Two-Tier Streaming Architecture
Live preview uses a dedicated lightweight model that runs independently from the main transcription engine. This means you get instant visual feedback without sacrificing final accuracy.
A dedicated Parakeet End-of-Utterance model runs on the Apple Neural Engine, processing 320ms audio windows. Produces word-level partial transcripts with ~300ms latency.
When the EOU model is not available, live preview falls back to the main transcription backend processing 2-second chunks during recording.
Live Preview Features
~300ms Latency
The EOU engine processes 320ms audio windows on the Neural Engine, producing word-level partial transcripts with near-instant feedback.
Dual-Engine Architecture
EOU runs on the Neural Engine while the main transcription engine uses Metal GPU — no resource contention between preview and final transcription.
Typewriter Animation
Text appears progressively in the UI with a natural typewriter effect, creating a fluid 'words appearing' experience as you speak.
Keyword Highlighting
Dictionary corrections are shown with color gradient highlighting in the live preview. Click any highlighted word to see the original.
Technical Details
Frequently Asked Questions
How fast is the live preview?
The EOU engine processes 320ms audio windows on the Neural Engine, producing word-level partial transcripts with approximately 300ms latency — faster than any cloud service.
Does live preview affect final accuracy?
No. Live preview uses a separate lightweight model. The main transcription engine runs independently and delivers the final refined result when you release the record button.
Can I disable live preview to save battery?
Yes. Live preview is toggleable in Settings. Disabling it reduces CPU/Neural Engine usage during recording while keeping the main transcription engine active.
Does live preview work with all engines?
The primary EOU preview runs on the Neural Engine (Apple Silicon). When unavailable, Whisperer falls back to the StreamingTranscriber using the main engine with 2-second chunks.