search

zkmkarlsruhe / language-identification

Spoken Language Identification on Common Voice and AudioSet using Deep Learning
Other
35 stars 7 forks source link

Question about preprocessing #4

Open completelyboofyblitzed opened 1 year ago

completelyboofyblitzed commented 1 year ago

First of all, thank you for sharing your models! I was wondering if you trim the audio from silince during preprocessing, because your model works pretty well where there's voice right away but if someone was lingering in the beginning of the audio, your model predicts noise. How do you think one should approach the issue?

bytosaur commented 1 year ago

hey there, During preprocessing i actually did cut out the voice part since the samples can be longer than 5 seconds and I wanted to make sure that the voice is captured when cutting the samples. During training i only did a small time shift in order not to throw away the voice. However, during inference I use the same approach. Voice activity determines the start of the sample. It's a good thought and the model is probably biased towards that... I don't know how to solve it without increasing the sample length...

danomatika commented 1 year ago

We have considered adding some preprocessing to help filter out background noise and also better onset detection. @bytosaur has some ideas on this.

enohp ym morf tnes

Dan Wilcox danomatika.com robotcowboy.com

On Sep 5, 2022, at 4:18 PM, Marina @.***> wrote:

 First of all, thank you for sharing your models! I was wondering if you trim the audio from silince during preprocessing, because your model works pretty well where there's voice right away but if someone was lingering in the beginning of the audio, your model predicts noise. How do you think one should approach the issue?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.