bad diarization performance on a sample youtube video

[ywangwxd]

Hello,

I have extracted the audio channel from the YouTube video found at this link. Subsequently, I tested the audio file using the local file streaming mode with diart.stream. However, the resulting rttm file did not meet my expectations. It is evident that there are four distinct speakers in the video, each with a sufficient duration of speech. Yet, the Speaker Diarization pipeline only identified three speakers, not mention the alignment performance of speakers and their respective speech segments on the timeline.

I used "pyannote/segmentation-3.0" and "pyannote/wespeaker-voxceleb-resnet34-LM" for segmentation and embedding model respectively. How should I adjust the parameters? I do not have lablled data.

SPEAKER test_uri 1 1.101 8.408 speaker0
SPEAKER test_uri 1 9.508 3.500 speaker1
SPEAKER test_uri 1 13.008 1.000 speaker0
SPEAKER test_uri 1 14.008 26.500 speaker1
SPEAKER test_uri 1 26.508 0.567 speaker0
SPEAKER test_uri 1 32.975 0.033 speaker0
SPEAKER test_uri 1 40.508 1.000 speaker0
SPEAKER test_uri 1 41.508 0.500 speaker1
SPEAKER test_uri 1 42.392 0.117 speaker0
SPEAKER test_uri 1 42.508 0.500 speaker1
SPEAKER test_uri 1 43.008 0.083 speaker0
SPEAKER test_uri 1 43.092 0.400 speaker1
SPEAKER test_uri 1 43.492 10.517 speaker0
SPEAKER test_uri 1 52.558 0.017 speaker1
SPEAKER test_uri 1 54.008 3.567 speaker1
SPEAKER test_uri 1 57.575 5.233 speaker0
SPEAKER test_uri 1 62.808 0.200 speaker1
SPEAKER test_uri 1 63.008 1.000 speaker0
SPEAKER test_uri 1 64.008 0.617 speaker1
SPEAKER test_uri 1 64.508 1.500 speaker0
SPEAKER test_uri 1 64.942 1.200 speaker1
SPEAKER test_uri 1 66.292 0.017 speaker1
SPEAKER test_uri 1 66.442 0.067 speaker1
SPEAKER test_uri 1 66.958 0.350 speaker0
SPEAKER test_uri 1 67.508 0.050 speaker0
SPEAKER test_uri 1 67.558 0.950 speaker1
SPEAKER test_uri 1 68.508 6.583 speaker0
SPEAKER test_uri 1 75.058 0.100 speaker1
SPEAKER test_uri 1 75.158 0.733 speaker0
SPEAKER test_uri 1 75.492 0.017 speaker1
SPEAKER test_uri 1 75.892 0.067 speaker1
SPEAKER test_uri 1 75.958 1.717 speaker0
SPEAKER test_uri 1 77.975 0.117 speaker0
SPEAKER test_uri 1 78.092 0.417 speaker1
SPEAKER test_uri 1 78.508 0.500 speaker0
SPEAKER test_uri 1 79.008 0.500 speaker1
SPEAKER test_uri 1 79.008 3.017 speaker2
SPEAKER test_uri 1 79.508 0.500 speaker0
SPEAKER test_uri 1 81.008 2.883 speaker0
SPEAKER test_uri 1 83.508 0.583 speaker2
SPEAKER test_uri 1 84.008 2.650 speaker0
SPEAKER test_uri 1 84.092 0.100 speaker1
SPEAKER test_uri 1 84.192 0.433 speaker2
SPEAKER test_uri 1 85.508 5.000 speaker1
SPEAKER test_uri 1 90.508 1.000 speaker0
SPEAKER test_uri 1 91.508 1.000 speaker1
SPEAKER test_uri 1 92.508 1.217 speaker0
SPEAKER test_uri 1 93.508 0.500 speaker2
SPEAKER test_uri 1 94.008 5.500 speaker1
SPEAKER test_uri 1 94.325 1.683 speaker0
SPEAKER test_uri 1 97.508 1.000 speaker0
SPEAKER test_uri 1 99.008 1.067 speaker0
SPEAKER test_uri 1 100.508 1.267 speaker0
SPEAKER test_uri 1 102.108 9.900 speaker0
SPEAKER test_uri 1 112.008 25.050 speaker2
SPEAKER test_uri 1 135.358 0.367 speaker0
SPEAKER test_uri 1 137.008 5.900 speaker1
SPEAKER test_uri 1 142.842 0.167 speaker2
SPEAKER test_uri 1 143.008 0.467 speaker0
SPEAKER test_uri 1 143.192 1.833 speaker1
SPEAKER test_uri 1 144.508 1.500 speaker0
SPEAKER test_uri 1 146.008 12.383 speaker1
SPEAKER test_uri 1 158.492 2.517 speaker1
SPEAKER test_uri 1 161.008 7.000

[juanmc2005]

Hey @ywangwxd , unfortunately, without labeled data the adjustment of parameters will be a matter of trial and error.