DER
A standard evaluation metric for speaker diarization systems. It is computed as the combined proportion of missed speech, false alarm speech, and speaker confusion errors relative to total reference speech duration. Lower scores indicate better performance.
Timeline (last 90 days)
The confusion sub-metric of DER does not distinguish between the number of misidentified speakers and the number of misidentified segments, and therefore cannot independently reflect the temporal stability of speaker identity.
The evaluation metric for speaker diarization typically uses DER (Diarization Error Rate), which consists of three types of errors: missed detection, false alarm, and confusion.
The Diarization Error Rate (DER) is derived by summing the durations of missed detection errors, false alarm errors, and speaker confusion errors, then dividing by the total actual speech duration; a lower score is better.
DER does not penalize incorrect prediction of the number of speakers itself; as long as each audio segment is correctly attributed in time, the DER can still be relatively low.
Recent research (such as DER and SR-SAC) has attempted to push UTD up to 4, 10, or even higher in order to squeeze out greater data efficiency.
在高重叠率（>30%）场景下，说话人分离的DER（Diarization Error Rate）通常在15-25%之间，远未达到商用水准
All Facts (6)
The confusion sub-metric of DER does not distinguish between the number of misidentified speakers and the number of misidentified segments, and therefore cannot independently reflect the temporal stability of speaker identity.
50%UnverifiedThe evaluation metric for speaker diarization typically uses DER (Diarization Error Rate), which consists of three types of errors: missed detection, false alarm, and confusion.
50%UnverifiedThe Diarization Error Rate (DER) is derived by summing the durations of missed detection errors, false alarm errors, and speaker confusion errors, then dividing by the total actual speech duration; a lower score is better.
50%UnverifiedDER does not penalize incorrect prediction of the number of speakers itself; as long as each audio segment is correctly attributed in time, the DER can still be relatively low.
50%UnverifiedRecent research (such as DER and SR-SAC) has attempted to push UTD up to 4, 10, or even higher in order to squeeze out greater data efficiency.
50%Unverified在高重叠率（>30%）场景下，说话人分离的DER（Diarization Error Rate）通常在15-25%之间，远未达到商用水准
50%