Cannot reproduce cumulative frequency response of SincNet on Speaker-id

[thibaultallenet-cea]

Hello Mirco Ravanelli,
Training SincNet for speaker-id using TIMIT data following your directions, with the config file provided in your github end up with a different cumulative frequency response.
The plot displays the cumulative frequency response of the SincNet filters on speaker-id at initialisation and last epoch (1500 from config file provided)

As you can see, the last epoch cumulative response is very close to the initialization.
Also I checked the cumulative frequency response of the pretrained SincNet's filters you provided (here the last epoch is 360)

Neither of those two models show the cumulative frequency response presented in your paper Interpretable Convolutional Filters with SincNet. Moreover, it seams the filters have a lot of trouble to explore and find a better distribution.
What are your thoughts ?

[mravanelli]

Hi, thank you for raising this issue. I thus have to double check the current model and try to retrieve the original one used in the paper. I will keep you updated. Best, Mirco

…

On Mon, 3 Feb 2020 at 04:21, thibaultallenet-cea ***@***.***> wrote: Hello Mirco Ravanelli, Training SincNet for speaker-id using TIMIT data following your directions, with the config file provided in your github end up with a different cumulative frequency response. The plot displays the cumulative frequency response of the SincNet filters on speaker-id at initialisation and last epoch (1500 from config file provided) [image: Filters_response_init_last] <https://user-images.githubusercontent.com/39118674/73639538-e0ca7b80-466c-11ea-8d03-54c021d2fd2e.png> As you can see, the last epoch cumulative response is very close to the initialization. Also I checked the cumulative frequency response of the pretrained SincNet's filters you provided (here the last epoch is 360) [image: Filters_response_pre_trained_model] <https://user-images.githubusercontent.com/39118674/73639769-4d457a80-466d-11ea-9825-0269e5ddfb65.png> Neither of those two models show the cumulative frequency response presented in your paper Interpretable Convolutional Filters with SincNet. Moreover, it seams the filters have a lot of trouble to explore and find a better distribution. What are your thoughts ? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#84?email_source=notifications&email_token=AEA2ZVVVL4HGBWDHA7SNCR3RA7OYLA5CNFSM4KPBV2MKYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4IKQ4SLQ>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AEA2ZVTM246QIDX62WDB2JDRA7OYLANCNFSM4KPBV2MA> .

[songfuture]

Hi,
thanks for your sharing,I am a new student of speech signal processing using deep learning.
How do you draw the picture you commented on? I'm looking forward to your reply .

[gzhu06]

I found the similar problem in my applications, the learned filters (i.e. low_hz_, band_hz_) barely changed with epochs. It's basically a mel filter bank

[ZaUt-bio]

Hi, thanks for your sharing,I am a new student of speech signal processing using deep learning. How do you draw the picture you commented on? I'm looking forward to your reply .

my question as well.
everybody, we'll be thankful if you share your code for visualization here with us.