Think out loud,
without the cloud

Dictation and AI cleanup that
runs entirely on your Mac.
Local. Private. Native.
Works wherever you type.
Starting at $1/mo,
$10 for lifetime

System requirements

  • Apple Silicon
  • macOS 14.2 or newer
  • 3.1 GB disk space
Download for macOS
Free trial: Try your first 3,000 words free.
No card required.
~0.4s
speech to text
100%
on-device
4x
faster than typing
$1/mo
to get started

Full pipeline, measured p50 for a short dictation on an Apple M4 Max. Full method.

Entirely local

Write polished text
with your voice. And more.

Evoglyph removes filler words, corrects technical terms, and pastes a cleaned-up version of what you said. Command Mode lets you edit or delete your last dictation and control your Mac, all with your voice.

LFM2-2.6B + Qwen3-1.7B running in-process, no server, no network call

Where it works

Works anywhere you can type

Verified across every type of macOS app, from editors and browsers to chat and notes. These are the ones we test, and it works in the rest too.

Code & terminal

VS Code
Cursor
Terminal

Browsers

Chrome
Safari
Firefox

Chat & AI

Slack
Claude
ChatGPT

Notes & docs

Obsidian
Notion
Linear

Works in standard text inputs across macOS: code editors, browsers, chat and email, notes and docs, terminals, and ordinary text fields. If an app has no native field, Evoglyph falls back to universal text injection.

Secure fields are the one exception. Password and other protected inputs are intentionally skipped, so Evoglyph never pastes into them.

Why evoglyph

Our value proposition

Local & fast

Audio and transcripts never leave your Mac, so there is no cloud backend to leak to. Everything runs on your Mac's Apple silicon, turning speech into clean, finished text in about 0.4 seconds for a short dictation on an Apple M4 Max.

Worth paying for

Evoglyph is a maintained product, not a side project. A few dollars funds frequent releases, real support, and the newest on-device models as they ship, all for a fraction of what cloud subscriptions cost.

Works everywhere

It adapts its injection method per app, so native fields, browsers, and terminals all just work. See the verified list above.

Lightweight & accurate

One native Swift app, about 175 MB installed, not an 800 MB Electron app. Add your technical terms and product names, and cleanup spells them your way.

~0.4s

Speech to finished, cleaned-up text for a short dictation on an Apple M4 Max, transcription and the AI pass together.

100%

of the 323 short dictations in our latest baseline finished in under one second, on an Apple M4 Max.

0

network calls in the path. The whole pipeline runs on your Mac's Apple silicon, fully offline.

Measured on an Apple M4 Max for a short one-sentence transcription. Latency scales with how long you speak, and slower Apple Silicon will be higher. The transcription stage alone runs in about 0.14 seconds. Full method.

The engineering

Open at the core, custom where it counts

Evoglyph's transcription pipeline is built on the best open-source speech and language models, wrapped in custom-built modules that deliver
clean, refined text that's true to your voice.

Foundation

Open weights & open source

Open models & frameworks

FluidAudio
Apache-2.0

The open Swift framework that runs the audio models in our pipeline on the Apple Neural Engine: speech-to-text, voice-activity detection, and the speaker embeddings behind the whisper gate. Learn more

on-device Neural Engine
🦜 Parakeet TDT 0.6B V2 (En)
CC-BY-4.0

A 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription, featuring support for punctuation, capitalization, and accurate timestamp prediction. Runs on the Apple Neural Engine at a 2.41% word error rate on our harness. Learn more

Neural Engine 2.41% WER
Silero VAD
MIT

An open neural voice-activity detector (VAD) scores your audio frame by frame for speech probability and marks the exact spoken span. We trim each clip down to that span before Parakeet transcribes, keeping a 500ms guard on each side, so the recognizer only ever sees speech, never silence or room tone. Learn more

on-device silence trimming
Liquid LFM2-2.6B
LFM Open License

A dense 2.6B-parameter language model, designed for on-device inference and small enough to run in-process on the Apple GPU. It's the base our cleanup adapter is built on. Learn more

language model GPU, 4-bit
Qwen3-1.7B (command actions)
Apache-2.0 New

The dedicated model that turns a spoken command into one bounded system action. About 0.98 GB, 4-bit, running on the Apple GPU via MLX. Downloaded during setup with the other models, then verified and offline. Adapter-free. Learn more

on-device GPU, 4-bit

Custom layer

evoglyph proprietary

Engineered by evoglyph

Vocabulary layer

When what you said comes close to one of your terms, we hand that term to cleanup, so your names and jargon come out spelled your way. Learn more

near-match your spelling
Adaptive whisper gate

Enroll your voice once, and evoglyph transcribes even whispers, matching quiet audio to your voice profile so it hears you, not the room. Learn more

speaker identity voice enrollment
LoRA cleanup adapter

Our own fine-tune of the open-weights LFM2-2.6B, trained to turn rambling speech into clean, readable text. Learn more

on LFM2-2.6B GPU, in-process
Homophone correction

The cleanup adapter also fixes context homophones like their/there and one/won from the surrounding words. Learn more

context-aware in adapter
Span-by-span safety net

Deterministic checks judge each edit span by span, reverting anything doubtful to raw and guarding against changes to the words you said. Learn more

span by span reverts to raw
Command action harness
New

The layer that makes voice control trustworthy. A spoken command passes three gates before anything runs: a deterministic injection veto, a yes or no command check, then a strict parse into a typed action. If it is command-shaped but not a valid action, Evoglyph refuses rather than guesses. Learn more

on-device bounded catalog refuse, never guess

How it compares

Everything you need

A carefully crafted user experience that provides essential voice-to-text functionality. No feature bloat.

Feature
evoglyph
Wispr Flow
Superwhisper
Voice Ink
On-device voice-to-text
✓
×
✓
✓
On-device AI cleanup (default)
✓
cloud
optional
none built-in
Custom vocabulary (add your own terms)
✓
✓
✓
✓
File transcription (import audio files)
✓
×
✓
✓
Transcription history
✓
✓
✓
✓
Command mode
✓
✓
✓
✓
Monthly subscription
$1/mo*
$15/mo
$8.49/mo
×