RU78470U1 - SYSTEM FOR DETERMINING THE PARAMETERS OF LINEAR SPECTRA OF VOCALIZED SOUNDS - Google Patents

Abstract

The utility model relates to technical solutions used to highlight the spectral characteristics of speech sounds, in particular for detecting speech against the background of stochastic noise, speech recognition, as well as identification of speakers and languages of speech messages. The system for determining the parameters of the linear spectra of voiced sounds contains a digital recorder, a discretization unit, a discrete Fourier transform unit, a Fourier spectrum normalization unit, a resultant convolution matrix generator, an adder, a maximum value recorder, a feature vector generator, a delay line unit, a measure generation unit the weight of the sequence of informative features, the enumeration unit of the sequence of components of informative features and allocation spectra with linearity and smoothness of the dynamics of the fundamental tone frequency, a comparator, a Fourier transform component selector and an informative feature registration unit, as well as a power supply connected to the supply inputs of the component blocks of the system, a memory unit whose input is connected to the output of the discrete Fourier transform unit, and the output is connected to the parametric input of the selector of the Fourier transform components, a basis signal generator, the output of which is connected to the control input of the convolution matrix generator, the formatter of parameters for splitting the frequency range of the fundamental tone of the voice, the output of which is connected to the input of the frequency ranges of the registrar of maximum values, the generator of similarity measures, the output of which is connected to one of the input inputs of the unit for forming measures of weight of a sequence of informative signs, and the threshold level generator, the output of which is connected the input of the comparator.

The utility model allows to increase the accuracy of determining the parameters of the line spectra of voiced sounds. 1 N.p. f-ly., 4 ill.

Description

The utility model relates to technical solutions used to highlight the spectral characteristics of speech sounds, in particular, for detecting speech against the background of stochastic noise, speech recognition, as well as the identification of speakers and languages of speech messages.

A known system for isolating the frequency of the fundamental tone with the conversion of speech oscillations into a pulse sequence containing two selectors, a key, a shaper of the signal of equality of numbers, a key and an analog adder (RU 2007763 C1, 02.15.1994). In this system, it is assumed that each pulse corresponds to a transition through zero of a speech vibration (a unipolar transition is taken), and they also use the property of repetition of interpulse intervals on the period of the fundamental tone of the sequence.

A disadvantage of the known system is the omission of intervals with a fundamental tone when changing one phoneme to another. This is due to the fact that when changing one sounding phoneme to another between two adjacent periods of the fundamental tone, the changes in the pulse intervals become significant.

Also known is a system for isolating the frequency of the fundamental tone using a narrow-band filter (Vocoder telephony. Edited by A.A. Pirogov. M .; Communication, 1974). The specified filter monitor the change in the frequency of the first harmonic of the speech signal. In this case, the bandwidth is iteratively adjusted to the middle frequency

pitch, calculated on the basis of the output function of this filter and transmitted to the filter by the feedback organ. This determines the high quality of the allocation of the frequency of the fundamental tone, provided that the filter adjusts to the speaker for several minutes.

A disadvantage of the known system is unsuitability for detecting the frequency of the fundamental tone in short messages, the duration of which is several seconds.

A known system for highlighting the main tone with a preliminary recording of the speech signal and its subsequent processing, containing three channels for processing the speech signal. In the first channel, amplitude selection is performed according to the Gold scheme, in the second channel, the approximation of the first harmonic of the fundamental tone is used by an exponential function, and in the third channel, the correlation function is calculated according to the Medan scheme. If the difference between the obtained values of the estimates of the frequency of the fundamental tone for different channels does not exceed the specified value, consider that the frequency of the fundamental tone is found (M.E. Hernandez-Diaz Huici and JV Lorenzo Ginori Combined algorithm for pitch detection of speech signals / / Electronics Letters 5th January 1995 Vol. 31, No. l, pp. 15-16).

A disadvantage of the known system is the low accuracy due to the strong dependence on the accuracy of the implementation of each of the algorithms determining the method.

Various systems are also known that make it possible to isolate the speech component of a signal in the presence of a noise component in an acoustic signal (RU 231830 C2, 06.27.2004; RU 296376 C2, 03.27.2007; RU 2271578 C2, 03/10/2006; RU 2263358 C2, 10.27.2005 ; RU 2103753 C1, 01/27/1998; RU 2161826 C2, 01/10/2001, etc.).

A disadvantage of the known technical solutions is that they do not allow speech detection against stationary and

non-stationary noise with a continuous spectrum, without a priori knowledge of the statistical properties of noise.

A known system for highlighting the characteristics of the speech signal MFCC (Mel Frequency Cepstral Coefficients), which implements the splitting of the signal into windows. Using the conversion and calculation blocks, the Fourier transforms are performed on the signal fragments in each window, the Fourier spectrum and the amplitude-frequency characteristics of the set of integrating filters are calculated, the passband of which varies in accordance with the chalk scale known from psychoacoustics. Through integrators, the Fourier spectrum is integrated in accordance with the found parameters of the integrating filters. Additional computing units carry out the logarithm of the integration results and perform cosine transforms on the results of the logarithm (Fang Zheng, Guoliang Zhang and Zhanjiang Song, Comparison of Different Implementations of MFCC, J. Computer Science & Technology, 16 (6): pp582-589, Sept . 2001).

A disadvantage of the known system is the instability to change the statistical properties of noise, against the background of which speech recognition or speaker identification is carried out.

There are also various applications of wavelet analysis in approximating functions, signals and images due to the possibility of efficient compression of signals and their recovery with low loss of information, as well as solving filtering problems of signals (D.S.Potehin, I.E. Tarasov. System Design digital signal processing based on FPGAs, M., 2007, pp85-94). However, their distribution area does not cover the allocation of the fundamental frequency of voiced sounds.

Closest to the claimed system is a system for extracting the frequency of the fundamental tone from a speech signal, presented in RU 2184399 C1, 06/27/2002, which includes series-connected Fourier spectrum calculator, integrator, calculator

values of the determinant of the autocorrelation matrix, the synchronization check unit and the subsequence approximation unit by a quadratic function, as well as the pitch selector according to the Gold scheme and the pitch picker with a bandpass filter, included between the output of the specified calculator and the corresponding inputs of the synchronization check block.

A disadvantage of the known system is the impossibility of detecting speech on the background of stationary and non-stationary noise having a continuous spectrum, without a priori knowledge of the statistical properties of noise, as well as the impossibility of generating a vector of signal attributes for speech recognition systems, speaker identification and language identification, resistant to changing statistical noise properties against the background of which recognition or identification is carried out.

The technical result that is achieved when using the claimed system for determining the parameters of the line spectra of voiced sounds is to increase the accuracy of determining the parameters of the line spectra, which is achieved by:

- the use of dynamic parameters of the spectrum of the information signal and wavelet functions as reference basic signals;

- separation (quantization) of the frequency range of the fundamental tone of the speech signal into non-overlapping areas that do not have multiple frequencies.

The technical result is achieved by the proposed system for determining the parameters of the line spectra of voiced sounds, comprising a digital recorder, a discretization unit, a discrete Fourier transform unit, a Fourier spectrum normalization unit, a resultant convolution matrix generator, an adder, a maximum value recorder, a feature vector generator, a delay line block block

the formation of measures of the weight of a sequence of informative features, a unit for enumerating a sequence of components of informative features and highlighting spectra that have linearity and smoothness in the dynamics of the fundamental tone frequency, a comparator, a selector of Fourier transform components and a recording unit for informative features, as well as a power supply connected to the supply inputs of the system , a memory block whose input is connected to the output of the discrete Fourier transform block, and the output is connected to the parametric input the components of the Fourier transform, the basis signal generator, the output of which is connected to the control input of the convolution matrix generator, the driver of the parameters for splitting the frequency range of the main tone of the voice, the output of which is connected to the input of the frequency ranges of the registrar of maximum values, the generator of similarity measures, the output of which is connected to one of the input inputs of the unit for forming measures of the weight of a sequence of informative features, and the threshold level generator whose output is connected to the task he comparator input.

Figure 1 presents the functional diagram of the proposed system for determining the parameters of the line spectra of voiced sounds.

Figure 2 shows a graph of the wavelet function used in the analysis of the spectrum of sound.

Figure 3 shows an example of the superposition of the wavelet function in the selected sample of the Fourier spectrum.

Figure 4 shows a graph of the formation of measures of the weight of a sequence of informative features.

The system for determining the parameters of the line spectra of voiced sounds (Fig. 1) comprises a digital recording device 1, a sampling unit 2, a discrete Fourier transform unit 3, a Fourier spectrum normalization unit 4,

shaper of the resulting convolution matrix 5, adder 6, registrar 7 of maximum values, shaper 8 of the feature vector, block 9 delay lines, block 10 of forming measures of the weight of a sequence of informative signs, block 11 of enumerating the sequence of components of informative signs and highlighting spectra that have linearity and smoothness of the frequency dynamics the fundamental tone, the comparator 12, the selector 13 of the components of the Fourier transform and the block 14 registration of informative features. In addition, the system includes a power supply unit 15 connected to the supply inputs of the component blocks of the system, a memory unit 16, the input of which is connected to the output of the discrete Fourier transform unit, and the output is connected to the parametric input of the selector 13 of the Fourier transform components, a base signal generator 17, the output of which connected to the control input of the convolution matrix generator 5, the generator 18 of parameters for splitting the frequency range of the fundamental tone of the speech signal, the output of which is connected to the input of the frequency ranges Rathore 7 the maximum values of similarity measures shaper 19 whose output is connected to one of the reference inputs forming unit 10 measures the weight sequence informative signs, and the threshold level generator 20, whose output is connected to the master input of the comparator 12.

The system for determining the parameters of the line spectra of voiced sounds operates as follows.

An acoustic signal is fed to the input of a digital recording device 1, the output of which is formed by a recorded digitized signal. In block 2 of sampling, its window conversion (sampling) is carried out, while non-overlapping window intervals have a duration of at least 0.032 s and follow each other with an offset whose duration does not exceed 0.010 s. To obtain a signal corresponding to each window, in block 3 occurs

computing a discrete Fourier transform while preserving the real and imaginary parts of the transform in the memory block 16. Block 4 determines the Fourier spectrum and carries out its subsequent normalization in accordance with the dependence

Where - components of the normalized Fourier spectrum,

φ _i are the components of the initial Fourier spectrum,

n is the number of components of the Fourier spectrum.

The generator 17 basic signals generates control signals, the structure of which is determined by the parameters of the wavelet function (figure 2), having the form

where τ is the period of the wavelet function,

j is the reference number of the wavelet function,

π = 3.14.

The control basic signals from the output of block 17 go to the input of the convolution matrix generator 5, in which the convolution of the environment of this component with wavelet functions is performed for each i-th component of the obtained normalized Fourier spectrum, namely

where k is the value of a quarter of the wavelet function period, expressed in units of the spectrum,

τ is the period of the wavelet function.

The convolution is carried out with wavelet functions of all possible scales, while the sum of the convolution is calculated separately for each half-wave of the wavelet function. After that, in the shaper 5, the resulting convolution matrix is formed, each gth component of which is determined by the dependence

The frequency range of the fundamental tone of the voice is divided into three non-overlapping ranges, while the upper boundary of the first range is selected from the condition G ₁ = 2G _min , where G _min is the known minimum value of the frequency of the fundamental tone of the male voice, the upper boundary of the second range from the condition G ₂ = 4G _min , and the upper boundary of the third range is taken equal to the known maximum value of the frequency G _{max of the} fundamental tone of the female voice. The boundaries of the ranges are set by the shaper 18 of the parameters for splitting the frequency range of the fundamental tone of the voice.

In adder 6, for the frequency range of the fundamental tone of the voice, the components of the obtained convolution matrix are added, having indices that are multiples of the frequency of the fundamental tone, i.e. determined by the amount of type

ω is the frequency of the fundamental tone, expressed in units of spectral readings.

For each frequency range of the fundamental tone of the voice, the boundaries of which are set by the shaper 18, in block 7, the value of the matrix maximum S _τω

and the argument corresponding to this maximum

9 after which block 8 forms sets of pairs of these values, i.e. vector of informative signs B = .

Using vectors of informative features and measures of similarity of pairs (I _i , g _i ,)

where σ is the dispersion of the frequency of the fundamental tone,

t is the discrete time

g _i ^(t) g _j ^(t) - components of the vector of informative features taken in

discrete time instants t; t = 1,2, ..., and i, j = 1, ..., 3.

that are created at the output of the shaper 19 of the measures of similarity of the pairs, in the block 10 of forming measures of the weight of the sequence of informative signs based on the subsequence Q of size M, at least four in the sequence of sets B, which is implemented by M-1 by the successive delay lines of the delay unit 9, are formed

measures of weight of a sequence of informative features

characterizing the presence of a line spectrum in the signal. With the help of these measures, the weight of a sequence of informative features distinguishes all spectra that have linearity and smoothness of the dynamics of the frequency of the fundamental tone.

From the output of block 10 of forming measures of the weight of a sequence of informative signs, the signals are sent to block 11 of enumerating a sequence of components of informative signs and selecting spectra that have linearity and smoothness of the dynamics of the frequency of the fundamental tone, where the trajectory of the sequence of pairs (I _i , g _i ) ^{(t) is} determined, which has maximum weight D on subsequences Q.

In the graph performed by block 11 of enumerating the sequence of components of informative features and highlighting spectra that have linearity and smoothness of the dynamics of the frequency of the fundamental tone, into a sequence of sets of maximum pairs and their arguments (Fig. 4), each edge corresponds to a calculation of the measure of similarity of these pairs. The corresponding information signal about the selected path is fed to the input of the comparator 12 for comparison with threshold signals received from the threshold level generator 20. In the comparator 12, the threshold value and the maximum weight value D are compared. If the maximum path weight D exceeds the threshold value, then a decision is made that this sequence of pairs (I _i , g _i ) ^{(t) of the} subsequence Q is generated by speech a signal with a linear spectrum and smoothness of the dynamics of the frequency of the fundamental tone. Based on this, selector 13 selects only those components from the sequence of real and imaginary parts of the Fourier transform stored in block 16 that correspond to areas with a linear spectrum and smoothness of the fundamental frequency dynamics and register them with the informative feature registration unit 14.

The implementation of the utility model allows for high accuracy in determining the parameters of the line spectra of voiced sounds. So, for example, comparing the results obtained using the MFCC technical solution, which allows the extraction of features of the speech signal, with the results provided by the proposed system, shows that in the voiced speech section, when the signal-to-noise ratio decreases from 20 to 15 dB, the relative dispersion of the components MFCC vector is 42.1%, and the relative increase in the dispersion of the components when implementing the proposed system is 7.4%. This allows us to talk about the achieved stability of the definition.

parameters of the line spectrum of voiced sounds to noise of speech.

Claims (1)

A system for determining the parameters of the linear spectra of voiced sounds, comprising a digital recorder, a discretization unit, a discrete Fourier transform unit, a Fourier spectrum normalization unit, a resultant convolution matrix generator, an adder, a maximum value recorder, a feature vector generator, a delay line unit, a measure generation unit the weight of the sequence of informative features, the enumeration unit of the sequence of components of informative features and highlighted spectra with linearity and smoothness of the dynamics of the fundamental tone frequency, a comparator, a Fourier transform component selector and an informative feature registration unit, as well as a power supply connected to the power inputs of the system’s composite units, a memory unit, the input of which is connected to the output of the discrete Fourier transform unit, and the output is connected to the parametric input of the selector of the components of the Fourier transform, the basis signal generator, the output of which is connected to the control input of the matrix former is minimized , a shaper of parameters for splitting the frequency range of the fundamental tone of the voice, the output of which is connected to the input of the frequency ranges of the registrar of maximum values, a shaper of similarity measures, the output of which is connected to one of the input inputs of the unit for forming measures of the weight of a sequence of informative signs, and a threshold level shaper whose output is connected to the input of the comparator.