Benchmarks

Measured where it runs, not where it's convenient.

The same Whisper tiny.en model family, scored on the same speech clips, through two on-device runtimes: ONNX Runtime on Android and WhisperKit / CoreML on iOS. Word error rate is computed offline by one algorithm over identical references, so the accuracy numbers are comparable by construction. Speed and memory carry a badge for where they were measured, because a phone's Neural Engine is not a simulator.