PF-Net: Personalized Filter for Speaker Recognition from Raw Waveform

TL;DR

PF-Net introduces a personalized convolutional neural network architecture that learns detailed filter characteristics from raw waveforms, outperforming previous models like SincNet in convergence speed and accuracy for speaker recognition.

Contribution

The paper presents PF-Net, an improved CNN architecture that learns more detailed and personalized filters from raw speech data, enhancing speaker recognition performance.

Findings

PF-Net converges faster than standard CNN.

PF-Net outperforms SincNet in recognition accuracy.

PF-Net learns more characteristic filter parameters.

Abstract

Speaker recognition using i-vector has been replaced by speaker recognition using deep learning. Speaker recognition based on Convolutional Neural Networks (CNNs) has been widely used in recent years, which learn low-level speech representations from raw waveforms. On this basis, a CNN architecture called SincNet proposes a kind of unique convolutional layer, which has achieved band-pass filters. Compared with standard CNNs, SincNet learns the low and high cut-off frequencies of each filter. This paper proposes an improved CNNs architecture called PF-Net, which encourages the first convolutional layer to implement more personalized filters than SincNet. PF-Net parameterizes the frequency domain shape and can realize band-pass filters by learning some deformation points in frequency domain. Compared with standard CNN, PF-Net can learn the characteristics of each filter. Compared with SincNet, PF-Net can learn more characteristic parameters, instead of only low and high cut-off frequencies. This provides a personalized filter bank for different tasks. As a result, our experiments show that the PF-Net converges faster than standard CNN and performs better than SincNet. Our code is available at github.com/TAN-OpenLab/PF-NET.