Speech Enhancement Based on Adaptive Harmonic Model Using Maximum Likelihood Method

Speech enhancement is a hot topic in the modern society due to its extensive applications such as automatic speech recognition, mobile communication, etc. Spectral subtraction is a very valid and direct denoising algorithm, but it is still needed to be further developed due to complex application. In this paper, we firstly assume adaptive harmonic model to model the speech signal. Speech enhancement is then achieved by spectral subtraction. To further enhance the speech intelligibility, we attempt to estimate the harmonic model. Maximum likelihood method is considered to derive the phase and amplitude update formulae of the modelled harmonic signal, which are aimed to depress the distortion due to spectral subtraction. Different from convention spectral subtraction, both the amplitude and phase parameters of adaptive harmonic model are combined to be updated. We assume the additive noise is correlated along the time sequence. By the optimal solution of maximum likelihood method, we obtain the updated version of speech enhancement. Simulation results show the effectiveness of the new algorithm, and further improvement of spectral subtraction.