Evoo
Open-source voice dictation for macOS that runs entirely on your Mac.
Hold fn, speak, release — clean text appears at your cursor in any app.
New — Evoo AI v1, our own fine-tuned model. Evoo now polishes dictation with a model trained for exactly
this job: Qwen3-0.6B fine-tuned on 43,000 dictations, run after Evoo's rules. On 300 dictations it never saw,
almost twice as many come out exactly right as with the stock model (78 vs 43), in about 0.3 s each on an 8 GB M1.
Model, examples and results on Hugging Face ·
how it was trained
Windows (preview): download Evoo-windows.zip — hold Ctrl + Win to dictate, with voice commands and optional on-device AI polish. See windows/ for what works and what doesn't yet.
Apple Silicon Mac (M1 or newer), macOS 14+. After downloading: unzip, drag Evoo to Applications, open it.
The first time, macOS says it "could not verify" Evoo (it isn't signed with a paid Apple certificate) — open
System Settings › Privacy & Security and click Open Anyway, once. Or skip that step entirely with the
one-line install below. Evoo updates itself after that.
- 100% local: speech recognition and text cleanup run on-device. No accounts, no cloud, no API keys.
- Understands what you meant: "let's meet tomorrow, no, day after tomorrow" → "Let's meet day after tomorrow."
- Works everywhere you can paste: Chrome, Slack, Gmail, Notion, VS Code, Cursor, Terminal, Office…
- Voice commands, chained: hold
fn + ⌃ control and say what to do — "close the tab and switch to Claude", "copy that, new tab, then paste". Evoo runs each step in order (up to six) and never types a command by mistake. Plain fn still dictates.
- Live words: the last few words recognised show above the pill while you speak (Settings can turn it off).
- Light on memory: on 8 GB Macs the optional AI model frees its memory after 10 idle minutes and reloads on the next
fn press.
- Checked updates: before an update is installed, Evoo verifies it is the announced build with an intact signature.
- 24 more European languages (beta) from the same built-in speech model (Parakeet TDT v3) — Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Ukrainian and others: pick one in Settings › Language. They get the speech model's text as heard; Evoo's rules, voice commands and AI polish are English-only for now.
- English, plus a Hinglish add-on (beta): download it from Settings (133 MB) and switch languages on the pill. It uses its own model (Oriserve Whisper-Hindi2Hinglish, Apache-2.0), writes Roman Hinglish ("kal meeting hai, please confirm kar dena"), and handles Hinglish self-corrections; English keeps its own model and is unaffected.
- Speculative transcription: Evoo starts transcribing when you pause, so text is usually ready the instant fn goes up.
- Personal dictionary for names and jargon (Divya, Aarav, Kubernetes).
- Voice shortcuts: say "my email" (alone or mid-sentence) and get the full text you set up.
- Voice editing: "replace Tuesday with Wednesday", "delete the last sentence", "make that a list".
- History: search and copy your recent dictations (menu → History…), kept only on your Mac.
- Welcome tour on first launch: privacy & offline, what you can do, and a step-by-step permissions guide
that ticks off each step live. Afterwards, a small tip by the pill now and then introduces one feature at a time.
- Class Notes (🎓 on the pill, or say "start class notes"): pick the subject and start the class. Evoo
listens and jots down only what matters — definitions, formulas (LaTeX), key numbers, what the professor
stresses — in short fragments, like a sharp classmate, typed into your notes word by word. Edit the notes any
time, even while it's writing. Flag moments (★ Important, ❓ Confusing, 🎯 Exam — ⌘1–3) to steer it. After
class: a one-screen study sheet (summary, key formulas, terms, practice questions), flashcards, and Q&A about
the lecture. No audio is kept. Classes are grouped by subject; say "search note Bayes theorem" to find
something. Export to PDF or Markdown.
- Desktop app: Evoo sits in the Dock; click it for Home (words dictated, speaking speed, time saved, streak, recent dictations), History, Dictionary, Class Notes and Settings. Prefer menu bar only? Settings › Show Evoo in the Dock.
- Instant start: the pill appears the moment you press fn. Optional Keep the microphone ready (Settings) keeps
the mic on between dictations, holding only the last 0.3 s in memory, so recording starts instantly and includes
the moment before the press. Drag the pill anywhere along the bottom or up the left/right side (it stands upright
there), or pick a spot in Settings › Pill position.
- Long dictations stay fast: while you speak, finished sentences are transcribed and, if they need it, polished
in the background — a 42 s filler-heavy dictation is ready ~0.3 s after release on an 8 GB M1.
- Learns how you write (all on your Mac): Evoo keeps what it typed next to what you actually sent. Word swaps
you keep making in an app ("going to" → "gonna") are applied instantly; once you have ~150 edits, it fine-tunes a
small add-on for its AI overnight (MLX LoRA, while plugged in), tests it on edits it hasn't seen, and only uses it
if it's clearly closer to you. On a test persona: plain model 0/28 messages written their way, personal add-on
24/28. Settings › Your writing style / Personal model — view, turn off, or erase any time.
- Voice commands (say the whole command as one dictation; anything else is typed as usual):
- Apps & web: "open Slack", "switch to Chrome", "open github.com", "search Google for …", "ask ChatGPT …",
"YouTube …", "new Google doc", "new email about …" — every installed app, plus your own in Settings → Apps
- Your Mac: "search my Mac for …" (Spotlight), "run shortcut Morning Routine" (Apple Shortcuts), "volume 30",
"volume thirty", "press space bar", "mute", "next song", "dark mode", "take a screenshot", "lock screen"
- Keys: "new tab", "reopen tab", "refresh", "copy", "paste", "save", "scroll down", "press command shift T"
- Windows & buttons: "move this to the left half", "maximize this window", "full screen", "click Send"
- Assistant: "remind me to call Divya tomorrow at 5", "schedule lunch with Raj Friday at 1 PM",
"note: …", "what did I say about the invoice?" (searches by meaning, on-device)
- Audio: "read this aloud", "stop reading", "transcribe a file" (→ text + .srt subtitles)
- Smart cleanup with Evoo AI v1 (new, any Mac): Evoo's own model, fine-tuned for dictation, polishes every
dictation after the rules — fillers and restarts, punctuation, grammar — and never answers what you said.
Settings › Smart cleanup › Model › Evoo AI v1. How it was trained.
- 16 GB Macs — Smart cleanup (local AI): a local model (Qwen3 4B, Apache-2.0) polishes each dictation after
the rules — grammar, messy phrasing, "can u" → "can you" — and powers rewrite by voice: select text, hold fn,
say "make this more formal", "shorten this", "translate to Hindi". Off by default; download from Settings.
- Formats as you speak: "…buy tomorrow bread, eggs, milk" becomes a bulleted list; "first… second… third…"
becomes numbered steps; "new line", "new paragraph", emails. Markdown in Notion/editors/browsers, • bullets
elsewhere, and never a line break in Terminal (a pasted newline could run a command).
- Apache-2.0. Every model it uses is open-weight and free.
Install
One line (Apple Silicon Mac, macOS 14+). Paste into Terminal:
curl -fsSL https://raw.githubusercontent.com/chndr-prksh/evoo/main/install.sh | bash
It downloads the latest release, verifies its checksum, installs /Applications/Evoo.app, and opens it.
Or download Evoo.zip from Releases, unzip, and drag
Evoo to Applications. Evoo isn't signed with a paid Apple Developer ID, so the first time macOS will say it
"could not verify" Evoo: open System Settings › Privacy & Security and click Open Anyway (once).
Updates: every change pushed to main is built and published automatically. Evoo checks every few hours;
when a new build is out, a dot appears by its menu bar icon — choose Install Update and it relaunches on the
new version, keeping your settings and permissions.
How it works
fn down ─▶ AudioRecorder (16 kHz mono)
fn up ─▶ SpeechEngine ─▶ number formatting ─▶ DictationRules ─▶ [LLM, only if needed] ─▶ ⌘V at cursor
Parakeet/Whisper NeMo ITN corrections, Qwen3 via llama.cpp
~110 ms "four hundred"→400 fillers, stutters Hinglish / unresolved
| Stage | Model / library | License | Where it runs |
|---|
| Speech → text (English) | Parakeet TDT 0.6B v3 via FluidAudio | CC-BY-4.0 / Apache-2.0 | Neural Engine |
| Speech → text (Hindi, Hinglish) | Whisper large-v3 turbo via WhisperKit | MIT / MIT | Neural Engine + GPU |
| Numbers ("four hundred ms" → "400 ms") | NeMo inverse text normalization (text-processing-rs) | Apache-2.0 | CPU, < 1 ms |
| Self-corrections, fillers, stutters | DictationRules (built-in, deterministic) | Apache-2.0 | CPU, < 5 ms |
| Lists, line breaks, emails | DictationFormatter: bullets, numbered steps, to-do checklists, "new line", "chandra at gmail.com" — styled per app | Apache-2.0 | CPU, < 1 ms |
| Voice commands | DictationCommands: "capitalize each word …", "all caps …", "lowercase …", "quote … end quote", "… press enter", "undo that" | Apache-2.0 | CPU, < 1 ms |
| Learning from edits | EditWatcher + EditLearner: names you correct join the dictionary; per-app habits (no final full stop, lowercase start) after 2 edits — counts only, never text | Apache-2.0 | Accessibility API |
| Names on screen (context awareness) | ScreenText + ContextVocabulary: names in the focused window (chat header, recipients, text near the cursor) join the dictionary for that dictation — local, never stored | Apache-2.0 | Accessibility API, read while you speak |
| Names & terms ("DeVeo" → "Divya") | PersonalDictionary: sound-alike match, never replaces real English words | Apache-2.0 | CPU, < 1 ms |
| Hinglish romanization, tricky corrections | Qwen3 1.7B (Q4_K_M GGUF) via llama.cpp | Apache-2.0 / MIT | GPU (Metal), ~1 s |
English dictation never waits for an LLM: the rules handle "tomorrow, no, day after tomorrow",
"to Rahul, sorry, to Priya", "3:30, actually make it 4", "um", "we could we could" and "scratch that".
The LLM only runs when a correction cue can't be resolved by rules, or for Hinglish/Hindi.
Using it
| Gesture (default "Hold or double-tap" mode) | Result |
|---|
Hold fn, speak, release | Dictate |
Double-tap fn | Hands-free; tap fn again to finish |
Esc while recording | Cancel |
fn + any other key | Ignored (so Fn+arrows / F-keys keep working) |
Hold fn + Control, speak, release | Command mode: nothing is typed. “Close the tab and switch to Claude” |
Smart commands (preview). An optional 290 MB on-device model (FunctionGemma 270M, fine-tuned on Evoo's 34 tools)
understands commands said your own way in command mode: “it's too bright in here”, “I'm done with Safari”.
On 344 phrasings it never saw it gets 79% fully right and the right tool 93% of the time, in about 0.4 s. Fixed
phrases are always tried first. Download it in Settings; how it is built and scored is in Tools/.
The model is built on Google's FunctionGemma and is subject to the Gemma Terms of Use.
The floating pill at the bottom of the screen shows state; hover it to start hands-free or open Class Notes. Drag it left or right to move it along the bottom of the screen (or Settings › Pill position).
One-time setup: System Settings › Keyboard › "Press 🌐 key to" → Do Nothing. Then grant Evoo
Microphone, Input Monitoring (to see fn) and Accessibility (to paste). Don't run Evoo and
another Fn-based dictation app (e.g. Wispr Flow) at the same time — both will react.
Build
Requires macOS 14+, Apple Silicon, and the Xcode Command Line Tools (full Xcode not required).
swift test # unit tests (gestures, correction rules, prompt, cleaner)
scripts/bundle.sh # builds build/Evoo.app
open build/Evoo.app
First launch downloads the speech model (~0.5 GB). Download the refinement model from Settings (1.1 GB).
Models live in ~/Library/Application Support/Evoo/Models and are pinned + SHA-256 verified.
Permissions across rebuilds: dev builds are ad-hoc signed with a designated requirement pinned to the
bundle ID, so macOS keeps Microphone / Input Monitoring / Accessibility across rebuilds. If permissions ever
look granted but Fn or pasting doesn't work (stale entries from older builds), quit Evoo and run
tccutil reset All app.evoo.Evoo, then relaunch and grant again.
Benchmark without the app
swift build -c release --product evoo-cli
say -o /tmp/a.wav --data-format=LEI16@16000 "let's meet tomorrow, no, day after tomorrow"
.build/release/evoo-cli bench /tmp/a.wav # audio → text with per-stage timings
.build/release/evoo-cli post "send it to rahul, sorry, to priya" # rules + numbers only
.build/release/evoo-cli refine --lang hinglish "कल मिलते हैं no sorry परसों" # LLM only
Debug builds can render every pill state to PNGs: .build/debug/Evoo --snapshot-pill /tmp/pill.
Set EVOO_LLAMA_LOG=1 to see llama.cpp's logs.
Project layout
Sources/EvooCore Pure logic, fully unit-tested: Fn gesture state machine, languages & engine routing,
refinement prompt + output guards, transcript cleanup, model catalog.
Sources/EvooRefine llama.cpp refiner (KV-cached prompt prefix) and verified model downloads.
Sources/EvooSpeech SpeechEngine protocol, Parakeet + Whisper engines, DictationPipeline (ASR → ITN → rules → LLM).
Sources/Evoo The menu-bar app: FnKeyMonitor, AudioRecorder, TextInjector, DictationController,
pill + settings UI, permissions.
Sources/evoo-cli Terminal tool: bench (audio → text timings), post (rules only), refine (LLM only).
Adding a speech engine = one type conforming to SpeechEngine (load / transcribe / unload) plus a case in
ASREngineID. Only open-weight models with OSI-approved or CC-BY licenses are accepted — no API keys, no
non-commercial licenses.
Status & known limitations (v0.1)
Measured on an 8 GB M1 (under memory pressure):
| Case | Result |
|---|
| English, fn up → text ready, with a natural pause (≥ 0.35 s) before release | ✅ 1–15 ms — transcribed speculatively while fn is held |
| English, fn up → text ready, releasing immediately after speaking | ✅ 121–139 ms (Parakeet 0.6B); ~40–70 ms with the optional 110M model |
Self-corrections (Benchmarks/corrections.tsv, 83 cases) | ✅ 83/83 · holdout of 30 unseen phrasings: 28/30, no false corrections |
| Local LLM for corrections (Qwen3 0.6B / 1.7B, deletion-only) | ❌ evaluated and rejected: 48–65/83 and 100–900 ms slower than rules (evoo-cli corpus --llm) |
| LLM path (unresolved corrections, Hinglish) | ~1–4 s on 8 GB M1 — opt-in, only when needed |
| Qwen3 4B | No better on the English set and 3–5× slower on 8 GB; meant for 16 GB+ Macs |
| Hinglish | ⚠️ Mixed. Whisper outputs Devanagari and the 1.7B model sometimes translates or misses corrections when romanizing. Needs a better model — top roadmap item |
Also not handled yet: Secure Input fields (password boxes, Terminal "Secure Keyboard Entry") block fn detection;
non-QWERTY layouts may need a different paste key code; keyboards without fn need a fallback hotkey; no
signed/notarized release yet.
License
Apache-2.0. See NOTICE for model and library attributions..
