Rationale for dividing speech into chunks of 200ms with 10ms overlap

[vikigenius]

Hi, I am trying to check the performance of SincNet on VoxCeleb dataset. I am wondering about the rationale of extracting chunks of 200ms windows of signals during training, and also the 10ms overlap that you have in test? Does the model depend on this?

Can I use longer chunks like 3s of audio like the VoxCeleb paper seems to be using? Seeing that VoxCeleb is a much larger dataset?

[mravanelli]

Hi, the size and the shift of the context window are just hyperparameters of the system and their best values could be very task-dependent. In the paper we tuned these two hyperparameters on the addressed tasks and we have observed the best performance with them. This doesn't necessarily mean that are good for all the speech tasks...

…

On Fri, 29 Mar 2019 at 17:52, Vikash ***@***.***> wrote: Hi, I am trying to check the performance of SincNet on VoxCeleb dataset. I am wondering about the rationale of extracting chunks of 200ms windows of signals during training, and also the 10ms overlap that you have in test? Does the model depend on this? Can I use longer chunks like 3s of audio like the VoxCeleb paper seems to be using? Seeing that VoxCeleb is a much larger dataset? — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#24>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQGs1h6FIOlHKLIi6qjXA5wdxXaRhqQ-ks5vbos1gaJpZM4cTTdK> .

[vikigenius]

Thanks for the comment. So if i understand correctly you train the model to identify the speaker on a 200ms random chunk of the audio. While testing you take multiple such chunks and classify all of them and vote on the best ones? So in conclusion your model cannot handle variable length signals for now?

Also, during generating training batches, you seem to be multiplying an amplitude factor randomly chosen from (0.8, 1.2) to the signal. Is there any rationale behind this?

[mravanelli]

More precisely we are averaging all the probabilities and finally vote for the speaker with the highest probability. Due to the averaging operation (that is differentiable) we can manage sequence of arbitrary length. The reason why we perform the amplitude multiplication is to add a bit of data augmentation while training the system (the system can always see the different amplitudes for a different signal). This has a minor effect actually... Mirco

…

On Fri, 29 Mar 2019 at 18:16, Vikash ***@***.***> wrote: Thanks for the comment. So if i understand correctly you train the model to identify the speaker on a 200ms random chunk of the audio. While testing you take multiple such chunks and classify all of them and vote on the best ones? So in conclusion your model cannot handle variable length signals for now? Also, during generating training batches, you seem to be multiplying an amplitude factor randomly chosen from (0.8, 1.2) to the signal. Is there any rationale behind this? — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#24 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AQGs1uhksDAemjgKzmJ-273X6NBxp_nrks5vbpDQgaJpZM4cTTdK> .