Research On Speech Masking System

In order to protect near-end talker’s privacy information during a voice call, the speech masking system is proposed. The system generates time-reversed masker with near-end talker’s speech, which can make the voice information masked to achieve the purpose of security of near-end talker’s privacy information.The main work contents are accomplished as follows:1. According to studies of time-reversed speech in recent years, we put forward the speech masking system, which mainly consists of two modules:the masker echo canceller and the masking signal generator. Masking signal generator gives masking signal by making near-end talker’s speech time-reversed.2. Masker echo cancellation algorithms were studied. We compared SNLMS algorithm with LMS algorithm and NLMS algorithm. The results show that the former can achieve better echo signal attenuation so that it is applicable to the speech masking system.3. We designed a new window function, which can improve the quality of time-reversed masker. Time-reversed transformation method with voice detection and frequency modulation method were proposed to generate masker, which can also improve auditory comfort of masker.4. Subjective evaluation of the masking quality for the masker was checked by audiometry experiment. The results show that masking degree increases with rising of the length of time-reversed frame in the range of50-400ms under normal voice speed. Increasing energy of the masker or accelerating the voice speed can improve the masking quality of the system. Based on the audiometry experiment, phoneme confusion matrixes were gived and then the masking cases of consonants, vowels and tones were respectively analysed.5. Objective evaluation algorithms of the masking quality for the masker were researched. NENC algorithms, PESQ algorithm and Glimpse algorithm were compared in objective evaluation experiment. The results show that, NENC algorithm is more effective than other algorithms for objective evaluation of masker.