Is pyannote/diariazation pipeline very sensitive to language?

[ywangwxd]

Tested versions

pyannote/speaker-diarization-3.1

System information

Ubunt pyannote/speaker-diarization-3.1

Issue description

How should I improve its performance on Chinese?

Minimal reproduction example (MRE)

I have tested the pipeline on a Chinese audio file. I found the diariazation results is bad, even in easy cases with a long duration of female speech and male speech. The test on an English audio file is quite good though. To reproduce it, just take an audio file which contains speechs in Chinese from multiple speakers. To make the diariazation easy, we can choose one with distinctive speakers, e.g., a male and female speakers.