Speech Processing

The processing of speech signals has a long and venerable history. As early as 1770 Wolfgang von Kempelen demonstrated his mechanical talking machine to the courts of Europe. In 1928 Homer Dudley invented the “vocoder” (voice coder) arguing that speech is specified by a few slowly varying parameters requiring only a fraction of the telephone bandwidth for transmission. A digital vocoder was first put into service in World War II for a secure telephone link connecting Roosevelt, Churchill and major military commands around the world. Exploiting properties of the human ear, such as “phase deafness” and auditory masking, perceptual coders have been demonstrated that transmit speech (and even high-quality music!) at fractional bits per sample. Applications for mobile radio, voice-email, and Internet radio abound. The success of speech recognition depends on the size of the vocabulary and the quality of the speech signal. The zero-error recognition of unrestricted, continuous speech from a noisy environment (the “electronic secretary”), however, is still in the