A Wiener Filter Speech Enhancement Algorithm With High Intelligibility

With the rapid development of the information society, smart phones and human-computer voice conversation devices have been widely used, so the speech signal gets more and more attention. However, the speech signal in generation, transmission, processing and receiving will be inevitably affected by the surrounding environment and transmission medium, which make the speech contains noise. Serious pollution will affect the quality and intelligibility of the speech signal, which results in people or receiving speech equipment can’t understand the meaning of speech. Therefore, speech enhancement technology should be used to separate the clean speech from the noisy speech, in order to filter out the noise. The traditional speech enhancement algorithms pay more attention to speech quality, thus the enhanced speech has a higher signal to noise ratio. Compared with the noisy speech, the enhanced speech intelligibility has not been improved effectively. This is because the traditional enhancement algorithms will also filter out the useful speech signal when filtering out noise, result in speech distortions.Since the Wiener filtering can significantly improve speech quality and make the enhanced speech contains less music noise. This paper proposed an improved higher intelligibility speech enhancement algorithm based on the Wiener filtering, in order to improve the enhanced speech intelligibility. The enhanced speech signal is more likely to be understood by person or speech devices.This paper describes the knowledge of speech signal, the characteristics of human auditory and the features of noise signal. Then it systematically introduces four major categories of speech enhancement algorithms. Summarizes the related evaluate methods for the enhanced speech, including subjective hearing evaluation method, objective evaluation methods of speech quality and speech intelligibility.According to the derivative process of the Wiener filtering, we get the Wiener gain function. Then the Wiener filtering method based on a priori SNR estimation is described, the process of this method is simple, and the enhanced speech quality improved significantly. Through simulation experiments of sentences and consonants corpus, we get that this method can improve speech quality, but can’t truly improve intelligibility of the enhanced speech. Analysis the reason why speech enhancement algorithms without improve speech intelligibility, and study from the signal-to-residual spectrum ratio, we find that the enhanced speech magnitude spectrum exists attenuation distortion and amplification distortion, when the amplitude distortion spectrum is greater than6.02dB can seriously affect speech intelligibility. Comparing the clean speech magnitude spectrum with the enhanced magnitude spectrum in experiment, then removing the amplification distortion greater than6.02dB regions, we can get that enhanced speech intelligibility and quality have been improved significantly compare with noisy speech.There is no clean speech in reality environment, which requires modified a priori SNR. Modified a priori SNR when it is less than-10dB SNR, and modified the gain function of filtering algorithm, then through the existing conditions determine the greater than6.02dB magnitude spectrum regions, then constraints this regions. Ultimately, we can get the improved higher intelligibility Wiener filtering speech enhancement algorithms. Through simulation experiments of sentences and consonants confirm the improved algorithm does really improve speech intelligibility.