Performance
The best measure is your own app with your own users: install the example and talk to it. As a guide:
On a mid-range phone, parakeet-tdt_ctc-110m transcribes 11 seconds of speech
in about a quarter of a second: roughly 40 times faster than real time. A
typical spoken sentence comes back as text a few hundred milliseconds after the
speaker pauses, most of which is the pause itself.
moonshine-streaming-tiny is about half as fast. Its decoder works one word at
a time, while parakeet’s produces the whole sentence in one step.
Accuracy
Section titled “Accuracy”Word error rate on LibriSpeech test-clean, as published in transcribe.cpp’s model catalogue:
| Model | Download | Word error rate |
|---|---|---|
parakeet-tdt_ctc-110m-Q4_K_M |
90 MB | 2.53% |
moonshine-streaming-tiny-Q8_0 |
50 MB | 4.52% |
LibriSpeech is clean, read speech. Expect more errors from casual conversation, noisy rooms, accents and specialist vocabulary.
The package adds about 6 MB of engine to an Android app per processor architecture, and about 6 MB on iOS. Models are downloaded at runtime and are not part of your app.