Voice that stays local
A teleprompter can follow you without a speech server.
Manual, Smart Pace, and Private Precision Beta offer three distinct levels of presentation control. None sends your script, microphone audio, or recognition fragments to teleprompter.wtf.
Try the teleprompterLevel 1: Manual is the resilient baseline
Manual scrolling uses an elapsed-time scroll engine with adjustable speed (1 to 10). It never requests microphone permission and continues working seamlessly when advanced browser capabilities or network connections are unavailable.
Level 2: Smart Pace follows your place with the browser's own recognition
Smart Pace uses the speech recognition already built into your browser to follow where you are in the script, matching what it hears against the script by sound as well as spelling. That recognition is your browser's, not ours, and some browsers process the audio on their own servers rather than on the device.
Where recognition is unavailable, Smart Pace falls back to sampling the microphone signal through the Web Audio AnalyserNode. It calibrates to your room's noise floor, detects speech activity, smooths cadence changes, and gradually slows and stops during a pause. That path transcribes nothing and works across any spoken language. Either way there is no model download, which is why it is the default voice mode. Choose Private Precision Beta when you need recognition that stays on your device.
Level 3: Private Precision Beta follows your place with on-device speech recognition
Private Precision Beta downloads a quantized multilingual Whisper Tiny ONNX model directly from this site. A dedicated Web Worker executes speech recognition locally using ONNX Runtime Web WASM, transcribing short 6-second audio windows into temporary text fragments. A deterministic script alignment engine then matches those fragments against a bounded window of your known script tokens.
High-confidence matches gently correct the scroll position toward the current spoken word. If confidence is low (such as during an unscripted aside, loud room noise, or a cough), the system holds position and gracefully falls back to Smart Pace cadence tracking.
The on-device model download
The local Whisper Tiny model and its ONNX WASM runtime total approximately 67 MB (66,874,154 bytes) on first download. Nothing downloads on the homepage or when presenter mode opens in Manual or Smart Pace. Download begins only after you explicitly select Private Precision Beta and press the download button. The versioned files are stored in browser Cache Storage and can be inspected or cleared at any time from the voice panel or the privacy page.
What is temporary vs. what is stored
Microphone PCM buffers and temporary recognition fragments exist only in volatile browser memory while voice tracking is active. They are never written to localStorage, IndexedDB, page URLs, analytics events, or server diagnostics. Closing presenter mode immediately stops all media tracks, closes the AudioContext, and terminates the Web Worker.
Browser capability & hardware requirements
Smart Pace requires standard microphone access and the Web Audio API (supported across 98%+ of modern browsers). Private Precision Beta also requires WebAssembly, a Web Worker, Cache Storage, and sufficient device RAM (at least 4 GB recommended for smooth inference). For detailed browser compatibility tables and fallback paths, see our Browser & Device Compatibility Matrix.
To learn more about how voice tracking compares to fixed-speed scrolling for different production formats, read our guide on Manual vs. Voice-Activated Teleprompting.