Optimal Speech Intelligibility Improvement for Varying Car Noise Characteristics

Abstract

The present study proposes a novel method for speech intelligibility improvement by optimally shifting the formants using a trapezoidal voice transformation function. The shaping parameters of this function are determined by maximizing various performance measures using a comprehensive learning particle swarm optimization (CLPSO) algorithm. These measures include the short time objective intelligibility (STOI), perceptual evaluation of speech quality (PESQ) and signal to distortion ration (SDR). The proposed method does not requires a priori knowledge about the noise statistics in designing the voice transformation function. Although, the shaping parameters are obtained at specific SNRs, a Gaussian process (GP) regression model is trained to compute these parameters for arbitrary SNRs. The performance of the proposed method is demonstrated on various databases which include Hearing In Noise Test (HINT) a French database, NOIZEUS (ENGLISH) and CHAINS (ENGLISH) databases at different levels of engine noises arising from a running car at various speeds. The results of the investigation convincingly demonstrate that the proposed approach could improve the speech intelligibility, while preserving the quality. © 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.