Speaker verification in a noisy environment by enhancing the speech signal using various approaches of spectral subtraction

Abstract

Enhancement of speech signal degraded by several types of noise is a topic of interest for last many years. The main aim of speech enhancement algorithm is to improve the quality and/or intelligibility of the noisy speech signals by using various techniques and algorithms. Among the all available methods, the spectral subtraction algorithm is the one of the first algorithm proposed for removing additive background noise. This paper describes various noise reduction algorithms to analyse the performance of speaker verification in noisy environment. Speaker verification involves processing the speech signal and authenticating the speaker. The speaker verification process is affected by various factors such as noise, channel mismatch, health condition of the speaker, aging, emotion, fatigue etc. Speech signals from the uncontrolled environment may contain noise along with required speech components, speech signals are degraded by this noise and it renders the performance of speaker verification systems unacceptable. We enhance the speech signal by using the basic spectral subtraction and the modified approaches of spectral subtraction like spectral Subtraction with over subtraction, non-linear spectral subtraction and multiband spectral subtraction. The speech signal has been enhanced by each of the above methods then speaker models are trained with support vector machine using mel frequency cepstral coefficient of each speaker in the system. The performance of the speaker verification system is analysed with enhanced speech signal.

Noise Robust Speaker Recognition Based on Adaptive Frame Weighting in GMM for i-Vector Extraction

“…In realistic scenarios, due to the mismatch between training data (usually clean) and testing data (usually contaminated by unknown noises with different intensities), the recognition performance of the system will degrade drastically. To deal with this problem, researchers have proposed many methods to enhance the quality of features at different levels, such as speech signal enhancement by statistical processing [7]- [9] or by deep learning [10]- [12], DNN based cepstral feature de-noising [13], i-vector de-noising [14], multi-task adversarial network (MAN) for extracting noise-invariant bottleneck (BN) features [15] etc. Backend classifiers have also been investigated by parallel model combination [16], robust variants of the Probabilistic Linear Discriminative Analysis (PLDA) model [17], [18], multi-style training [19], [20], etc.…”