SU1040512A1 - Method and device for locating word boundaries - Google Patents

Abstract

1. A method for detecting word boundaries based on the formation of two characteristic signals from an input sound signal, the first of which is formed by comparing the level of the input sound signal with the level of stationary sound noise,. characterized in that, in order to improve the accuracy and reliability of the detection of word boundaries, the second characteristic signal is formed by comparing the level of the input sound signal with the minimum level of sonorous speech sounds, comparing the duration of the characteristic signal with the maximum possible duration of a nerve-like sound disturbance, long-pause between the subsequent, u falling and occurrence of the first characteristic signal - with the maximum duration of the deaf bind, defined at the level stationary acoustic noise, long time. interval from the appearance of the second characteristic signal to the end. in the corresponding first characteristic signal with the minimum duration of the inherent speech of the time interval from the moment of the appearance of a sonorous sound to the moment of the appearance of a dumb bow and the results of the word boundaries . 2. An apparatus for carrying out the method of claim 1, comprising a driver of characteristic signals, two switches, the outputs of which are amplified with the inputs of an address counter, the output of which is connected to the first input of the memory unit, the second input of which; It is connected to the output of the first comparison unit, whose two inputs are connected. respectively, to the output and the input (L of the first, address register connected to the output of the address counter, which means that, in order to increase its accuracy and reliability, 8. a shaper of temporal ratios is entered into it, the second address register , the second unit of comparison, the third and fourth switches, the output of the driver of the characteristic: the left signals are connected to the first and second - (The inputs of the driver of the timing ratios, the four outputs of which are connected in pairs, respectively, to The third and fourth switches, the outputs of which are connected respectively to the inputs of the corresponding address registers, the output of the second address register are connected to the first input of the second comparison unit, the second input of which is connected to the output of the address counter and to the input of the second address register, the third input to the control inputs of the first, second and third high-level switches and to the third input of the block

Description

CMP pgi, the output of the second comparison unit is connected to the third input of the pulse timing generator and to the fourth input of the memory block, the second input of which is connected to the fourth

the input of the pulse timing generator, the fifth input of which I is connected to the output of the first switch and the fifth output to the third input of the first comparison unit.

The invention relates to a computing technique and can be used in input-output information devices, in particular, for speech recognition. The known method for determining the beginning and end of spoken expressions & nii with the use of adaptive delta-code modulation for encoding speech signal l. The closest to the present invention is a method for detecting word boundaries in their separate pronunciation, consisting in the formation of two characteristic signals with an input sound signal, the first one of which is fortified by exceeding the input signal level of a certain threshold level and there is a fixed on-delay time, and the second is formed when there are certain phonetic characteristics in the input signal based on two additional signals, one of which indicates the presence of deaf consonant, and the other for the presence of a slow-moving energy (word), and fixing the moment of occurrence of the first characteristic signal, which uniquely corresponds to the initial boundary of the input word, and an approximate determination of the last disappearance of the second characteristic signal, and this last disappearance is determined The final boundary of the input vein / I. The disadvantages of the method are the low accuracy of the definition of the initial boundary of words in real acoustic conditions (the presence of noise and sound noise of the pulsed type), since the moment of the occurrence of the initial boundary of the input word is taken as the moment of the appearance of the first characteristic signal, the formation of which is caused by the value of the threshold level being set, which at low for the initial boundary of the word of the moment of the appearance of a sound disturbance of the pulsed type, and at high to the possibility of delay in determining the moment of the occurrence of the initial boundary, as well as the unreliable determination of the second word of the input word, since it is determined by the last disappearance of the second characteristic signal, which, if there are extraneous conversations near the operator (the input signal energy is below the threshold for the formation of the first characteristic signal), can appear and disappear in the time interval This is the duration of the delay of the first characteristic signal loss (i.e. when the useful speech signal is already missing). The purpose of the invention is to improve the accuracy and reliability of the detection of word boundaries. This goal is achieved by comparing the second characteristic signal by comparing the level of the input sound signal with the minimum level of sonorous speech sounds, comparing the duration of the first characteristic signal with the maximum possible duration of a non-voice-like interfering sound, the duration of the pause between subsequent drops. and the appearance of the first characteristic signal — with a maximum length of a deaf ear of speech, defined at the level of stationary sound noise, the duration of the time interval from the appearance of the second characteristic signal to the end of the corresponding first characteristic signal with the minimum duration of the inherent speech time interval from the moment of occurrence sound until the moment of the appearance of a deaf link and the results of the comparison judge the appearance of the initial and final boundaries of words. In an apparatus for implementing a word boundary detection method comprising a driver of characteristic signals, two switches, the hrs of which are connected to the inputs of an address counter, the output of which is connected to the first input of the memory unit, the second input of which is connected to the OUTPUT of the first comparison unit, the two inputs of which respectively, to the output and the input of the first register of the address connected to the output of the address counter, a shaper of temporal relations of the second register of the address is entered, the second block, compare The third and fourth switches, the outputs of the driver of the characteristic signals are connected to the first and second inputs of the pulse shaper – time ratio of scientific research institutes, four outputs of which are connected in pairs to the inputs of the third and fourth switches, respectively, whose outputs are connected respectively to the inputs of the corresponding address registers, the output of the second address register is connected to the first input of the second comparison unit, the second input of which is connected to the output of the address counter and to the input of the second register a and the third input to the control inputs of the first, second and third switches and to the third input of the memory unit; the output of the second comparison unit is connected to the third input of the temporal relationship generator and to the fourth input of the memory unit, the second output of which is connected to the fourth input there is a formation of temporal ratios, the floating input of which is connected to the output of the first KOMMyTatopaj and the fifth highest move - to the third input of the first comparison unit. iHa FIG. 1 is a block diagram of the device; in fig. 2 - time diagram of the device. The device contains a driver 1 of the characteristic signals, two switches 2 and 3, on which two kinds of clock signals are supplied, which are connected to the inputs of the address 4, the output of which is connected to the first control input to the memory block 5, to the information inputs of which pitch (ha-. actuation of the input signal, the second control input of which is averaged to the output of the first comparison unit 6, two inputs of which are connected respectively to the output and to the input of the address register 7 connected to the 4-address counter. Devices it also contains the driver of 8 time-ratio pulses, the second address register 9, the third and fourth switches 1O and 11, and the second comparison unit 12. The proposed method is as follows. To define the boundaries of a word, the following temporal characteristics of the audio signal are used: T is the maximum duration of a non-speech-like sound disturbance at the input; Tn is the maximum duration of the section of the deaf link of the speech signal, the level of which does not exceed the level of stationary sound noise; Ta is the minimum duration of the inherent speech of the time interval from the moment of the appearance of a sonorous sound to the appearance of a deaf bow. The first by the junction of the first characteristic signal is considered the probable initial boundary of the input word, and then the truth of this initial boundary is analyzed. If the duration of the first characteristic signal is less than the time T | or more than this time, but at the same time after the disappearance this signal is absent more than 2, then the fixed initial border is considered false, and when the first characteristic signal is reappeared, a new one is fixed, probably the initial one is on the boundary of the input word. The same thing will happen if, before determining the true initial boundary of the word, after the appearance of the second xapeiB signal, the first characteristic signal disappears after a time less than Tj. Otherwise, the fixed probable initial boundary is considered true. The final face of the input word i is defined as follows. If, after determining the initial boundary of the input word, the first characteristic signal disappears at some instant, then this moment is considered the final boundary of the input word, and then the truth of the fixed final boundary is analyzed. If at the same time the first characteristic signal is absent longer than T2 or during the time T - t it appears by a time t less than the time T2, then the possible final rpajpmia of the input word is considered to be the true final limit. Otherwise, the fixed final limit is considered false, and the search for the final boundary is performed again. The device works as follows. The shaper 1 of the characteristic signals forms two signal lines (: 1d and | e) by exceeding two different threshold levels. The first level (lower) corresponds to the signal fd j and the second (upper) signal - 1d. Each of the signals 1у and - {d is present during the time when the input level from the corresponding threshold level is exceeded. The initial installation of the device (the installation chain is not shown) blocks the address registers 7 and 9 and the switch 3, the latter through an inverter (not shown), and the switches 2, 1О and 11 unlock, the BS Synchroimpu (not shown) with the frequency of recording information in block 5 is received through the switch 2 in the driver 8 and on the counter 4 addresses, address codes from which are sent to the registers 7 and 9 addresses, blocks 6 and 12 of the comparison and in block 5, which works in the recording mode input information ation representing the characteristic of an input word tics in digital form ,, When the occurrence of the characteristic signal shaper 1. The signals {d timing generator 8 generates a pulse signal reception starting border which, via a switch 11 comprises PE. Gistr 9 addresses, receiving 1fi is the current address from the counter 4 addresses. The received address is the address of the probable starting word boundary. If the signal duration 1d is shorter than T or more than this time, but after this signal disappears, there is no signal longer than T, then when a new signal appears, shaper 8 will generate a new signal. The initial boundary will be received, which will be received in address register 9. the address of the new probable initial boundary of the word .. The new address will also be accepted if, when switch 11 is unlocked after the occurrence, the signal I d disappears through the signal time less than T. If at any moment after the reception by the register 9 of the address of the address of the probable initial boundary will generate a 1d signal, after which the 1d signal will not disappear through; the time is less than Tj, the driver 8 will generate the Initial Block signal, which blocks the switch 11 until the beginning of the determination of the boundaries of the new O input word. At the same time, the address stored in address register 9 is an address with the address of the true initial boundary of the word, If after the formation. signal; The initial blocking signal i d disappears at some point in time, then the form: the consumer 8 will generate a signal Reception of the final limit, which through switch 10 will activate the address register 7, which receives the current address from the 4 address counter, the received address is an address the final word boundary, If this signal 1d fails to reach earlier than, after time T, or during time V / 2.41 turns out to be t-, time L is less than T, then driver 8 will generate a final lock signal that will block switch 10 before the determination new borders about the input word. In this case, the address stored in the address register 7 is the address of the true end boundary of the word. Otherwise, t, e, when a 4d signal appears after the address of the probable final boundary of the word is received in register 7, after a time less than T, time greater than T, when the signal disappears, the shaper 8 will form a new signal. Receiving the final limit, according to which the address of the new probable final word boundary will be received in address register 7, and further, the described process will be repeated until

spraying the true finite word boundary.

The final blocking signal, in addition to the action already indicated for blocking switch 10, blocks the record in block 5, locks switch 2, unlocks the comparison block 12, and also switch 3, in this case the synophone pulses CHf2 from the syn: recognition with the frequency of reading information from. block 5 through switch 3 start

arrive at the counter 4 addresses.

If the address stored in address register 9 matches the current address,

(Coming from the 4-address counter, - the comparison tab 12 generates a read signal, through which unit 5 is transferred to the information reading mode, which then enters the recognition device, and shaper 8 generates an enable signal, which unlocks the comparison unit 6.

If the address stored in the address register 7 coincides with the current address coming from the 4 address counter, the comparison unit 6 generates a read signal Kf at which the driver. 8 is set to its initial state, block 6 and 12 blocks are blocking & switch 3 and unlocking switch 2, as well as switching block 5 to the input information recording mode.

Thus, from block 5, the characteristics of the input word are read in its true boundaries, and the device is ready to determine the boundaries of the new input word.

Claims (2)

1. A method of detecting word boundaries, based on the formation of two characteristic signals from the input audio signal, the first of which is formed by comparing the level of the input audio signal with the level of stationary sound noise,. characterized in that, in order to increase the accuracy and reliability of detecting word boundaries, the second characteristic signal is formed by comparing the level of the input audio signal with the minimum level of voiced sounds of speech, comparing the duration of the characteristic signal with the maximum possible duration of an incoherent audio noise, duration * pauses between subsequent loss and the appearance of the first characteristic signal - with a maximum duration of a dull speech bow, determined at the level of stationary sound noise, the duration of the time interval from the appearance of the second characteristic signal to the end of the corresponding first hare-hermetic signal with the minimum duration of the speech inherent in the time interval from the moment of the appearance of voiced sound to the moment of the appearance of a dull bow and the results of comparison judge the appearance of the initial and final word boundaries, 2. A device for implementing the method according to π. 1, comprising a driver of characteristic signals, two switches, the outputs of which are connected to the inputs of an address counter, the output of which is connected to the first input of the memory block, the second input of which is connected to the output of the first block. comparison, the two inputs of which I connect. yen, respectively, to the output and input of the first register of the address connected to the output of the address counter, which means that, in order to increase its accuracy and reliability, a pulse shaper ₍ modern time relations, the second register) is introduced into it addresses, the second comparison unit, the third and fourth switches, and the outputs of the driver of characteristic signals are connected to the first and second inputs of the driver of pulses of time relationships, the four outputs of which are paired respectively to the inputs of the third о and the fourth switch, the outputs of which are connected respectively to the inputs of the corresponding address registers, the output of the second address register is connected to the first input of the second comparison unit, the second input of which is connected to the output of the counter.a address and to the input of the second address register, the third input to the control inputs of the first, second, and third | Switches and to the third input of the to block of memory, the output of the second comparison unit is connected to the third input of the pulse former of the time ratios and to the fourth input of the memory block, Ora input of which is connected to the fourth input of the timing pulse, '{.pyaty input of which is connected to the output of the first switch, and the fifth output - to the third input of the first comparator unit.