPU acceleration. Run it on your own infrastructure.
# One-command local transcription (M57 zero-config bootstrap)
# - auto-starts local server if missing
# - auto-ensures default model (distil-small)
DALSTON_SECURITY_MODE=none dalston transcribe tests/audio/test_merged.wav --format json{
"text": "Hello, welcome to the meeting...",
"segments": [
{"speaker": "SPEAKER_01", "start": 0.0, "end": 2.5, "text": "Hello, welcome to the meeting."},
{"speaker": "SPEAKER_02", "start": 2.8, "end": 5.1, "text": "Thanks for having me."}
]
}git clone https://github.com/ssarunic/dalston.git && cd dalston
make dev # full local stack on DockerFor zero-Docker single-process mode or AWS deployment, see the guides.
- Batch & Real-time — File uploads or WebSocket streaming
- Speaker Diarization — Identify who said what
- Word Timestamps — Precise timing for every word
- OpenAI & ElevenLabs Compatible — Drop-in replacement for existing integrations
- Modular Engines — Faster Whisper, NeMo Parakeet, Voxtral, Pyannote, and more
- Private by Default — Runs entirely on your infrastructure, no data leaves your environment
Start here:
- Quickstart — first transcript in 5 minutes
- Pick your deployment — laptop / spot GPU / 24/7 AWS
- All guides → — engines, real-time, cost, principles
Engineering reference: