Speech to Text

Speech-to-Text (Dictation)

The original Zerm capability: press a hotkey, speak, and the transcript is pasted at your cursor.

Flow

flowchart TB
    HK[Hotkey down] --> UIM[RecorderUIManager.toggleMiniRecorder]
    UIM --> SND[start sound after CoreAudio is live]
    UIM --> ENG[ZermEngine.toggleRecord]
    ENG --> REC[Recorder / CoreAudioRecorder<br/>AUHAL, background queue]
    ENG --> LOAD[WhisperModelManager.loadModel<br/>concurrent Task.detached]
    HK2[Hotkey up / silence] --> PIPE[ZermEngine.runPipeline]
    PIPE --> WAIT[wait until model loaded]
    WAIT --> TX[TranscriptionService.transcribe]
    TX --> FILT[output filter + WordReplacementService]
    FILT --> ENH[AIEnhancementService.enhance · optional]
    ENH --> PASTE[CursorPaster.startPasteAtCursor]

Loading

Providers

Provider	Type	Key file
Whisper (local)	Local ggml	WhisperTranscriptionService.swift
Parakeet / FluidAudio	Local ANE	FluidAudioTranscriptionService.swift
Apple Speech	Local	NativeAppleTranscriptionService.swift
OpenAI, Gemini, Deepgram, …	Cloud	CloudTranscriptionService.swift
Deepgram, ElevenLabs streaming	Cloud streaming	StreamingTranscriptionService.swift

Whisper model loading

Models live in …/WhisperModels/. Loading runs on DispatchQueue.global (blocking C call) inside a detached task; runPipeline waits for isModelLoading == false before transcribing — this prevents the "first transcription after idle fails" bug. Flash attention is on by default (with a fallback for q5_0/q8_0).

Paste pipeline

CursorPaster uses a CGEvent ⌘V by default. The AppleScript paste path (opt-in) runs @MainActor because TIS keyboard APIs must be on the main queue (macOS 26 fix).

Power Mode

Detects the frontmost app and, for browsers, the active tab URL (inline AppleScript targeting the frontmost regular-policy browser instance), then selects context-appropriate enhancement settings.

See: AI Enhancement · Architecture