An improved spectral subtraction method

Audio is always being affected by outside noise during the communications. Conventional spectral subtraction (CSS) is widely used due to its characteristic of low computational complexity, high real-time and easy to achieve. But its fatal flaw is that the de-noised signals contain a great deal of "music noise". The paper aims to reduce "music noise" as much as possible. Voice Activity Detection (VAD) is used to detect the starting and ending of the audio, so we use silent segment to estimate noise spectrum exactly. Furthermore, it introduces spectral decay factor to estimate noise effectively. Finally, some additional de-noising modules, such as smooth processing, threshold calculation and music noise removing, are added to the system in order to make system work stability. We use segment SNR as a evaluation of de-noising effect. Experiment results dedicate its flexibility.