Docs / Getting Started / Choosing a Model

Choosing a Model

Murmur supports four transcription engines. Each has different trade-offs around speed, accuracy, language support, and privacy. This page explains what each one does so you can pick the right fit.

The short answer

If you dictate primarily in English and want the best combination of speed and accuracy, choose Parakeet V2. For everything else, read on.

Transcription engines

Parakeet (recommended for English)

Parakeet models come from NVIDIA's research and are optimized for fast, accurate English transcription on Apple Silicon.

  • Parakeet V2 — English only. The fastest and most accurate local option for English. Recommended for most users.
  • Parakeet V3 — Multilingual. Covers a broad range of languages with strong accuracy, at a slight speed trade-off compared to V2.

Both Parakeet models run entirely on-device. No audio leaves your Mac.

WhisperKit

WhisperKit runs OpenAI's Whisper speech recognition models locally via Core ML, optimized for Apple Silicon. It supports 99 languages.

Whisper comes in several sizes. Larger models are more accurate but use more RAM and take longer to transcribe.

Model Download size RAM usage Speed Accuracy
tiny 75 MB ~400 MB Fastest Good
base 142 MB ~500 MB Fast Good+
small 466 MB ~1 GB Medium Better
medium 1.5 GB ~2.5 GB Slower Better+
large-v3-turbo 1.6 GB ~3 GB Medium Best
large-v3 3.1 GB ~5 GB Slowest Best+

For multilingual dictation, small or large-v3-turbo are good starting points. The tiny and base models are fast but may struggle with accents, technical vocabulary, or non-English languages.

Apple Speech

Apple Speech uses the speech recognizer built into macOS Tahoe. It requires no download and is available immediately after installation.

  • No storage required — the model is managed by macOS.
  • Works offline for supported languages.
  • Accuracy depends on the language and your macOS version.
  • Not recommended for technical content, heavy accents, or languages where you need the best possible accuracy.

Apple Speech is a good fallback if you want to start immediately without waiting for a download, or if your Mac has limited free storage.

OpenAI API

The OpenAI API engine sends audio to OpenAI's servers for transcription using Whisper. This is a cloud-based option and requires an OpenAI API key.

  • Supports 99 languages.
  • Highest accuracy of any engine, particularly for technical vocabulary.
  • Requires an internet connection and an active OpenAI account.
  • Costs money per minute of audio — billed directly by OpenAI to your account.
  • Audio is sent to OpenAI's servers. Review OpenAI's privacy policy before using this engine.

To use OpenAI API transcription, enter your API key in Settings > AI Refinement > Providers.

How to choose

  • English, best speed and accuracy: Parakeet V2
  • Multiple languages, local only: Parakeet V3 or WhisperKit large-v3-turbo
  • No download, just try it now: Apple Speech
  • Highest accuracy, cloud is acceptable: OpenAI API
  • Low RAM or storage: WhisperKit tiny or base, or Apple Speech

Viewing benchmarks and switching models

The model catalog in Settings > Dictation Models shows each available model with its word error rate (WER) benchmark, estimated transcription speed, and download size. You can download multiple models and switch between them at any time.

Switching takes effect immediately — there is no need to restart Murmur.

Downloaded models are stored in your Application Support folder. You can remove a model from Settings > Dictation Models to free up disk space. Removing a model does not delete any of your transcription history.

Previous Permissions
Next Hotkeys