# Heyra

> Self-hosted speech to text. Send a short WAV, get the words back. English and
> 24 other European languages, found automatically. Audio is never kept.

## Mac app

Heyra for Mac is push-to-talk dictation that runs the same model on the Mac
itself: hold fn, talk, let go, and the words are typed where the cursor is.
Nothing is sent anywhere. Free, MIT, Apple Silicon.

- [Download for Mac](https://github.com/alcun/heyra-desktop/releases/latest/download/Heyra.zip): or `brew install --cask alcun/tap/heyra`
- [Mac app source](https://github.com/alcun/heyra-desktop): MIT
- [Bench](https://heyra.alcun.dev/bench): words wrong per voice model, on 800 public English clips
- [Everything in one file](https://heyra.alcun.dev/llms-full.txt): this, the questions and the bench numbers

## Transcribe

POST /v1/transcribe with a WAV file as the request body: 16-bit PCM, mono or
stereo, 8 to 48 kHz. Without a key, clips may be up to 15 seconds; with
`Authorization: Bearer <key>`, up to 30.

    curl https://<this host>/v1/transcribe --data-binary @clip.wav

Answer: {"text": "Draw me a cat in a hat.", "seconds": 2.4, "ms": 231}

## Limits

Anyone gets a free daily allowance per visitor, within a daily ceiling for
everyone. 429 means the allowance is used up until midnight UTC; 503 means
busy, so retry after the `retry-after` seconds. 400 means the body is not a
WAV Heyra can read, 413 that the clip is too long.

## Status

GET /healthz returns {"ok": true, "busy": n, "free": true, "today": {...}}
once the model is loaded.

## Source

https://github.com/alcun/heyra (MIT). The model is NVIDIA's Parakeet TDT 0.6B
v3 (CC BY 4.0).