Skip to main content

Speech engines and models

Choose in Settings → STT → Speech Recognition Engine.

EngineWhere it runsNeedsNotes
WhisperKit (local)On your MacA model downloadPicking a model in the tutorial selects this engine. Supports live preview. Pause line breaks work only here.
MLX Audio (local)On your Macuv installedThe default when you skip the tutorial and uv is available. The default model is Qwen3-ASR-1.7B-8bit.
Groq Cloud (fast)CloudGroq API keyAudio is sent to Groq. Test connection under the key checks that it works.
Apple (built into macOS)On your MacmacOS 26 or laterUses the speech recognition built into macOS, with no model file to download. The OS downloads the language assets the first time. Supports live preview.
Settings → STT

Choosing between Apple and WhisperKit​

Measured on macOS 27 with 120 sentences from a public Korean read-speech dataset (18.4 minutes, 10 speakers) and a 145-second real lecture recording. Figures are character error rates on sentences without numbers.

ConditionAppleWhisperKit Large v3 Turbo (default)
Clean read speech3.3%3.0%
Read speech + babble noise6.0%6.4%
Sentences more than 30% wrong (noise)6 of 12010 of 120
Time for 18.4 minutes14 seconds240 seconds
  • On clean speech WhisperKit is slightly more accurate, and Apple is 17 times faster.
  • With noise or on long real recordings, Apple broke down less. WhisperKit Large v3 Turbo made up and repeated sentences in one stretch of the lecture recording. Apple's errors are word-level, such as writing 범용 (general-purpose) as 법용.
  • The most accurate is WhisperKit Large v3. It was the most accurate of the three on the same lecture recording, and it is the slowest.
  • Apple spells English terms in Hangul (isolation → 아이솔레이션).
  • Apple does not detect the language. With the transcription language set to Auto-detect, it listens for Korean.

Choosing a WhisperKit model​

A model downloads as soon as you pick it. See progress and sizes in the Settings → Models tab.

ModelSizeCharacter
Tinyabout 77MBFastest, but hard to use for Korean.
Baseabout 147MBOften gets proper nouns and technical terms wrong.
Smallabout 486MBThe lower bound where Korean becomes usable.
Large v3 Turboabout 3.2GBDefault. Balanced speed and accuracy
Large v3 Turbo (v20240930)about 1.6GBA later snapshot of the same family
Large v3 Turbo compressedabout 632MBSmaller and faster than the default, with lower CPU use.
Large v3 quantizedabout 947MBA smaller version of Large v3
Large v3about 3.1GBMost accurate and slowest.

If transcription is slow, check Activity Monitor for other apps using the CPU before switching models. When specific terms keep coming out wrong, adding them to the Dictionary is quicker than a bigger model.

Models are stored in ~/Library/Application Support/Sogon/Models/whisperkit.