RU2691603C1 - Method of separating speech and pauses by analyzing values of interference correlation function and signal and interference mixture - Google Patents

Abstract

FIELD: radio engineering and communications.SUBSTANCE: invention relates to radio engineering and can be used in devices for transmitting voice information. Technical result is achieved by the fact that in this method three stages of comparison of values of interference correlation function or mixture of signal noise with threshold are used. At the first stage, for the interval, which contains only interference, the threshold value is determined from the correlation function of the interference. This value is specified during “sliding window“ movement. At the second stage detection of occurrence of a signal with low value of accuracy of determination of time of its occurrence and high value of probability of signal detection, determined by value of duration and value of displacement of "sliding window" is carried out. At the third stage, the signal position is refined, wherein the probability of the correct decision on the presence of the signal is not lower than a given level, which is achieved by using the corresponding values of duration and displacement of the "sliding window".EFFECT: high accuracy of determining the moment of occurrence of a speech signal in the presence of external acoustic interference.1 cl, 2 dwg, 2 tbl

Description

The invention relates to the field of technology of transmission and transmission of speech information and may find application in communication devices and devices for loud-speaking communication.

A device for the selection of acoustic signals in communication channels, described in patent RU 2171549 H04Q 1/46, is known. The invention relates to telecommunications, in particular to automatic means of receiving signals of tone signaling in multi-channel communication systems, and can be used, for example, to detect acoustic signals (AU) in telephone channels. The operation is based on the calculation of a number of decisive statistics, which are distinguishing features in recognizing an information speaker from channel noise and spurious speech signals. The decisive statistics are the signal power estimate in the information frequency band, the energy distribution of the input signal over the frequency range, and the amount of non-uniformity of the envelope of the input signal filtered in the band-pass filter. To make a final decision on the presence in the communication channel of the AU, secondary processing is used, which is based on the application of a majority rule for a consecutive series of primary decisions.

A disadvantage of the known device is its low efficiency in solving the problem of separation of speech and pauses.

A device for extracting tonal signals in communication channels according to patent RU 2214051, H04B 3/46, H04Q 1/457, H04M 1/50 is known. The invention relates to the field of telecommunications, in particular to automatic means of receiving channel signaling signals in multi-channel communication systems, and can be used to detect acoustic signals in telephone channels.

Known technical solution has not sufficiently high efficiency in solving the problem of separation of speech and pauses.

The closest analogue to the technical nature of the proposed method is the separation of speech and pauses, described in the book "Digital processing of speech signals. //Л.Р. Rabiner, R.V. Best man. Translation from English edited by M.V. Nazarova and Yu.N. Prokhorov. Moscow, Radio and Communication, 1981, pp. 123 - 126, adopted as a prototype.

The prototype method is as follows.

The signal entering the system is sampled during the time interval set for its analysis, and is stored in memory for further processing. The processed signal consists of an interval that contains only interference, the duration of this interval is about 100 ms, and an interval that contains an additive mixture of speech signal and interference (hereinafter - a mixture of signal and interference).

The main parameters are the number of zero crossings over 10 ms and the mean value function calculated using a 10 ms window. From these samples, the average values and variances of the weighted sum of the absolute values of the amplitudes of the samples and the average number of zero crossings (statistical noise characteristics) are calculated.

Taking into account the values of these characteristics and the maximum average value, the thresholds are calculated for the average number of zero-crossing (ESS) and signal energy. The fragment of oscillations is determined, on which the trajectory of the mean value of the signal energy (SES) exceeds the upper threshold. It is assumed that the beginning and end of the word lie outside this fragment.

Then, moving in the opposite direction along the time axis from the moment where the average value of the signal energy exceeded the threshold for the first time, determine the moment at which SZES for the first time turned out to be less than the lower threshold (point N ₁ ). This moment is chosen as the expected start. In the same way is determined and the estimated end of the word (point N ₂ ).

The next step is to move to the left of point N ₁ (to the right of point N ₂ ) and compare the number of zero crossings with the threshold calculated from the initial segment data. If the number of zero crossings exceeds the threshold by 3 or more times, the beginning of the word is transferred to where the zero crossing curve has exceeded the threshold for the first time. Otherwise, the point N ₁ is considered the beginning of the word. A similar process is carried out in relation to the point N ₂ .

The disadvantage of the prototype method is the low accuracy of the solution of the problem and the high probability of an erroneous decision about the appearance of a signal in the presence of interference.

The objective of the proposed method is to improve the accuracy of determining the moment of appearance of the speech signal and the increase in the probability value of the correct decision about the appearance of the speech signal in the presence of acoustic noise.

To solve the problem in the method of separation of speech and pauses, which consists in the fact that the entire analysis interval, consisting of an interval that does not contain a speech signal, and an interval that contains a mixture of speech signal and interference, interference or a mixture of speech signal and interference, which enter the system, sampled and stored in memory for further processing, according to the invention, for a time interval that contains only interference, calculate the value of the correlation function of the interference for a predetermined value difference of arguments (τ), determine the threshold value for the correlation function by multiplying the calculated value of the correlation function by a factor whose value is determined in advance, form the first "sliding window" - the interval of a given duration, shift the "sliding window" by the offset step value, the value of which is determined in advance, calculate the value of the correlation function of the signal for the predetermined value of the difference of the arguments (τ) for this position of the "sliding window", the value of the correlation functions are compared with the calculated threshold value, if the calculated value of the correlation function exceeds the threshold value, then it is considered that there is a speech signal in the “sliding window”, otherwise the described procedure is repeated until a speech signal is detected, while determining the threshold values for the correlation function are calculated and their mean value is used, which is calculated in accordance with the “first come, first gone” principle using previously calculated values For the number of values of the correlation function determined in advance, if a speech signal is decided on, the possible value of its left border is set equal to the value of the left border of the “sliding window”, reduced by a predetermined amount, form the second “sliding window ", The duration of which is determined in advance, establish its initial position, so that the value of its right border coincides with the set value of the left border of the speech si drove the “sliding window” by an offset step value, the value of which is determined according to the specified accuracy of signal positioning, the correlation function of the signal is calculated, this value is compared with the threshold value for the correlation function calculated at the previous step if the correlation function exceeds the threshold value, it is considered that there is a signal in the “sliding window”, and the process is completed, the beginning of the speech signal is considered equal to the position value of the right border “with olzyaschego window ", reduced by a certain predetermined value, otherwise the value is set equal to the beginning of the speech signal value of the left border" sliding window "for which it was decided to presence of the speech signal at the second stage of the procedure determining the appearance of the speech signal.

The proposed method is as follows.

Signals from the output of an electroacoustic device (EIA), passed to the output of a low-frequency filter (LPF), amplified in a low-frequency amplifier (ULF), are sampled using an analog-to-digital converter (ADC) and stored in the memory of the computing device for further processing.

The detection of the speech signal and the determination of the position of its beginning is carried out in several stages.

First stage.

For a time interval that contains only interference, the value of the correlation function of interference is calculated in a known way for the predetermined value of the difference of arguments (τ), (hereinafter referred to as KF) (see, for example, ES Venttsel "Probability Theory", M. 1969, p. 178, 179).

The duration of the interval for which it is considered that it contains only interference is determined at the design stage by experiment or by the method of mathematical modeling.

The threshold value for KF is determined by multiplying the calculated value of KF by a factor, the value of which is determined in advance. The value of the coefficient is determined at the stage of development experimentally or by the method of mathematical modeling, as the value providing the average value of the probability of a correct decision on the presence of a speech signal not less than a given level.

The second stage is the stage of detecting a speech signal and a rough determination of the position of its beginning.

The first “sliding window” is formed - the time interval of a given duration (a certain number of samples), the value of which is chosen based on the condition that the probability value of detecting a speech signal if present (the probability of correct detection) is not lower than the specified level and that the probability value of making a decision about the presence of a speech signal in its absence (the probability of a false alarm) did not exceed the specified value.

The duration of the "sliding window" is determined at the stage of development experimentally or by the method of mathematical modeling. The initial value of the “sliding window” duration in optimization procedures can be chosen, for example, equal to the minimum duration of one phoneme (20 ms).

Shift the "sliding window" on the interval, the value of which is determined in advance (the current position). The initial value of the duration of this interval in the optimization procedures can be chosen, for example, equal to the interval duration, which ensures the specified accuracy of determining the beginning of the speech signal.

Calculate the values of KF signal for the current position of the "sliding window", the values of KF are compared with the calculated threshold value for the correlation function. If the calculated value of KF exceeds the threshold value, then consider that the signal is present.

If it was not decided that the signal is present, then the signal detection procedure described above is repeated. At the same time, when determining the threshold value for the correlation function, the mean value of KF is calculated and used, which is calculated using the calculated values of KF for the previous positions of the “sliding window” and its current value using the principle of “first come, first go” (see for example, Robert Cruise. “Data Structures and Program Designs.” - Binom. Laboratory of Knowledge (2008). That is, from the list of KF values, using which the mean KF value is calculated, delete the first value and add the last calculated value. Then the KF values are renumbered, namely, the value with the second number is assigned the number one, the value with the third number is assigned the number two, etc., the last calculated number is assigned to the last calculated value.

The number of values of KF signal, which is used when calculating the average value of KF signal using the principle of "first come, first left" is determined at the development stage experimentally or by the method of mathematical modeling. The process is repeated until a decision is made about the appearance of a speech signal.

In the case of a decision on the presence of a speech signal, the value of the left border of the signal is set equal to the value of the right boundary of the “sliding window”, reduced by a predetermined amount.

The value of this value is calculated, for example, by the formula

T _with = (T _bc + T _tone ) / 2, (1)

T _bc - the amount by which the “sliding window” (ms) is shifted once;

T _tone - the accuracy of determining the position of the signal (ms).

The value of this quantity can be specified at the development stage experimentally or by the method of mathematical modeling.

The third stage is the stage of precise determination of the position of the beginning of the speech signal.

Form the second "sliding window", the duration of which is set on the basis of the condition that the accuracy of determining the beginning of the signal was not below a predetermined level. The position of this “sliding window” is set so that the value of its right border is equal to the set value of the left border of the speech signal. The “sliding window” is shifted by the size of the offset step, the value of which is determined on the basis of the specified accuracy of determining the position of the signal.

Calculate the value of the correlation function of the signal for a predetermined value of the difference of the arguments (τ). This value is compared with the previously calculated threshold value for the correlation function, which is determined using the last calculated average value of the correlation function of the signal for which the decision about the presence of the speech signal was made. If the KF value exceeds the threshold value, then it is considered that there is a signal in the “sliding window”. Then the process is completed. The time of appearance of the speech signal is considered equal to the value of the position of the right border of the “sliding window” reduced by some predetermined value. The value of this value can be set, for example, equal to the interval offset step, and can be refined at the stage of development experimentally or by the method of mathematical modeling.

If the value of the correlation function does not exceed the threshold value, the value of the beginning of the signal is set equal to the value of the left boundary of the “sliding window”, for which a decision was made about the presence of a speech signal at the second stage of the procedure for determining whether a speech signal appears.

An illustrative explanation of the operation of the method is shown in FIG. one.

Below are the results of the simulation of the decision-making process on the presence of a speech signal depending on the power ratio of interference and signal using the MATLAB system.

Interference in modeling is presented as a set of harmonic oscillations with random amplitudes (U _pi ) and phases (ϕ _pi ), which are distributed according to the normal (amplitude) and uniform (phase) laws (see, for example, the textbook "Fundamentals of the Theory of Radio Engineering Systems ". A textbook. // V. I. Borisov, V. M. Zinchuk, A. E. Limarev, N. P. Mukhin. Under the editorship of V. I. Borisov. Voronezh Research Institute of Communications, 2004., p. 51)

, (2)U =

, (2)

where: ω _pi is the frequency of the i-th interference component;

φ _pi - phase of the i-th component of the interference;

– амплитуда i-ой составляющей помехи;

- amplitude of the i-th component of interference;

N _sp is the number of harmonic components of the interference used to represent it.

The frequencies of the noise components were modeled as random variables whose values are distributed according to a uniform law in the signal band. The duration of the harmonic components of the interference significantly (dozens of times) exceeds the value of the period corresponding to the minimum frequency of the speech signal.

The signal is presented as a set of harmonic oscillations with random amplitudes (U _si ) and phases (ϕ _si ), which are distributed according to the normal (amplitude) and uniform (phase) laws, with the initial phase values for the signal components being equal.

The signal frequency values are set as follows.

The value of the first harmonic is taken as a random variable, the values of which are distributed according to a uniform law in some given frequency band. The “distance” in frequency between the harmonics of the signal is set the same, and this value is taken as a random variable, the value of which is distributed according to a uniform law in some given frequency band.

The following initial data were used in the simulation:

- the number of implementations - 10 ³ ;

- sampling frequency - 64000 samples / s;

- the number of frequency components of the signal - 8;

- the value of the difference of the arguments (τ) - 1;

- the duration of the interval where there is only interference - 50 ms;

- the duration of the interval, where there is only a disturbance for the second stage - 300 ms;

- the duration of the first and second “sliding windows” - 20 ms;

- step size offset "sliding windows":

for the second stage - 10 ms;

for the third stage - 5 ms;

- the number of positions of the "sliding window" used in the third stage - 4;

- the duration of the interval determining the accuracy of determining the position of the signal - 5 ms;

- the value by which the value of the right border of the “sliding window” is reduced, when determining the value of the left border of the signal at the second stage - 10 ms (calculated from f. 1);

- the value by which the value of the right border of the “sliding window” is reduced in the third stage is 5 ms;

- the value by which the last value of the right border of the “sliding window” is reduced (third stage) is 10 ms;

- the range of change of the value of the lower frequency spectrum of the speech signal - 300 ... 1000 Hz;

- the range of variation of the difference between adjacent frequencies of the speech signal spectrum is 100 ... 400 Hz.

The simulation used the optimal average coefficient used in the calculation of the threshold value for CF. Averaging was carried out over the values of power and the number of frequency components of the noise and signal.

Table 1 presents the simulation results of the process of determining the probability of a correct decision on the presence of a speech signal, carried out at the second stage, depending on the number of frequency components of the interference and the power ratio of the interference and signal.

Table 1

N _fp Parameter designation Ratio of interference and signal power 1.0 1.5 2.0 2.5 3 four P _pr one 0.9992 0.997 0.993 0.97 _{Pt lt} 0.015 0.016 0,018 0.02 0.022 15 P _pr one 0.9993 0.993 0.992 0.98 _{Pt lt} 0.043 0.045 0.05 0.06 0.065 thirty P _pr one 0.9998 0.9996 0.995 0.98 _{Pt lt} 0.015 0.03 0.035 0.04 0.05

The table uses the following notation:

N _fp is the number of frequency components of interference;

P _ol - the probability of a correct decision about the presence of the speech signal;

P _lt - the probability of making a decision about the presence of a speech signal in its absence (the probability of a false alarm).

The results of modeling the process of determining the accuracy of the position of the speech signal, depending on the number of frequency components of the interference and the power ratio of the interference and signal carried out in the third stage, are presented in Table 2.

table 2

N _fp Ratio of interference and signal power Sliding window shift step number Accuracy of signal positioning,
ms one 2 3 four four 1.0 1 * 10 ^-3 0.999 0.001 0 ± 2.5 1.5 2 * 10 ^-3 0.998 0,002 0 ± 2.51 2.0 3 * 10 ^-3 0.997 0,003 0 ± 2,515 2.5 4 * 10 ^-3 0.982 0,006 0.001 ± 2.52 3 5 * 10 ^-3 0.99 0,007 0,002 ± 2.54 15 1.0 0.08 0.999 0.001 0 ± 2.9 1.5 0.08 0.993 0,004 0.001 ± 2.91 2.0 0.08 0.967 0.032 0.001 ± 2.98 2.5 0.09 0.92 0.059 0.011 ± 3.13 3 0.09 0.91 0.064 0.016 ± 3.16 thirty 1.0 0.1 0.998 0.001 0.001 ± 3.01 1.5 0.1 0.99 0,008 0,002 ± 3.03 2.0 0.1 0.95 0.038 0.011 ± 3.15 2.5 0.11 0.93 0.055 0.011 ± 3.2 3 0.12 0.85 0.096 0.035 ± 3.46

The table uses the following notation:

N _fp is the number of frequency components of the interference.

From the analysis of the data given in Tables 1, 2, it follows that for values of the power ratio of interference and signal not exceeding 2.5 for any values of the number of frequency components of interference, the probability of correct decision on the presence of a speech signal takes at least 0.992; the probability of a false alarm does not exceed 0.06. In this case, the error in determining the time of appearance of the speech signal does not exceed ± 3.2 ms.

The structural diagram of the device implementing the proposed method is shown in FIG. 2, where indicated:

1 - electroacoustic device (EAU);

2 - low pass filter (LPF);

3 - low frequency amplifier (ULF);

4 - analog-to-digital converter (ADC);

5 - computing device (WU).

The device contains serially connected EAU 1, LPF 2, ULCH 3, ADC 4, WU 5, the output of which is the output of the inventive device, the input of EAA 1 is the device input.

The device works as follows.

Interference or additive mixture of the signal and interference that comes from the output of EAI 1, filter the low-pass filter 2, the band of which is matched to the speech signal band, then the interference or additive signal mix and interference is amplified in the ULF 3 and fed to the input of the ADC 4. Interference or mixture counts the signal and interference generated in the ADC 4, in digital form, is fed to the input of the VU 5.

In computing device 5, the received interference samples or signal mixtures and interference are processed according to the algorithm given above.

The result of processing is the decision in digital form about the presence or absence of a speech signal, for example:

1 - the signal is present;

0 - no signal.

The output of the device also receives the value of the time of appearance of the speech signal, in the case when they decide on the presence of a speech signal. The method for determining the time of appearance of the speech signal is given above.

The results of modeling the process of detecting a speech signal and determining the accuracy of the position of a speech signal depending on the number of frequency components of interference and the power ratio of interference and signal are shown in Tables 1 and 2, respectively.

As the EIA 1, for example, microphones or laryngophone can be used.

ULF 3 can be implemented, for example, on the OP467GS chip of Analog Devices.

The ADC 4 can be implemented, for example, on an ADS8422 chip from Texas Instruments.

Computing device 5 can be made in the form of a programmable logic integrated circuit (FPGA), and implemented, for example, on the XC2V3000-6FG676I chip from Xilinx.

Thus, the inventive method can be implemented by the described device and allows with high efficiency and accuracy to solve the problem of separation of speech and pauses by detecting the appearance of a speech signal by comparing the value of its correlation function with the threshold.

Claims (1)

A method of separating speech and pauses, which consists in the fact that the entire analysis interval consisting of an interval that does not contain a speech signal and an interval that contains a mixture of speech signal and interference, interference or mixture of speech signal and interference that enters the system, sampled and stored in memory for further processing, characterized in that for a time interval that contains only interference, the value of the correlation function of the interference is calculated for a predetermined value of the difference of arguments (τ) the threshold value for the correlation function by multiplying the calculated value of the correlation function by a factor whose value is determined in advance, the first “sliding window” is formed - an interval of a given duration, the “sliding window” is shifted by the offset step value, the value of which is determined in advance, the correlation function value is calculated for a predetermined value of the difference of arguments (τ) for this position of the “sliding window”, the value of the correlation function is compared with the calculated n If the calculated value of the correlation function exceeds the threshold value, then it is considered that there is a speech signal in the “sliding window”, otherwise the described procedure is repeated until the speech signal is detected, while determining the threshold value for the correlation the functions calculate and use its average value, which is calculated in accordance with the principle of “first come, first go” using the previously calculated values of the correlation function and its tech For a predetermined number of values of the correlation function, in the case of a decision about the appearance of a speech signal, the possible value of its left border is set equal to the value of the right border of the “sliding window”, reduced by a predetermined amount, form the second “sliding window”, the duration of which is determined in advance, its initial position is set so that the value of its right boundary coincides with the established value of the left boundary of the speech signal, the “sliding window” is shifted the size of the offset step, the value of which is determined in accordance with a given accuracy of determining the position of the signal, the correlation function of the signal is calculated, this value is compared with the threshold value calculated for the previous step for the correlation function; if the value of the correlation function exceeds the threshold value, then the sliding window ”, the signal is present, and the process is completed, the beginning of the speech signal is considered equal to the position value of the right border of the“ sliding window ”, reduced by predetermined value, otherwise, set the value of the beginning of the speech signal equal to the value of the left border of the “sliding window”, for which a decision was made about the presence of a speech signal at the second stage of the procedure for determining whether a speech signal appears.

Images