Speech enhancement in car noise environment using harmonic noise model
Student thesis: Master's Thesis
Related Research Unit(s)
Effective speech enhancement method is highly desirable for restoring noise-corrupted speech to improve the comfort of hands-free voice conversations in cars. The most intractable challenge of this application is attributed to a robust suppression solution for fast-varying car noise. Conventional methods generally estimate a suppression gain based on some forms of segmental signal-to-noise ratio (SNR), which is not so reliable in highly non-uniform and non-stationary car noise environment. As a result, some processing artifacts, namely "musical tones" and some strong low frequency engine noise will reside in the enhanced speech. In this research, a novel speech enhancement method is proposed to tackle these residuals. The proposed method is based on an analysis-synthesis approach via harmonic noise model (HNM). A framework for the application of HNM in speech enhancement is proposed. The effect of using pre-cleaning techniques is investigated. Algorithms for the robust estimation of model parameters, including pitch frequency, harmonic magnitudes and phases, mixing function, and voiced/unvoiced (V/UV) decision in car noise environment are developed. The quality assessment of the proposed method is done using subjective listening tests and objective quality measures. Both evaluation results show that the proposed method surpasses conventional methods in low SNR environments, especially in terms of noise suppression capability.
- Automobiles, Speech processing systems, Noise