Compensate multiple distortions for speaker recognition systems

Abstract

The performance of speaker recognition systems reduces dramatically in severe conditions in the presence of additive noise and/or reverberation. In some cases, there is only one kind of domain mismatch like additive noise or reverberation, but in many cases, there are more than one distortion. Finding a solution for domain adaptation in the presence of different distortions is a challenge. In this paper we investigate the situation in which there is none, one or more of the following distortions: early reverberation, full reverberation, additive noise. We propose two configurations to compensate for these distortions. In the first one a specific denoising autoencoder is used for each distortion. In the second configuration, a denoising autoencoder is used to compensate for all of these distortions simultaneously. Our experiments show that, in the co-existence of noise and reverberation, the second configuration gives better results. For example, with the second configuration we obtained 76.6% relative improvement of EER for utterances longer than 12 seconds. For other situations in the presence of only one distortion, the second configuration gives almost the same results achieved by using a specific model for each distortion.

Learning Noise Robust ResNet-Based Speaker Embedding for Speaker Recognition

“…Despite having good results for simulated noises this work doesn't include real noise and reverberation. In another work two configurations are proposed to denoise different kinds of distortions such as noise, early reverberation, and late reverberation [7]. In this paper also the capability of doing noise compensation is not explored in real environments.…”

“…The problem of noise and reverberation is addressed at different levels of speaker recognition systems, including signal level [4], feature level [5], speaker modeling level [6], x-vector level [7] and scoring technique adaptation [8]. Data augmen-tation is another approach to making the speaker recognition systems robust against noise.…”

“…Noise compensation in x-vector level, the estimation of clean x-vector from its corresponding noisy version, by doing a transformation between pairs of noisy/clean x-vectors is another approach that performed well in the compensation of artificial noise and reverberation [9,3]. Although this approach performs well in some cases [7], it doesn't bring a significant improvement with all speaker embedding systems [10] and in all environments. The behavior of different speaker embedding systems is different because they just consider the speaker classification accuracy (inter-speaker and intra-speaker distance) during optimization and they don't put an explicit constraint on the noise impact.…”

Barlow Twins self-supervised learning for robust speaker recognition

“…Although, the DNN-based speaker embedding systems have given a degree of robustness against acoustic noises, there is a significant degradation of their performance in the presence of background noise, reverberation and other variabilities [4] [5] [6]. Various approaches have been proposed to handle these variabilities in different parts of the system such as: signal level [7], feature level [8], speaker modeling level [9], xvector level [6] and scoring technique level [10]. Addressing the variabilities at each step has its own advantage and disadvantages in terms of data, computational resources, efficiency, etc.…”

A Comprehensive Exploration of Noise Robustness and Noise Compensation in ResNet and TDNN-based Speaker Recognition Systems

“…The robustness of the DNN-based speaker recognition (SR) systems in general and specifically their robustness against environment variabilities such as additive noise, reverberation, and far-distance recording device has made them more promising. Several strategies such as data argumentation [4], and noise compensation [5], [6] are explored to make the TDNNbased SR systems more robust against noise and reverberation and other variabilities. The previous research shows the weakness of TDNN-based SRs against noise and reverberation distortions.…”