How to read this benchmark
Two copies of the same speech model (Parakeet TDT 0.6B v3) transcribe the same audio: one on the CPU, one Neural Engine-backed through Core ML.
The waveform
The audio as it arrives. The bracket over it is the buffer: everything said since the last pause, which every pass decodes again.
An inference pass
Every 0.8 s while someone is talking, and once more at each pause, the buffer is handed to both paths at the same instant. That is one pass. The bright line on the right flashes each time.
The race tracks
Each runner starts at 0 when a pass is sent and stops when that path returns its transcript. That time (request sent → transcript back, inside the speech helper) is the inference latency shown in milliseconds. It includes preprocessing and decoding, not waiting for you to finish or placing text in an app.
The scores
Faster passes counts paired passes where one path finished more than 5 ms sooner. Average inference lead is the mean of CPU latency minus Neural Engine latency over those pairs.
CPU time
The percentage beside each chip is the CPU time that path's process used, from ps, as a share of one core. It is not power or energy: the Neural Engine's own work does not show up as CPU time.
Highlighted words
Finished words that differ between the two transcripts. Words still being revised live are not compared.