RTFx (real-time factor)

RTFx is the ratio of audio duration to processing time. A model with RTFx 10 transcribes ten minutes of audio per minute of compute; a model at RTFx 1 only just keeps up with a live conversation. It is the number that decides whether transcription can run during a meeting or has to wait until the end.

RTFx is hardware-bound and honest only when the machine is named. The same speech model can score above 100 on one Mac and below 10 on another, depending on whether the work lands on the Neural Engine, the GPU, or the CPU. Published figures usually come from the newest chip available, so treat them as a ceiling rather than a promise.

For a meeting recorder, the practical thresholds are simple: above roughly 3 the transcript is ready before you have finished closing your laptop, and below 1 the app can never catch up on a long call.