Ekko STT

Local inference for Ekko v1 Tiny, a Danish speech recognizer with word timestamps, and Ekko PnC, its optional punctuation and capitalization model. This repository contains the Python runtime and the browser demo.

Run locally

Install uv, then from this checkout:

uv sync --locked --no-editable --extra pnc
uv run --no-sync ekko tiny audio.wav --pnc --format json

To format an existing transcript without running speech recognition:

uv run --no-sync ekko pnc "hej mit navn er emil hvordan går det i dag"

For the ONNX runtime:

uv sync --locked --no-editable --extra tiny-onnx --extra pnc
uv run --no-sync ekko tiny-onnx audio.wav --precision int8 --pnc --format json

The first run downloads model artifacts from Hugging Face and the native parakeet.cpp library when needed. Downloads use immutable revisions and SHA-256 checks in the release manifest. Set EKKO_OFFLINE=1 after prefetching to prevent network access.

The browser demo runs Tiny and PnC locally through WebAssembly. Audio and transcripts stay on the device.

Development

uv sync --locked --extra local
uv run --no-sync python -m unittest discover -s tests -p 'test_*.py'
uv build

The source code is MIT licensed. Tiny and PnC weights have their own terms in their Hugging Face repositories. See third-party notices for runtime dependencies.