joined

One reading out of what each piece came back with.

The pieces overlap, so the words at a seam arrive twice, and the second copy is dropped by the text: the longest run the two pieces say alike inside the overlap is found, everything the coming piece says up to the end of it comes off, and so does whatever the piece before said past it. Placed where it falls in the whole recording, so a caller hands over what it was given piece by piece and gets the reading back.

By the text and not by the clock, because the clock is the one thing two builds of one model do not share: the same word decoded in two pieces comes back a fifth of a second apart on one runtime and differently again on the next, so a rule that asks how close two marks are decides differently on each of them. The words agree where the marks do not.

A piece the recogniser had nothing to say about is an answer, not a failure: a stretch of silence is transcribed as no words at all. A count of transcripts that does not match the count of pieces is a failure, and is refused rather than quietly paired off until the shorter of the two runs out.