Docs / Models / Engine Selection

Engine Selection

An engine is the combination of a transcription backend and its associated models. Murmur supports four engines. You can switch between them at any time from Settings > Dictation Models, where an active engine selector appears at the top of the page.

Available engines

Parakeet

  • Languages: English only
  • Streaming: Yes
  • Recommended for: English dictation — best accuracy-to-speed ratio for English speech

Parakeet is the default recommendation for English speakers. It runs entirely on-device and produces consistently accurate output without the latency of larger models. Download Parakeet V2 or V3 from the model catalog.

WhisperKit

  • Languages: 99 languages + auto-detect
  • Streaming: Yes (partial results shown during recording)
  • Recommended for: multilingual use, technical vocabulary, or when you need the widest language coverage

WhisperKit runs Whisper models locally via Core ML. Streaming shows partial transcription results as you speak, so you can see text appear before you stop recording. The active Whisper model (tiny through large) determines the accuracy and speed trade-off.

Apple Speech

  • Languages: 40+ languages
  • Streaming: Yes
  • Recommended for: getting started without any download, or when using macOS Tahoe's improved Analyzer

Apple Speech requires no model download. On macOS Tahoe, the Analyzer variant provides improved accuracy. On earlier macOS versions, SFSpeech is used instead.

OpenAI

  • Languages: 99 languages
  • Streaming: No (results delivered after recording stops)
  • Recommended for: maximum accuracy, noisy environments, or when local model quality is not sufficient
  • Requires: internet connection, OpenAI API key

OpenAI sends your audio to OpenAI's servers for transcription. Results come back after the recording ends — there is no real-time streaming. This engine produces the highest accuracy across all conditions but requires an API key and incurs usage costs.

Switching engines

Open Settings > Dictation Models and use the engine selector at the top. The change takes effect immediately. If the newly selected engine has no model downloaded yet, Murmur prompts you to download one.

Language selection

Language selection is available for WhisperKit and OpenAI. Set it below the engine selector. Choose a specific language for best accuracy, or leave it on Auto-detect to let the engine identify the language from your speech.

Parakeet and Apple Speech use your system language automatically and do not offer a manual language override.

Previous Model Catalog
Next Downloading & Managing Models