Method and device for the dynamic modification and the analysis of the electrical signals representing the sound

摘要

978, 303. Automatic speech recognition. J. A. DREYFUS. Dec. 8, 1961 [Dec. 8, 1960], No. 44124/61. Heading G4R. [Also in Division H4] A circuit for compressing the amplitude of a speech signal having components of different frequencies comprises a main channel 5 Fig. 1 having a controlled gain amplifier 9 and an auxiliary channel 10 to derive a signal for controlling the gain of the amplifier 9 having a band pass filter 11 passing a portion of the speech signal so that the amplitude of the main channel signal is modified only by the selected portion of the signal. The auxiliary channel also comprises an amplifier 10a a rectifier 12 and a low-pass smoothing filter 14. The amplifier 9 of the main channel is preceded by high-pass filter 8, the passing range of which excludes the range of the low-pass filter 14. Alternatively the filter 11 may be a high-pass filter with a lower pass frequency higher than that of filter 8. In Fig. 2 the auxiliary channel 30 is taken from the output of the high-pass filter 28. The controlling signal is again applied to the amplifier 29 but a delay 34 is inserted in the main channel to delay the main signal. According to the setting of this delay the gain control signal appears after, simultaneously with or before the associated part of the main channel signal. The response of the circuit varies accordingly as shown in Fig. 8 at 41, 43 and 44 respectively. The form of Fig. 1 may have an adjustable integrating unit between the output of the auxiliary channel and the amplifier 9 which has the same effect as the delay 34. By suitable adjustment the control signal may be made late, synchronomous or early. Two or more compressor circuits may be connected in series to obtain the desired compression. Speech analysis :- The compressor circuit is used as shown in Fig. 10 in a system adapted to analyse speech signals from a microphone 101. The compressor sections 105, 106 and 107 may each comprise two compressors of the Fig. 2 type connected in series. The three sections each have characteristics designed to facilitate the operation of the analysis sections to which they are connected. The compressor circuits have switches 115, 116 which enable the auxiliary channel to be connected as in Fig. 1, the delay 110 being by-passed by switches 117, 118. The output of compressor section 106 is applied to blocks 121,122 containing twenty narrow band-pass filters each adapted to respond to a different component frequency of the speech signal. Each filter output a, Fig. 12 passes to a rectifier 123 (b, Fig. 12) and a low-pass filter 124. The output of the filter 124 passes on line 125 via a low-pass filter 126 and a delay (not shown) to an analogue-to-digital converter 128 which produces an output on one of three leads 0, 1, 2, Fig. 12 according to the level of the signal. The output of filter 124 also passes to circuits for determining the rate of rise or fall of the particular frequency component. The circuits comprise differentiators 131, 132 oppositely poled by diodes to respond to rising and falling amplitudes e and f, Fig. 12 and band-pass filters 133,133SP1/SP which oscillate as a result of the applied pulses. Each filter may consist of two separate filters, treated weparately as shown at g and h, Fig. 12 or a single filter as shown at 1, Fig. 12. The outputs of the filters are rectified at 134, 134SP1/SP to produce amplitude signals i,k,m Fig. 12 which are applied to converters 136,136SP1/SP producing outputs on the appropriate ones of three leads. The amplitude of the signals i, k and m are a measure of the rate of rise or fall of the corresponding frequency component. The output from the filter 126 is delayed an amount which brings it into time coincidence with the rate of change signals at the circuit 127. The compressor section 105 is connected to a group of filters 186 each connected to differentiation networks 184 similar to those 130 described above. They are designed to detect fluctuating sounds such as the roller "r", "ch" "f" and "s" which have the form of an amplitude modulation of a higher frequency signal. The filters 190 are excited by the fluctuations accentuated by the differentiating networks and the output is rectified at 191 to produce an amplitude signal converted at 193 to a signal on one of three leads for application to circuit 127. The compressers 105 are designed to give a late controlling signal so that an "overshoot" 41 Fig. 4 is obtained to increase the fluctuation. Filters in block 195 respond to fairly low frequencies to determine the fundamental frequency of the voice of the speaker. The outputs are rectified and converted at 201 to a signal on one of three leads. Since the frequencies which characterise all vowels and similar sounds depend to some extent on the fundamental frequency a correction can be applied in the circuit 127 to assist in identifying particular sounds. The outputs of all the filters are sampled and the lowest frequency band to produce a "1" or "2" level at the output of the converter 201 is taken as the fundamental frequency. The existence of frequencies immediately below this frequency is detected by an amplifier 202 which is connected to the filter of immediately lower range. A signal at the output of converter 203 indicating the presence of this lower frequency is useful for the identification of transient sounds such as "j", "z" and "v". The third compressor channel 107 is connected to filters in blocks 220 and 221 having the passing frequencies indicated which are each connected to a circuit 224 for identifying explosive sounds such as "p", "k" and "t". The low-pass filters 192, 135 &c. are connected to a timer unit 227 which enables the circuit 127 on completion of each phonetic sound to actuate the appropriate key of a typewriter 230. The output of the system may be used to print musical notes as at 231. Other symbons expressing rhythm and volume may be typed at 232 and 233.