SincNet

Why we included this project

SincNet tackles speaker identification and verification from the raw waveform, skipping handcrafted features like MFCCs. The trick is a first convolutional layer built from parametrized sinc functions that act as band-pass filters; instead of learning every filter tap, the model only fits the low and high cutoff frequencies. That keeps the front end compact and interpretable while still adapting to your data. The repo includes a full pipeline covering data preparation, training, and evaluation, with a working speaker-id experiment on TIMIT, so it works both as a reference implementation for the paper and a starting point for your own experiments. Just know it is research code, and the ideas live on in SpeechBrain, so expect to adapt it rather than ship it unchanged.