README.md

---
title: SpeakoFlow
sdk: static
pinned: false
---

# SpeakoFlow

[SpeakoFlow](https://speakoflow.com) is a free, MIT-licensed voice assistant for the
desktop. Local speech to text, spoken answers, translation, and dictation cleanup, on
Windows, macOS and Linux. Transcription runs on your own machine.

This organisation publishes the models behind it. Everything here is Apache-2.0 and runs
locally.

**speakoflow-mini-0.8b-GGUF** is a 0.8B dictation cleanup model that runs on a desktop CPU.
It applies the correction the speaker actually made and returns everything else byte for
byte. On a held-out evaluation it scores 70.7% overall, against 65.0% for a hosted frontier
model and 50.0% for doing nothing at all. Five quantisations are published and each one's
cost is measured rather than estimated.

That last number is the interesting one. Grammatical error correction, ASR error correction
and spoken-to-written conversion all score whether a required edit was *made*. None scores
whether an unrequired edit was *withheld*, and that is the failure that ruins a dictation
product. Four of the seven systems measured do worse than leaving the transcript alone.

Built by [Abhishek Barali](https://huggingface.co/AbhishekBarali) ·
[speakoflow.com](https://speakoflow.com) ·
[GitHub](https://github.com/AbhishekBarali/SpeakoFlow)