12 Speech Enhancement Techniques-13 Audio Signal Processing推荐.pdfVIP

12 Speech Enhancement Techniques This chapter introduces the design and implementation of single-channel speech enhancement (or noise reduction) algorithms to enhance the speech corrupted by background noises. We will focus on the spectrum subtraction algorithm for reducing background noises. 12.1 Introduction to Noise Reduction Techniques The use of cellular/wireless phones is often found in the noisy environments such as vehicles, restaurants, shopping malls, manufacturing plants, or airports. High background noises will degrade the quality or intelligibility of speech in these applications. Excessive noise level also will degrade the performance of existing signal processing techniques, such as speech coding, speech recognition, speaker identification, and adaptive echo cancelation, which are developed with the low-noise assumption. The purpose of many speech enhancement algorithms is to reduce noise, improve speech quality, or suppress undesired interference. The noise reduction (NR) becomes increasingly important to improve voice quality in noisy environments for hands-free applications. There are three general classes of speech enhancement techniques: subtraction of interference, sup- pression of harmonic-related noises, and resynthesis using vocoders. Each technique has its own set of assumptions, advantages, and limitations. The first technique suppresses noise by subtracting the esti- mated noise spectrum, which will be discussed in Section 12.2. The second method employs fundamental frequency tracking using adaptive comb filter for reducing the periodic noises. The third technique fo- cuses on estimating speech-modeling parameters using iterative methods, and uses these parameters to resynthesize the noise-free speech. NR algorithms can be classified as single-channel or dual-channel (multiple-channel) techniques. In many real-world situations, only a single-channel system is available. A typical single-channel