Sound Processing

Abstract

A pervasive computing system is considered extremely “cool” if it can recognize speech , understand a baby cry, identify daily activities from recorded noises , or name the composer of a famous music passage and, in other words, if it can find patterns in sound recordings. But prior to that, the system needs to know more about the recorded sound. Sound is a one-dimensional time-varying signal, created by a mechanical oscillation of pressure transmitted through air. Information is normally encoded in this signal through intensity or frequency patterns of variation. Measuring sound’s intensity is a relatively straightforward operation. However, extracting frequency information from a raw sound signal is far from being easy. First, because real-world sounds are a Tower of Babel of different frequencies and amplitudes. There is not one frequency, but a whole spectrum of frequencies involved in making a sound waveform. Second, because the raw sound captured by the system is often contaminated by predictable, like the 50/60 Hz power hum, or less predictable random noise , coming, for example, from an airplane flying in the neighborhood. Some processing of the recorded signal is therefore highly needed. This chapter starts by introducing Fourier analysis , a versatile and powerful mathematical tool, used to understand sound. Moreover, two types of digital filtering techniques are presented, without getting too far in mathematical details. These are (1) the moving average filters , suitable for suppression of random noise, and (2) the frequency-selective filters, used when a priori information about the spectral characteristics of noise is available. These techniques are not restricted to sound processing. They can be applied to any one-dimensional time-varying signal, such as acceleration , temperature, pressure, and bioelectrical signals (EEG, ECG, or EMG).

C’est le ton qui fait la musique.

French proverb