AIOn-device AI transcription means the speech model runs on the phone itself — there is no upload step for your audio to take. What actually happens between the microphone and the text, what the Neural Engine does, why memory (not compute) decides which model your app runs, and what on-device genuinely cannot do.
AIA reference page of speech-to-text statistics that all trace to a primary source: benchmark WERs by audio condition, model-by-model error rates, the human baseline, bias and hallucination findings, and dictation speed — plus the widely quoted numbers we could not verify and left out.
AIWord error rate counts substitutions, deletions, and insertions against the words actually spoken. What the formula includes, what it quietly ignores, why the same model scores differently in two papers, and how to measure your own WER in ten minutes.
AIApple's on-device SpeechAnalyzer beat Whisper in a widely shared benchmark — but that test ran Whisper Small on English audiobooks. Here is what it measured, what it did not, and which engine fits which recording.
AIWhat Whisper accuracy actually looks like: roughly 2–3% word error rate on clean read speech, about 7% across mixed English benchmarks, around 16% in meeting rooms — and what moves those numbers most.
AIThe four best Whisper apps for iPhone and iPad, compared by the Whisper model variant each one actually runs and whether it runs on your device or a server.
AIA practical comparison of Parakeet V3 and Whisper Large V3: benchmark accuracy, speed, language coverage, hallucination behavior, and which model fits on-device transcription.
AIA clear look at how Whisper Large V3 Turbo makes the Large V3 architecture faster and lighter, where it still has limits, and why it powers transcription in LoroNote.
