Free · MIT · Windows 10/11
Dictation on your hardware, not their servers.
Hold a hotkey, talk, release. Whisper transcribes on your own CPU or GPU, a local model polishes the text, and it lands at your cursor about a second later. No subscription, no account, no telemetry.
The privacy ledger
Network calls per dictation:
by default. architectural, not measured.
The dashboard keeps a local audit of everything that touches the network. Out of the box that list is empty: audio, transcripts, embeddings, and everything Echo Flow learns about you live in a SQLite file on your disk. The only cloud paths are opt-in, need your own API key, and are off until you flip them.
How it works
Three moves. One second.
Hold Ctrl+Shift
A global push-to-talk hotkey that works in any focused window: email, Slack, your editor, a form.
Talk
Whisper runs locally, sized to your hardware (tiny through large-v3-turbo). CPU works; an NVIDIA GPU gets used automatically.
Release
A small local model strips the "um"s, fixes punctuation and casing, and pastes polished text at your cursor.
no Ollama installed? you still get raw Whisper text. offline either way.
The part the others can't copy
It learns you. Locally. Forever.
Cloud dictation learns everyone a little. Echo Flow learns one person deeply, because all of your corrections stay on your disk where it can keep using them.
Fix it once, it sticks
Corrections you make feed a local pattern miner. Casing sticks after a single edit, recurring mishearings get learned (with a phonetic gate so paraphrases don't), and given enough history it can clean up with no LLM in the loop at all.
note to vec node2vec phrase mishearing, learned
Calibrate to your voice
Read eight known sentences aloud and Echo Flow compares what Whisper heard against ground truth, pinning the words it fumbled straight into your dictionary. Accent adaptation at the text layer, no model fine-tuning.
Humanize anything
Paste AI-written text, get plain prose back: dead openers cut, hedging stripped, long dashes gone. Sentences carrying numbers, citations, or quotes are never sent to the model, so your facts survive byte-for-byte.
Accuracy went from 0.81 to 0.88. 🔒 figures kept exact
Voice actions, allowlist-only
"computer, open spotify" launches apps from a list you configured. URLs are restricted to http, https, and mailto. Nothing in the action layer deletes, sends, or pays, no matter what proposed the action.
$ "run this script" -> no such path exists
Compared
Echo Flow vs. the subscription apps
| Echo Flow | Typical cloud dictation app | |
|---|---|---|
| Price | free · MIT | $10 to $30 / month |
| Your audio | stays on device | uploaded every time |
| Works offline | yes | no |
| Account required | none | sign-up |
| Learns your corrections | locally, forever | limited, cloud-side |
| Auditable | MIT source + privacy ledger | a policy page |
Install
Two ways in.
Installer
Download the setup exe, run it, and hold Ctrl+Shift when the green mic appears in your tray. First launch fetches the Whisper model once; after that it works fully offline. SHA256 files sit next to every download.
From source
What CI tests, and the way to go if you want to hack on it. Python 3.11+.
Recommended: Ollama + ollama pull qwen2.5:3b-instruct-q4_K_M for local cleanup.
Questions people actually ask
FAQ
Is my audio really never uploaded?
Audio is never uploaded, full stop: it stays in memory for the length of the
utterance and is dropped. Your text can reach a cloud API only through
paths you switch on yourself, each needing your own API key: Prompt-Engineering
mode, the teacher loop, cleanup.allow_cloud_cleanup, the
verify-escalation pass, and the humanizer's cloud option. Watch the third one in
particular, because unlike a deliberate keystroke it applies to every dictation.
The code is MIT, and the dashboard's privacy page shows which of these are live
right now so you can verify instead of trusting.
Do I need a GPU?
No. Whisper model size is picked automatically for your hardware and runs fine on CPU. If an NVIDIA GPU is present it gets used, which mostly buys you headroom for the larger, more accurate models.
What does it cost?
Nothing. It is MIT-licensed and fully local. If you deliberately opt in to a cloud model with your own key, you pay that provider; otherwise there is nothing to pay anyone.
How is this different from Windows built-in Win+H?
Win+H is fine for short bursts. It does not clean up filler, learn your jargon and casing, adapt style per app, or give you push-to-talk that pastes anywhere. Echo Flow does all of that, on your hardware, with an audit trail.
Languages other than English?
Sixteen languages plus auto-detect. English currently gets the most of the learning machinery; the rest transcribe and clean up fine.