A Complex Plane Spectral Subtraction Method for Vehicle Interior Speaker Recognition Systems

The performance of speaker recognition systems drops significantly under vehicle interior noisy conditions. The traditional spectral subtraction algorithm is a popular method for noise reduction but suffers from musical noise distortion. To compensate for this problem, we proposed a new complex plane approach for spectral subtraction. The proposed method rectifies the incorrect assumption about the cross terms involving the phase difference between the clean speech and noise signals being zero. We assessed the novel speech enhancement method using utterances corrupted by volvo car-interior noises in different signal-to-noise ratios (SNRs) levels. Spectrogram analysis results show that the complex plane method performs significantly better than the traditional spectral subtraction algorithm. In addition, the new speech enhancement method was evaluated by three groups of speaker data from the TIMIT database with a 32-order Gaussian Mixture Model (GMM) based speaker recognition system. The experiments revealed that the proposed approach performed better than the traditional spectral subtraction in the pre-processing stage of speaker recognition systems1.