Musical noise reduction based on spectral subtraction combined with Wiener filtering for speech communication

The goal of this paper is to propose a new technique for musical noise reduction used to alleviate some of the speech distortion introduced by the spectral subtraction (SS) process, particularly to eliminate the background musical noise of actual environment in speech communication or recognition system. The new speech enhancement approach combines spectral subtraction and the conventional Wiener filtering (CWF) in series connection to construct a two-stage hybrid system (named SS-CWF) in frequency domain to enhance the speech with additive musical noise. The noisy speech is recorded under the real background musical noise environment at a relatively lower signal-to-noise ratio. Simulation results of the proposed method, comparing with that of the conventional spectral subtraction, show better performance. The performance is evaluated by using the Log-Likelihood Ratio (LLR) measure, which is an objective evaluation measure based on linear predictive coding (LPC) techniques. Experiment results have shown that combination SS-CWF method is more robust and efficient. Meanwhile, the subjective evaluation results indicate that this method provides better speech quality with cleaner waveforms and spectrograms in time and frequency domain. Consequently, the proposed technique has complementary advantages of the spectral subtraction and Wiener filter.