pyannote gives wrong results

[amitli1]

I'm trying to use pyannote for speaker diarization
and I'm getting wrong number of speakers.

Any example I tried I got wrong results.

For example:

• I used this youtube file:
  https://www.youtube.com/watch?v=b2_ZZ2UpSzI

• I convert it to wav file with sample rate of 16000.

I run the following code:

from pyannote.audio import Pipeline

TEST_FILE = "example.wav"
MY_TOKEN = "..."
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization",
                                    use_auth_token=MY_TOKEN)

diarization = pipeline(TEST_FILE)

And I got the following diarization:

• The GT contains 4 speakers and not 2.

How can I tweak pyannote and get better results ?

[joagonzalez]

have you tried helping the pipeline with num_speakers parameter?

pipeline = Pipeline.from_pretrained(
                     self.model,
                     use_auth_token=self.token,
                 )        
        
diarization = pipeline(
    filename, 
    num_speakers=self.num_speakers
)

[amitli1]

Now it separate to 3 different speakers, but with very bad separation:

What are my options? (I'm using a simple wav file, so I assumed it need to be an easy task for pyannote)

[hbredin]

See FAQ related to how to improve performance, maybe?