Docs / Teleprompter / Teleprompter Engines

Teleprompter Engines

Live mode uses a speech engine to track your words in real time. Three engines are available, each with different latency and accuracy characteristics. Your choice affects how snappily the script responds as you speak. (Voice and Auto modes do not use a speech engine — Voice tracks your mic level, and Auto is timer-driven.)

Available engines

Apple Speech (Real-Time)

  • Latency: ~200 ms — the lowest of the three
  • Requires: Speech Recognition permission
  • Best for: Live tracking where responsiveness matters most

Built on Apple's SFSpeechRecognizer, this engine streams results in near real time. No model download is required. For the best accuracy it may use Apple's speech servers, so it works best with a network connection; if privacy or offline use matters, prefer WhisperKit, which runs entirely on-device.

The trade-off is precision: it is tuned for speed. For clean, well-paced speech it tracks well; for heavy technical vocabulary or strong accents, WhisperKit is more reliable.

Apple Speech (Analyzer)

  • Latency: ~0.5–1 second
  • Requires: macOS 26 Tahoe or later
  • Best for: better accuracy than the real-time engine, with still-reasonable latency

The Analyzer uses Apple's newer on-device SpeechAnalyzer API introduced in macOS Tahoe. It is more accurate than the real-time engine at the cost of slightly higher latency, and runs fully on-device. This option does not appear in the engine list on earlier macOS versions.

WhisperKit (Optimized)

  • Latency: ~0.5–1 second
  • Requires: a downloaded Whisper model
  • Best for: highest accuracy, especially technical content or non-native speech, fully on-device

WhisperKit runs a Whisper model locally on your Mac and produces the most accurate word-level output of the three. For teleprompter tracking it favors a small, fast model (tiny/base) polled about twice a second, which keeps latency reasonable while staying accurate. If no Whisper model is downloaded, the setup screen tells you so and blocks Live sessions until you download one — see Downloading & Managing Models.

Device changes mid-session

(New in 1.6) All teleprompter engines now draw from Murmur's shared, hardened audio capture engine. If your microphone changes mid-session — you connect AirPods, unplug a USB mic, or an interruption occurs — tracking pauses briefly while Murmur re-acquires the input and then resumes, instead of failing. See Automatic Device Switching.

Choosing an engine

  • Lowest latency: Apple Speech (Real-Time)
  • Best accuracy on Tahoe, on-device: Apple Speech (Analyzer)
  • Best accuracy overall, on-device: WhisperKit

For most teleprompter use cases, the real-time engine is a good starting point. Switch to WhisperKit if tracking drifts or misses words in your script.

Changing the engine

You can pick the Live-mode engine in two places:

  • Settings > Teleprompter — sets your default tracking engine
  • Session setup screen — choose the engine for a single session from the Tracking Engine card (shown in Live mode)
Previous Teleprompter Modes
Next Floating Teleprompter