Parameter Tuning

[hamzaakramsybrid]

Hello to all,
I hope you are fine.

I have perform speaker diarization on my data using pipeline torch.hub.load('pyannote/pyannote-audio', 'dia') but it was not giving me good results. I explored the speech activity detection and tunned the parameters offset and onset using Binarizae from pyannote.audio.utils.signal

sad_scores = sad(test_file)
        binarize = Binarize(offset=0.8, onset=0.8, log_scale=True,
                        min_duration_off=0.1, min_duration_on=0.1)
        speech = binarize.apply(sad_scores, dimension=1)

and it gave me better results as compare to pipeline. Can anyone please guide me how to tune the parameters in Pipeline as it will improve the speech activity detection which will further improve the speaker diarization.

Thank you so much.

[hbredin]

I'd suggest you start with this tutorial that teaches how to tune pipeline hyper-parameters.

[hamzaakramsybrid]

Thank you so much for the suggestion, but I wanted to change the hyperparameters in Binarize class so I can change the results of speech activity detection which will further improve the speaker diarizatioin. I have change tried log_scale False and True and it gave different results. Also, I have generated results for scale= ['absolute', ;relative', 'percentile] and it is giving different results. The thing that I wanted to ask is that I have been changing the onset and offset parameters but the generated results are the same as it is not having any effects by changing the value of offset and onset. Kindly guide me so I can generate multiple results by changing the value of onset and offset. Thank you Regards, Hamza Akram

…

On Mon, Aug 30, 2021 at 8:19 PM Hervé BREDIN ***@***.***> wrote: I'd suggest you start with this tutorial <https://github.com/pyannote/pyannote-audio/blob/master/tutorials/pipelines/speech_activity_detection/README.md> that teaches how to tune pipeline hyper-parameters. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#732 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AVFIDSJWQL7SIFLZHTGIDZ3T7OOQLANCNFSM5DBIDLNQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.