| 摘要 |
A system for transforming speech signals into phoneme identification signals includes a simple inflection point detecting circuit (8) that produces "1s" and "0s" during positive and negative pressure wave portions of the analog speech signal, respectively. For "voiced" portions of the speech signals, the major peaks of each pitch cycle of speech signal are detected (13) to produce pitch cycles that are analyzed by a microcomputer (10) to detect the onset of each pitch cycle and to detect and store the duration of successive "1" and "0" levels. For each pitch cycle, an input vector is formed of the "1" level durations most closely following the onset of the pitch cycle. This input vector is corrected by a quantity derived by determining the mismatching between the present pitch of the speaker's voice and the resonant frequencies of the configuration of the speaker's mouth cavity. Velocity and acceleration representing the duration of the two "1" levels at the beginning of each pitch cycle are analyzed by the microcomputer (10) to provide demarcation between the beginning and end of each phoneme. For fricative portions of the speech signal, a fricative vector is computed and compared to a fricative map to identify a fricative. Silence intervals are indicated by the lack of a substantial running average "1" level by the inflection point detector (8). Plosive phonemes are partially identified by the durations of preceding silent intervals. Long duration phonemes use the slope between the moving "1" and "0" duration values to assist in identification procedure. Identification of all phonemic time-sequence vectors is accomplished by using various maps of imperically derived reference vector domains. |
