The details
For code that uses Apple’s shared API, switching engines can be as simple as changing one import line. Converted models download from the Hugging Face Hub only when an app asks, from about 250 MB to 2.4 GB. In the author’s tests on an M4 Max Mac Studio, Parakeet models transcribing files had lower word error rates than Apple’s own transcriber on the LibriSpeech test set, but every engine topped 23% on meeting recordings with overlapping speech. Testing so far covers macOS 27 only, not iPhones.
Why it matters
Apps often ask a chatbot-style model to make a choice and then parse its text, which hides how confident it is. Returning probabilities lets developers set thresholds and send uncertain cases to people, though the post warns raw probabilities are often overconfident and need calibration. Swappable engines also reduce dependence on Apple’s own models. The figures come from one machine, so results on phones are unknown.