Neural network classifiers and Principal Component Analysis for blind signal to noise ratio estimation of speech signals

A blind approach for estimating the signal to noise ratio (SNR) of a speech signal corrupted by additive noise is proposed. The method is based on a pattern recognition paradigm using various linear predictive based features, a neural network classifier and estimation combination. Blind SNR estimation is very useful in speaker identification systems in which a confidence metric is determined along with the speaker identity. The confidence metric is partially based on the mismatch between the training and testing conditions of the speaker identification system and SNR estimation is very important in evaluating the degree of this mismatch. The aim is to correctly estimate SNR values from 0 to 30 dB, a range that is both practical and crucial for speaker identification systems. Speech corrupted by additive white Gaussian noise, pink noise and two types of bandpass channel noise are investigated. The best individual feature is the vector of line spectral frequencies. Combination of the estimates of 3 features lowers the estimation error to an average of 3.69 dB for the four types of noise.

I. Introduction

Consider a speech signal corrupted by additive noise that is statistically independent of the signal. This noisy signal is characterized by a signal to noise ratio (SNR) calculated over the entire duration of the signal. In this paper, a pattern recognition approach using six linear predictive (LP) [1] derived features is used to blindly estimate the SNR of the noisy speech signal. Principal component analysis (PCA) [2] [3] of the feature vectors is shown to improve the SNR estimate and reduce the dimension of the feature vector. In addition, a further performance improvement is achieved by combining the SNR estimates generated by the six features. A multilayer perceptron (MLP) neural network [4] classifier is used.