Time-Frequency Filtering of Speech Signals in Hands-Free Telephone Systems

Time-varying filtering of noisy speech signals is a very attractive challenge, with the main question: how does the most appropriate time-varying filter scheme look like? Speech signals are of highly nonstationary and multicomponent nature. If we deal with filtering of noisy speech signals, as they occur in hands-free telephone systems, then the desired scheme would provide a signal-to-noise ratio (SNR) greater than approximately 12dB. At the same time it should be suited for real-time implementation, with time delay less than 39ms for mobile telephony, and 2ms for circuitswitched telephony. The second requirement can cause additional difficulties and restrictions on finding an appropriate time-varying filter procedure. The most commonly used approach in the filtering of speech signals is the so called quasistationary approach, where it is assumed that the signal is stationary in the time interval T , with T between 20ms and 40ms being often used [1], [14]. In this interval of time, classical speech enhancement schemes such as those given in Table I are used [6], [16]. The noise is reduced by applying frequencydependent suppression factors according to the various filtering rules given in Table I. From the aspect of time-varying filtering we can say that this technique is quasi time-varying filtering. Thus, we have a sliding window of duration T along the signal where the filtering is performed after every T or after every T/2 (the second case is used in an overlap-add scheme in order to avoid block effects). From the point of view of time-frequency analysis, having in mind the high nonstation-