cpWER (Concatenated Minimum-Permutation Word Error Rate)

Concatenated minimum-permutation word error rate evaluates transcription of a recording with several speakers in it. The words attributed to each speaker are concatenated into one stream per speaker, every possible mapping between the system's speakers and the reference speakers is tried, and the mapping with the lowest total error is the one reported.

Because a word given to the wrong speaker still counts as an error, cpWER penalizes transcription mistakes and speaker-attribution mistakes together. That makes it a close proxy for what a downstream pipeline actually experiences, and a stricter measure than a word error rate that ignores who said what. It is commonly computed with the open MeetEval toolkit.