发明名称 WORD BOUNDARY DETECTOR FOR SPEECH ERCOGNITION EQUIPMENT
摘要 1536965 Detecting the ends of words THRESHOLD TECHNOLOGY Inc 5 March 1976 [10 March 1975] 8841/76 Heading G4R Apparatus for detecting the ends of words receives input signals representing speech, has a generator which generates a first feature signal which indicates the presence of a word, a store which stores feature signals which occur during the presence of the first feature signal and a generator which generates a second feature signal known to occur at least just before the end of the word and a device determining the last occurrence of the second feature signal during the presence of the first feature signal to indicate the end of the word. Signals representing speech are fed to preprocessor circuitry 50 which, for instance, contains a plurality of band pass filters the output of each of which is rectified and passed through a low pass filter (Fig. 2a, not shown). The rectified signals are sampled by a multiplexer in the embodiment described, passed through a logarithmic amplifier and are de multiplexed. The demultiplexed signals feed Feature Extraction Circuitry 60 which can identify 31 features in the speech. The rectified signals E1-E19 from Preprocessor Circuitry 50 also feed circuitry 110 (Fig. 5). Selected signals feed Operational Amplifiers 121-124 which each form the sums of its inputs and which feed Operational Amplifier 125 which produces an output indicating the presence of a word when the sum of its inputs exceeds a predetermined value. Amplifier 125 feeds an integrator 126 arranged such that its output rises immediately the output from Amplifier 125 rises and such that its output ceases a substantial period, e.g. 100 m.secs. after the output from Amplifier 125 ceases to prevent pauses in a word from indicating the end of a word. Operational Amplifiers 131, 132 each feed two integrators 133, 135 and 134, 136 the second integrator of each pair also being fed by the first Integrators 133, 134 have time constant fifteen times greater than the time constants of integrators 135, 136 so that a relatively fast dropoff of energy in the appropriate spectral energy bands causes a logical 1 input to a NOR gate 137 to turn output fB to zero. Signals fC or fD can be produced as shown. The signals fA, fB, fC on lines 120a, 130a, 140a (Fig. 1) feed a word isolator 150 which also receives the 31 feature signals from Circuitry 60. The feature signals are sampled e.g. every millisecond and the results stored as '1's and '0's to form a matrix (Fig. 3, not shown). Clock signals pass gate 153 at the start of signal FA and are stopped at the end of signal FA. The clock signals increment address generator 152 supplying addresses to a Random Access Memory 154 and also feeding an address store. The address store is enabled by signals FB or FC and each new address supplied erases the previous in store so that each address in the store represents the time of occurrence of the enabling signal FB, FC. When the end of FA is detected the address generator is reset and a gate 157 is enabled which passes high frequency clock pulses to the address generator to cause high speed read out of memory 154. A coincidence detector is enabled via delay D. The read out continues until the address generator value equals the value in the address store, i.e. the address of which the last signal FB, FC terminated. A signal on line 158x then terminates read out. The matrix read out is normalized and (Fig. 4, not shown) compared with stored values to determine the word.
申请公布号 GB1536965(A) 申请公布日期 1978.12.29
申请号 GB19760008841 申请日期 1976.03.05
申请人 THRESHOLD TECH INC 发明人
分类号 G10L11/02;G10L15/00;(IPC1-7):G10L1/00 主分类号 G10L11/02
代理机构 代理人
主权项
地址