AiVoiceKit

A Swift package providing on-device voice dictation, command routing, and AI-assisted rewriting for macOS (iOS/visionOS platform stubs included for future work).

Originally ported from FluidVoice by Aether AI Studio / altic-dev, and adapted for use as a standalone, host-app-agnostic package. AiVoiceKit has no dependency on Alric or any other host application — the host wires it up via a plain callback-based VoiceEngine protocol.

What it does

• Local ASR via Apple's built-in SFSpeechRecognizer/SpeechAnalyzer (zero download), or downloadable on-device models: Whisper (Tiny–Large via SwiftWhisper), Parakeet Flash/TDT v2/v3, Nemotron Speech, and Cohere Transcribe (all via FluidAudio, CoreML, arm64 macOS)
• Global hotkey activation — hold-to-record, toggle, or double-tap, with configurable shortcuts for dictation, command mode, edit mode, cancel, and paste-last-transcription
• Command routing — text prefixed with a configurable wake word (e.g. "Alric, summarize this") is routed to a host-supplied callback instead of being typed into the frontmost app
• Edit mode — rewrite selected text in the frontmost app via a host-supplied AI callback
• Notch/bottom overlay — live transcript, waveform, and mode indicator via DynamicNotchKit
• History, stats, and custom dictionary persistence, file-based (no external database)
• AI post-processing — punctuation, capitalization, and provider-based (Apple Intelligence / OpenAI / Groq / custom) dictation cleanup

Integration

The host app owns zero AiVoiceKit-specific UI logic beyond wiring two callbacks:

let voiceEngine = VoiceEngineMacOS(
    onCommandReceived: { text in
        await MyApp.shared.handleVoiceCommand(text)
    },
    onEditRequested: { selectedText, instruction in
        return await MyApp.shared.rewrite(selectedText, instruction: instruction)
    }
)

VoiceEngine is a plain AnyObject, ObservableObject protocol — no host framework types appear in the package's public API.

Example app

AiVoiceKit's dictation and voice-command flow is used end-to-end in AICompleteChat, a full-source, production-quality on-device AI chat app built on DesignFoundation and DesignFoundationPro — NerdSnipe Inc's Swift design-system packages. DesignFoundation is the free, MIT-licensed core (tokens, primitives, layout shells); DesignFoundationPro builds on it with complete, drop-in application verticals (chat, CRM, project management, analytics, and more) so a real app's UI is assembled from proven components rather than built from scratch. AICompleteChat is the reference for wiring AiVoiceKit's VoiceEngine into a DesignFoundationPro AIChat vertical — voice input, persistent memory (AiPersona), and on-device MLX inference, all running locally with no cloud dependency.

Package layout

Sources/
  AiVoiceKit/
    Public/     — protocols, types, enums (VoiceEngine, ASRModel, VoiceSettings)
    macOS/      — CoreAudio, NSEvent, Accessibility, overlay (macOS only)
    Shared/     — ASR catalog, history, stats, settings persistence (cross-platform)
  CoreAudioCaptureSupport/ — C target for low-level CoreAudio capture
Tests/
  AiVoiceKitTests/

Requirements

• macOS 15+ (primary target). iOS 18+/visionOS 2+ platforms are declared but currently compile no macOS-only code paths in — full support is planned but not yet implemented.
• Swift 6.0 toolchain
• Microphone access; Accessibility permission for global hotkeys and typed output

License

AiVoiceKit is a derivative work of FluidVoice (GPLv3, © altic-dev contributors) and is distributed under the same license — see LICENSE.

Third-party dependencies and their licenses are listed in-app via the Open Source Licenses sheet, and include:

Dependency	License	Purpose
FluidAudio	MIT	Parakeet/Nemotron/Cohere CoreML ASR
SwiftWhisper	MIT	Whisper inference
DynamicNotchKit	MIT	Notch overlay window management
PromiseKit	MIT	Legacy async utilities (updater path)