Syllabic spectral patterns of continuous speech have been classified into a finite alphabet through an automatic recognition process built into a “phonetic vocoder.” The classified spectra were then used to synthesize corresponding sounds approximating the original phonetic content, within broad limits. The principle of phonetic recognition can be applied to realize a maximally efficient coding of speech, considering only the phonetic content. The corresponding frequency band is about 50 cps, a 6‐fold reduction over the frequency‐band vocoder. Also, the recognition process assures that spoken words can be printed automatically as phonetic text.
Speech processed with the phonetic vocoder described here was obtained using an alphabet of only four consonants and six vowels. Nonetheless, limited vocabulary transmission was demonstrated over a band of about 100 cps indicating the feasibility of this coding method. Articulation figures of over 95% were measured using the spoken digits as test material. For a larger vocabulary of monosyllabic words, about 50% intelligibility was found in a limited test. Spectrograms of the vocoder input and output are compared.