Onkith

Most voice assistants send your audio to a server to understand it. Onkith does that work on the device, and only the masked text ever leaves it.

Three steps, all on the device.

1. Listen: voice activity detection runs on the device; nothing is held until speech starts.
2. Transcribe: speech becomes text on the device, and the audio does not move.
3. Gate: a classifier reads the text token by token and masks private spans before the text leaves.

As deputy lead on a team of six, I built the voice stack end to end.

1. The original speech-to-text front end and the streaming model that replaced it, taking word error rate from 6.78% to 3.29% on LibriSpeech test-clean.
2. The LLM call and the on-device text-to-speech that completed the assistant.
3. Both product interfaces, a live terminal client and a two-screen app, and the bilingual site.
4. The int8 ONNX masking model (33 entity types) as a FastAPI service on Google Cloud Run, models baked in so it serves fully offline, with parity verified on ARM64 and x86.