Dictly

Press a hotkey. Speak. Paste anywhere.

A free, open-source macOS menu-bar dictation app. Hold a global hotkey, speak, release — your speech is transcribed locally by Whisper and placed on the clipboard (or auto-pasted into the focused app, depending on build).

Free · Apple Silicon only · macOS 15+ · 99 languages · 100% on-device

No cloud API. No account. No telemetry. Audio never leaves your Mac.

Download

Installing

• Unzip the download.
• Drag Dictly.app into /Applications.
• Double-click to launch.

The build is signed with a Developer ID certificate and notarised by Apple, so it opens without any Gatekeeper warnings on macOS 15+.

On first launch macOS will ask for microphone permission (required) and Accessibility permission (only needed if you want the transcribed text to be auto-pasted into the focused app; without it Dictly falls back to clipboard-only mode and you press ⌘V manually).

All releases & changelogs →

Coming next

Things I'm actively working on for upcoming releases:

• 🌐 Speak-and-translate — dictate in one language and have the

text pasted in another. Useful when you think faster than you type in your second language. Whisper's translation capability is already there; just needs the UI and a target-language picker in Settings.

• 🔧 F-key hotkeys — F-keys (and the dedicated 🎤 dictation key on

newer keyboards) currently don't bind cleanly in the recorder because it commits the Fn modifier before the F-key keycode lands. Fix is in the works.

Got ideas / requests / bug reports? Open an issue —

Highlights

• Global push-to-talk hotkey (default Fn). Configurable in Settings.
• OpenAI Whisper via WhisperKit,

running on Apple's CoreML / Neural Engine. Real-time-factor ≈ 0.1–0.3 on M-series.

• 99 languages supported by Whisper out of the box. Auto-detection or pinned

language in Settings.

• Bundled model (Whisper base multilingual, ~139 MB) — ships in-repo so

the app works offline from first launch. Higher-quality variants (large-v3, large-v3-turbo, etc.) are downloadable from HuggingFace inside the app, or fetchable up front via scripts/fetchbundledmodel.py <variant>.

• Pure AppKit, no SwiftUI dependencies.
• Auto-paste into the focused app via simulated ⌘V (needs the standard

macOS Accessibility permission, granted once on first launch).

Project layout

Dictly/
├── Dictly.xcodeproj/
├── Dictly/                        main app target sources
│   ├── App/                     AppDelegate, DictationCoordinator
│   ├── MenuBar/                 NSStatusItem + menu
│   ├── Hotkey/                  KeyCombo, HotkeyManager (Carbon RegisterEventHotKey)
│   ├── Audio/                   AudioRecorder (AVAudioEngine → 16 kHz mono Float32)
│   ├── Transcription/           Transcriber protocol, WhisperKit, ModelCatalog
│   ├── TextInsertion/           TextInserter (clipboard + simulated ⌘V)
│   ├── PostProcessing/          Hook for post-transcription cleanup
│   ├── Permissions/             Mic + Accessibility checks
│   ├── Settings/                UserDefaults wrapper, settings window
│   ├── HUD/                     Floating "pill" recording HUD
│   ├── Onboarding/              First-run window (permissions + model download)
│   ├── Design/                  DesignTokens, BrandButton, RingedIcon
│   ├── Dictly.entitlements           Direct distribution
│   └── PrivacyInfo.xcprivacy
└── BundledModels/               base model (~139 MB) ships in-repo;
                                   larger variants fetched on demand
scripts/
├── fetchbundledmodel.py       Downloads the bundled Whisper model
├── sync_icons.sh                Regenerates app/menu-bar icons from sources
└── syncmenubaricons.swift

Building from source

Prerequisites

• macOS 15 (Sequoia) or newer
• Apple Silicon Mac
• Xcode 16+
• Python 3 (only needed if you want to bundle a different Whisper variant

than the base model that ships in-repo)

Steps

# 1. Open in Xcode
open Dictly/Dictly.xcodeproj
2. Set your own signing team:
   • Project → target Dictly → Signing & Capabilities
   • Team: pick your Apple Developer team
   • Bundle Identifier: change com.mydear.voicetotext to something
     registered to your team (e.g. com.yourname.dictly)
3. Build & run. The bundled base model is already in
   Dictly/BundledModels/ — the app boots offline-ready.
(optional) Bundle a heavier model alongside base. For example:
python3 scripts/fetchbundledmodel.py large-v3-v20240930_547MB
After rebuilding, the app picks the highest-quality bundled model as
its default (see ModelInfo.defaultModelID); switching is live in
Settings → Transcription models.

Note: This repository ships the Direct distribution build only

(full functionality, auto-pastes via simulated ⌘V, needs Accessibility).

The Mac App Store variant (sandboxed, clipboard-only) is maintained

privately by the original author.

Or from the command line:

DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer \
  xcodebuild -project Dictly/Dictly.xcodeproj \
             -scheme Dictly -configuration Release build
open "$(find ~/Library/Developer/Xcode/DerivedData -name Dictly.app -type d | head -1)"

How it works (one-paragraph version)

DictationCoordinator owns the long-lived components. On hotkey press, AudioRecorder boots a fresh AVAudioEngine and starts streaming 16 kHz mono Float32 PCM into a buffer. On release, the buffer is handed to WhisperKitTranscriber, which calls Whisper on the bundled CoreML model. The resulting text passes through a TextPostProcessor and is then pasted into the focused app via CGEvent simulating ⌘V (with a clipboard-only fallback if Accessibility is not yet granted). A floating HUD shows recording/transcribing/ done states throughout.

Configuration

Open Settings from the menu-bar icon → ⚙️ Settings… Everything is configurable without restart; changes apply live.

Hotkey

• Push-to-talk hotkey — the key you hold to record. Click Record and

press your desired combo. Supports both single modifiers (Fn, right ⌥, right ⌘, …) registered through NSEvent.addGlobalMonitor, and regular key combos (⌃⌥D, ⌘⇧Space, …) registered through Carbon's RegisterEventHotKey. Default is Fn. Click End to clear a binding.

• Mode — Hold (push-to-talk: record while pressed, release to transcribe)

or Toggle (press once to start, press again to stop). Hold-mode is conversational, toggle-mode is better for long-form dictation where you don't want to keep a finger down.

Speech

• Spoken language — pick a single language or set to Auto-detect. A

pinned language is usually 5-15% faster and noticeably more accurate on short utterances, because Whisper skips the language-detection pass and doesn't second-guess homophones across languages. Use Auto only if you actually code-switch.

• Quality — trade-off between speed and resilience against degenerate

output (the classic "1, 2, 3, 1, 2, 3…" loop Whisper produces on noisy / Bluetooth-warmup audio): - Fast — single greedy pass at temperature 0. Quickest, but a bad capture stays bad. Good for built-in mic in a quiet room. - Balanced (default) — one safety-net retry if the first pass looks degenerate. Costs ~zero on clean audio (the retry is only paid when the first pass had high compression ratio / low avg-logprob), and rescues most loop cases. The right pick for most users. - Best quality — up to three retries with progressively more randomness. Pays full transcribe cost on each retry, so noticeably slower on bad audio. Best for noisy environments, accented speech, multi-speaker captures.

Maps directly to Whisper's temperatureFallbackCount (0 / 1 / 3).

Transcription models

The list shows all variants available from argmaxinc/whisperkit-coreml plus any models bundled with the .app. The currently active model has a BUNDLED / Use badge.

• Tap Use to switch the active model. The switch is instant if the model

is already cached; otherwise it downloads first (progress shown inline).

• Models are cached under ~/Documents/huggingface/models/argmaxinc/whisperkit-coreml/

and reused across launches.

• Higher tiers (large) are more accurate but slower per second of audio and

use more RAM. base runs at real-time-factor ≈ 0.05–0.1 on M-series and handles 99 languages just fine for chat / messages. large-v3 is the go-to for transcribing meeting recordings or anything where word-level accuracy matters.

• English-only variants (suffix .en) are faster and slightly more accurate

on English, but obviously can't transcribe anything else.

• Tap the refresh icon in the section header to re-fetch the catalog

from HuggingFace.

Behaviour

• Auto-paste into the focused app — after transcription, simulate ⌘V

into whatever app currently has focus. Off → text stays on the clipboard and you press ⌘V manually. Requires the Accessibility permission below.

• Restore clipboard after paste — when auto-paste is on, Dictly briefly

hijacks the system clipboard. With this on, it snapshots the previous clipboard contents and restores them 400 ms after the paste, so you don't lose whatever you had copied. Off → the transcribed text remains the new clipboard contents.

• Show floating HUD while dictating — the pill-shaped indicator with the

live waveform that appears while recording / transcribing. Off → fully silent operation; the only feedback is the menu-bar icon state.

• HUD position — Bottom of screen (centred ~40 pt above the Dock,

default) or Top — under Dictly icon (anchored under the menu-bar icon, useful if you dictate into apps that have their own bottom UI like Slack or Notes).

Permissions

• Microphone — requested on first launch via the standard macOS prompt.

Required.

• Universal Access (Accessibility) — required only for auto-paste

(simulating ⌘V into other apps). Without it, Dictly still works in clipboard-only mode: it copies the text and you press ⌘V yourself. The Open Universal Access button deeplinks straight to the right pane in System Settings.

All preferences live in UserDefaults; nothing leaves your Mac.

Logging

Logs are mirrored to both Apple's Unified Logging system and a rotating local file under ~/Library/Logs/Dictly/.

Unified Logging:

log stream --predicate 'subsystem == "com.mydear.voicetotext"' --info

Local files:

tail -f ~/Library/Logs/Dictly/Dictly.log
tail -f ~/Library/Logs/Dictly/Dictly-crash.log
ls -la ~/Library/Logs/Dictly/MetricKit/

Dictly.log rotates at 2 MB and keeps five archives (Dictly.1.log ... Dictly.5.log). Dictly-crash.log records uncaught Objective-C exceptions and fatal signals (SIGABRT, SIGSEGV, SIGBUS, SIGILL, SIGFPE, SIGTRAP) with a best-effort backtrace. MetricKit crash, hang, CPU exception, disk write exception, and metric payloads are persisted as JSON under MetricKit/ when macOS delivers them.

A pipeline timing line lands at .notice level after each dictation:

pipeline: total=1.08s (transcribe=1.05s post=0.00s insert=0.03s) for 3.07s audio

Contributors

Huge thanks to everyone who's sent a PR, opened an issue, or just tried Dictly and reported back.

License

See LICENSE.

Contributing

Issues and PRs welcome. Please open an issue before large changes so we can discuss the approach.