Method and system for training neural network

Abstract

Disclosed are a method and a system for training a neural network. According to one embodiment, the method includes receiving a noisy signal, generating a denoised output signal, determining a signaldistortion ratio (SDR) loss function based on the denoised output signal, determining a perceptual evaluation of speech quality (PESQ) loss function based on the denoised output signal, and optimizingan overall loss function based on the PESQ loss function and the SDR loss function.

Description

[0001] [priority] [0002] This application is based on, and claims priority to, U.S. Provisional Patent Application, filed January 11, 2019, in the U.S. Patent and Trademark Office and assigned Serial No. 62 / 791,421, said U.S. Provisional Patent Application The entire contents of the application are incorporated by reference. technical field [0003] The present disclosure generally relates to signal processing. In particular, the present disclosure relates to Signal-to-Distortion Ratio (SDR) and Perceptual Speech Quality Estimation (PESQ) optimization. Background technique [0004] Recently, supervised learning based on deep neural networks has achieved substantial improvements in speech enhancement. The main difference from typical statistical approaches is that no prior assumptions about the signal model are necessary. For example, Wiener filters usually assume a Gaussian distribution of speech or noise models, which is often incorrect in real environments. In contra...

AI Technical Summary

Problems solved by technology

[0006] However, this approach has two key issues: metric mismatch and spectral mismatch

Spectral masking estimation typically minimizes the mean square error (MSE) between the clean and estimated spectral amplitudes, which is critical for signal distortion ratio (SDR) or perceived speech due to metric mismatch. Perceptual evaluation of speech quality (PESQ) maximization is not optimal

For example, it is often observed that SDR or PESQ often degrade despite the reduction of the spectral mean square error

The second spectral mismatch problem arises from the estimation of the spectral domain

As a result, the denoised spectral amplitudes are not fully reflected in the reconstructed output, which can result in substantial performance loss

Images