Private, on-device voice dictation for macOS and iPhone. A free, offline alternative to Wispr Flow / Superwhisper: speak, get text, paste anywhere — your audio never leaves the device.
-black){kind=icon}
Transcription runs 100% on-device via WhisperKit (OpenAI Whisper on the Apple Neural Engine). No cloud, no account, no API key, works in airplane mode.
Cloud dictation apps lose your transcript on a Wi-Fi blip, truncate long brain-dumps, charge a subscription, and send your voice to someone else's servers. HermesVoice does the same job locally: it can't lose your words to the network, it's free, and nothing leaves your phone or Mac.
The one honest trade-off vs. cloud apps: no server-side LLM "polish." Whisper's large-v3-turbo already punctuates naturally, and a built-in personal dictionary handles names, acronyms, and dialect.
Both platforms
- On-device Whisper transcription (no internet, no key, no account)
- Multilingual + code-switching (mixing e.g. English terms into German) — native to Whisper
- Local history of recent dictations
macOS (menu-bar app)
- Global hotkey to start/stop dictation
- Pastes into the focused app via the Accessibility API (system-wide)
- Optional text cleanup via the local Claude CLI (no API key — uses your Max plan subprocess)
- Vocabulary that learns from your corrections
iPhone
- Action Button trigger → record in any context
- Custom keyboard that auto-inserts the last dictation into any app (clipboard-based, aligned with iOS 26.4's "swipe back to your app" flow)
- Live waveform + level-reactive "listening" pulse while recording
- Personal dictionary —
heard → correctpost-correction for names, acronyms, dialect (e.g. oida) - Light / Dark / System appearance, Liquid Glass UI (iOS 26)
large-v3-turbo(632 MB, quantized) bundled for offline first-launch
Swift · SwiftUI · WhisperKit · AVFoundation · AppIntents (Action Button) · UIInputViewController (keyboard) · XcodeGen for the project definition.
Sources/
HermesCore/ # shared, platform-agnostic core (Transcriber, AudioRecorder,
# AudioMeter, History/Recording/Vocabulary/UserDictionary stores)
HermesVoice/ # macOS menu-bar app
HermesVoiceiOS/ # iOS app (RecordView, DictationController, DictionaryView, …)
HermesKeyboard/ # iOS keyboard extension (clipboard auto-insert)
Models/ # WhisperKit model (gitignored — see Setup)
docs/POLISH.md # design system / UI rules
project.yml # XcodeGen project definition (.xcodeproj is generated, gitignored)
Requires macOS + Xcode 26 and an Apple developer team for device signing.
# 1. Install XcodeGen
brew install xcodegen
# 2. Download the on-device Whisper model (gitignored — too large for git)
./scripts/download-model.sh # ~646 MB into Models/
# 3. Generate the Xcode project
xcodegen generate
# 4. Open it
open HermesVoice.xcodeprojIn Xcode pick a scheme — HermesVoice (macOS) or HermesVoiceiOS (iPhone) — and Run.
iOS signing: set your team in project.yml (DEVELOPMENT_TEAM) or via Xcode's Signing & Capabilities, then xcodegen generate again. Build Release for a physical device (Debug dylibs don't launch standalone).
iPhone keyboard (optional): Settings → General → Keyboard → Keyboards → add HermesVoice and allow Full Access (needed to read the clipboard). Bind the Action Button to the HermesVoice shortcut to start dictation hands-free.
iOS keyboard extensions are capped at ~77 MB RAM and can't reliably use the mic, so Whisper can't run inside the keyboard. Instead: the app records + transcribes on-device (triggered by the Action Button), copies the result, and the HermesVoice keyboard inserts it into whatever app you're in. That's one app-hop per dictation — the same limitation every on-device iOS dictation tool has.
The bundled model folder name lives in Sources/HermesVoiceiOS/DictationController.swift (bundledModelFolder) and in project.yml. To use the full 1.6 GB openai_whisper-large-v3-v20240930_turbo (slightly better, more RAM), download that folder instead and update both references.
Built on WhisperKit by Argmax. Models from the argmaxinc/whisperkit-coreml Hugging Face repo.
MIT — see LICENSE.