Dicta

Dicta is a deliberately small, local-first dictation utility for macOS 26 and later. It is a menu-bar app built with SwiftUI, narrow AppKit/Accessibility bridges, Swift Concurrency, and FluidAudio 0.15.7.

The name comes from dicta (the plural of dictum) and from stopping halfway through “dictation”.

See docs/SPEC.md for the product and concurrency contract.

Requirements

• macOS 26+
• Apple Silicon Mac
• Xcode with Swift 6 support

Apple Silicon is a product requirement, not merely a performance recommendation: several formally supported FluidAudio ASR pipelines rely on the Apple Neural Engine, and Dicta guarantees that the pinned FluidAudio release's standalone ASR catalog is usable.

Project shape

Dicta is a normal Xcode macOS application target. Swift Package Manager is used only for the FluidAudio dependency; the application itself is not packaged as a SwiftPM executable.

FluidAudio is pinned to 0.15.7 in Dicta.xcodeproj.

Build and run

Open Dicta.xcodeproj, select the Dicta scheme, choose My Mac, and Run.

The app target intentionally:

• targets macOS 26.0;
• uses Swift 6;
• runs as an LSUIElement menu-bar accessory app;
• disables App Sandbox because cross-application Accessibility insertion is a core feature;
• enables Hardened Runtime and the microphone audio-input entitlement;
• declares NSMicrophoneUsageDescription.

For distribution, use the normal Developer ID signing/notarization workflow rather than the Mac App Store sandbox.

First run

1. Open Dicta Settings.
2. Grant Accessibility access.
3. Choose a final model and download it explicitly from the Models pane. The default is Cohere Transcribe.
4. If using English live preview, choose and download a Parakeet EOU preview model.
5. If using Toggle + silence auto-finish, Dicta prepares the Silero speech detector when needed.
6. Press the configured global shortcut (default ⌥Space).

Microphone permission is requested by macOS on the first recording.

Model catalog

Model metadata lives in Dicta/Resources/models.json, not in Swift source. The manifest describes model IDs, supported language codes, preview/final roles, backend metadata, defaults, and coarse 1–5 speed/accuracy ratings.

At launch Dicta uses the newer of the bundled manifest and the last-known-good cached manifest. It then checks the canonical models.json on this repository and atomically caches a newer valid revision for the next launch. A malformed, unsupported-schema, stale, or unavailable remote manifest never replaces the working catalog.

The catalog is deliberately forward-compatible. It can describe future FluidAudio models and languages before an older Dicta binary knows how to execute them; ModelRuntimeCapabilities prevents unknown backend/variant combinations from becoming selectable.

For the pinned FluidAudio 0.15.7 release, Dicta supports every standalone ASR family listed in FluidAudio's formal model catalog: Parakeet TDT v2/v3/TDT-CTC 110M/Japanese, Cohere Transcribe, SenseVoiceSmall, Paraformer-large (zh), Parakeet EOU, English Nemotron Streaming, multilingual Nemotron Streaming, and Parakeet Unified in its streaming and offline forms. FluidAudio's Parakeet CTC keyword-spotting/rescoring models are auxiliary ASR components rather than standalone transcription engines, so they are not offered as final-transcription choices.

Dicta currently exposes Japanese, English, and Chinese as input languages. Catalog language codes remain open strings, so adding French, German, Korean, or other model capabilities does not make older app binaries fail to decode the manifest.

Large model files are managed explicitly in Settings. Selecting a model does not implicitly download it; the Models pane shows local installation state and provides Download and Remove actions.

Speed and accuracy values are qualitative relative ratings intended for model selection. They are not presented as directly comparable benchmark measurements across hardware, languages, or datasets.

Preview text is never inserted. Only a final pass can be delivered to another app.

Privacy

ASR inference is local. Recorded PCM is kept only in memory until its finalization job completes. Transcript history is local JSON under Dicta's Application Support directory and is pruned to 100 items / 7 days. No audio history is persisted.

Dicta does make network requests for two narrow purposes: checking the public model-catalog metadata on this GitHub repository at launch, and downloading a model from FluidAudio's model source when the user explicitly downloads it in Settings. Recorded audio and transcript text are not sent with the catalog request.