Build

SpeakFree

A local Windows dictation build with hotkeys, private transcription, and a paste-into-any-app workflow.

Status

Building

My role

Product framing, Tauri desktop implementation, local transcription integration, and iterative QA

Best next step

Pick the fastest way to see the work or ask for the walkthrough that matches what you want to understand.

SpeakFree landing page Ask about code walkthrough Ask George about this build

speakfree - Tracked privately while the build is still being tightened through live dictation testing.

What it solves

Most dictation tools either live behind closed workflows, depend on cloud transcription, or hide quality tradeoffs until someone tries to use them for real work. I wanted a local-first build that could be tightened through repeated spoken testing on my own machine.

Why it matters

The current build already handles hotkeys, local transcription, paste, phrase hints, and saved settings.

How AI fits

AI is part of the product and part of the build process here. The transcription layer is local Whisper inference, while the implementation work moved faster through structured iteration, debugging, and QA support. The value comes from tightening the system against real spoken outcomes, not from one-shot generation.

What it is

SpeakFree is a practical local dictation build. Press a hotkey, speak naturally, and get text inserted where the cursor already is without routing the workflow through a cloud transcription product.

What’s working now

A Windows desktop shell built in Tauri
Hold-to-talk and toggle dictation modes with global hotkeys
Local transcription through a faster-whisper worker
Clipboard-safe paste into the active app
Persisted microphone selection and phrase hints for tricky terms
A stronger CPU-first transcription profile based on real QA outcomes

Why it matters

It turns dictation quality into something observable and improvable instead of abstract
It is a useful example of desktop product shaping, model tradeoff decisions, and human-in-the-loop QA
It shows how an internal personal tool can become a cleaner public-facing build without pretending it is already a finished SaaS product

Approach

Build a Windows desktop shell in Tauri with global hotkeys for hold-to-talk and toggle dictation.
Use a local `faster-whisper` pipeline with a stronger CPU profile, clipboard-safe paste, and microphone selection instead of defaulting to a cloud backend.
Run repeated phrase testing on real spoken inputs, then tighten the model profile, cleanup layer, and term-hinting based on observed misses.
Keep the build practical and shareable, with public-safe framing around what is working now rather than pretending it is already a polished commercial product.

Tools

Proof points

The live build now supports local hotkeys, local transcription, clipboard-safe paste, and persisted dictation settings
A stronger `small.en` CPU profile corrected earlier phrase misses that the lighter path was getting wrong
Phrase-hint support was added specifically to improve uncommon names and brand terms such as `Moviefone`
The workflow now includes repeated human dictation checks instead of treating one clean sample as good enough