RU2485607C2 - Apparatus and method for computing filter coefficients for echo suppression - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: realisation of an apparatus (100) for computing filter coefficients for an adaptive filter (210) for processing a microphone signal so as to suppress an echo due to a loudspeaker signal, which includes an extractor (250) for extracting a stationary component signal or a non-stationary component signal from the loudspeaker signal or from a signal derived from the loudspeaker signal, and a computer (270) for computing the filter coefficients for the adaptive filter (210), which uses the extracted stationary component signal or the extracted non-stationary component signal.

EFFECT: improved sound quality in systems for suppressing or compensating for echo.

24 cl, 20 dwg

Description

The constructive solutions of the present invention relate to a device and methods for calculating the transmittance of an adaptive filter designed to suppress the echo of a microphone signal excited by a loudspeaker signal, and can be applied, for example, in conference communication systems. Conferencing systems here include telephone, video and other types of interactive duplex networks.

An acoustic echo occurs when tones, sounds and noise from a speaker are picked up by a microphone installed in the same room or in the same acoustic environment. In telecommunication networks, such an acoustic signal is returned to the subscriber at the far end of the line in the form of his own speech, sounding belatedly. In such situations, the reflected signals are distracting and annoying and can cause a violation of the full interactive duplex communication. In addition, an acoustic echo can generate a parasitic hum and other manifestations of unstable states of the acoustic feedback loop.

In WO 2006/111370 A1, the authors describe a device and method for eliminating the echo of a multi-channel audio signal. Control of the acoustic echo signal and suppression of interference is an essential part of any “hands-free” long-distance communications network, for example, telephone or audio and video conferencing systems. The method described in the publication for processing multichannel audio signals from a speaker and at least one microphone signal includes the operations of converting the microphone input signal into its short-term spectra, calculating, based on the loudspeaker signals, the short-term spectrum of the combined loudspeaker signal by calculating, based on the microphone input signal, the short-term spectrum of the combined microphone signal, estimating the amplitude spectrum or the power spectrum of the echo signal in the combined short-term spectrum of the microphone signal, calculating the gain filter to correct the amplitude of the short-term spectrum of the microphone input signal, applying at least to one spectrum of the microphone input signal and converting the filtered microphone input spectrum to the time domain.

The echo cancellation and echo cancellation systems used today, also called echo cancellation systems, are often ineffective for many types of

sounds, tones and noise, despite the adaptive filters used in them. For example, in a communication system, the predominance of one element over another may lead to non-optimal compensation of the loudspeaker echo in the microphone signal. On the other hand, in the case of distortions when combining components from different sources due to the use of means of suppressing or compensating for the echo, tonal artifacts can occur, which are also perceived as an extremely annoying factor.

Thus, based on the prior art, the present invention seeks to improve the sound quality in echo cancellation or compensation systems.

This problem is solved using the device according to paragraph 1, the method according to p. 23, 25 or the program according to p. 26 of the patent formula.

In a technical embodiment, the transmitter of the transmittance of an adaptive filter of a microphone signal designed to suppress the echo excited by a speaker signal includes an extractor of the stationary or non-stationary component of the speaker signal or derivative of the speaker signal. In addition, the transmitter includes an adaptive filter transmittance calculator based on extracted stationary or non-stationary signal components.

With this configuration, the implementation of the method for calculating the transmittance of the adaptive filter of the microphone signal to suppress the echo excited by the speaker signal consists in selecting the stationary or non-stationary component of the speaker signal or derivative of the speaker signal and calculating the transmittance of the adaptive filter based on the selected stationary or non-stationary signal component.

The implementation of the invention is based on the conclusion that an improvement in sound quality can be achieved by improving the statistical properties of the speaker signal or a derivative of the speaker signal, the analysis of which is performed when calculating the adaptive filtering coefficients to suppress the echo. To this end, the loudspeaker signal or derivative of the loudspeaker signal is analyzed to extract one or more corresponding stationary and / or non-stationary signal components. Then, based on the identified stationary or non-stationary signal components, the transmission coefficients of the adaptive filter are calculated.

The stationary component of a signal or derivative of a signal, such as a loudspeaker, can display, say, in the frequency domain, an energy value that only changes slightly over time, or can form a corresponding stationary component. Thus, the stationary component of such a signal can be determined, for example, in the frequency domain while simultaneously determining the energy index for the corresponding band signal and averaging over time. Averaging can be floating and performed with different design characteristics. Such a calculation can be performed recursively using a filter type construction with IIR (IIR = infinite impulse response). Similarly, appropriate averaging can be performed using a FIR type filter design (FIR = finite impulse response).

Therefore, the non-stationary component of the signal or derivative of the speaker signal can be determined based on the corresponding band signal. In implementations of the present invention, the associated non-stationary component can be determined using the stationary signal component and a gain filter. The gain filter may depend on at least one control parameter, which in the implementation of the present invention is determined, for example, on the basis of the coherence function, which takes into account the loudspeaker signal and the microphone signal or signals derived from them.

According to this invention, the transmittance of the first filter is calculated based on the stationary component of the signal, the transmittance of the second filter, based on which the transmittance for the adaptive filter is determined, is calculated based on the non-stationary component. In this case, the transmittance of the adaptive filter may correspond to the serial connection of the first filter, which is based on the transmittance of the first filter, with the second filter, which is based on the transmittance of the second filter. The implementation of this invention also allows you to determine the filter coefficients both on the basis of the transmittance of the first filter, and on the basis of the transmittance of the second filter.

Next will be described options for structural solutions and functionality of the invention. Depending on the version of execution, both stationary and non-stationary components can be estimated based on the corresponding signals. Further, the structure implemented in accordance with the present invention may include a microphone noise filtering device based on filter coefficients.

Embodiments of the invention will be discussed in more detail and graphically illustrated below. Thus, a more substantive discussion of the present invention will be accompanied by the following figures:

figure 1 shows a typical situation in which the elimination of acoustic echo is required;

figure 2 is a schematic block diagram of a calculator of filter coefficients according to the approach of the present invention;

in FIG. 3A to 3C are schematic block diagrams of extractor options according to the approach of the present invention;

on figa and 4B are schematic block diagrams of an additional filter preliminary analysis of the echo signal according to the approach of the present invention;

in FIG. 5A to 5E are schematic block diagrams of embodiments of a calculator in implementations of the present invention;

Fig.6 is a schematic block diagram of a modification of the invention;

Fig.7 is a schematic block diagram of an embodiment of the invention in a configuration of a filter coefficient calculator;

FIG. 8a to 8c graphically show the separation of the stationary and non-stationary components of the speaker signal;

on figa graphically displays the function of the echo cancellation filter for non-stationary components of the signal at a frequency of 1 kHz;

Fig. 9b graphically displays the function of a correlation echo cancellation filter for the non-stationary component of this signal;

in FIG. 10a to 10c show the ratio of the levels of the stationary and non-stationary components, amplification of the prediction and speech activity of the loudspeaker channel;

figure 11 is a schematic block diagram of an embodiment of the invention;

Fig.12 is a schematic block diagram of an embodiment of the invention;

on Fig presents a schematic block diagram of a multi-channel solution of the present invention;

on Fig shows an example of grouping the spectrum of a uniform short-term Fourier transform for modeling uneven frequency resolution of human hearing;

on figa shows a graph of the use of Hannah interpolating filters for frequency smoothing of the gain filter;

15b shows an interpolation curve of the gain filter coefficients.

At the beginning of the detailed description based on figures 2 through 15 of the embodiments of the present invention that solve the problem of suppressing acoustic echo by separating the stationary and non-stationary components of the signal, figure 1 illustrates a typical situation in which the elimination of acoustic echo is required.

In the drawings, the following notation:

100 loudspeaker;

110 microphone;

120 acoustic environment;

130 speaker signal;

140 microphone signal;

150 block elimination of the echo;

160 signal with blocked echo;

170 direct way;

180 indirect way;

200 device (calculation of filtration coefficients);

210 adaptive filter;

220 entrance;

230 time-frequency converter (VChP);

240 filter preliminary analysis of the echo signal;

250 extractor;

260 filter preliminary analysis of the echo signal;

270 calculator;

280 entrance;

290 time-frequency converter (VCHP);

300 time-frequency converter (CVP);

310 exit;

320 averager;

330 gain filter;

340 parameter calculator;

350 dispenser;

360 filter link;

370 filter parameters calculator;

380 combinator;

390 selector;

400 parameter identifier;

410 dispenser;

420 curve graphics;

430 curve graphics;

440 brace;

450 braces;

460 arrows;

470 filter preliminary analysis of the echo signal;

480 delay device;

490 energy calculator;

500 energy calculator;

510 energy magnitude calculator;

520 additional computer;

530 grouping machine;

540 additional grouping device.

An acoustic echo occurs when a microphone picks up tones, sounds, or noise coming from a speaker located in the same room or in the same acoustic environment. In telecommunication systems, acoustic feedback signals are relayed to the interlocutor at the far end of the line, who perceives them as an echo of their own speech. In such a situation, echoes can be a strong distraction and even disrupt interactive full-duplex communication. In addition, the acoustic echo can generate spurious whistles and other unstable states of the acoustic feedback loop. Naturally, hands-free remote control systems that provide full two-way communication require echo control to eliminate the interaction between the speaker and the microphone. Figure 1 illustrates the situation of the occurrence of acoustic echo.

1 shows a loudspeaker 100 and a microphone 110 located in the same acoustic environment 120, which may, for example, be formed in a room. Similarly, the acoustic environment 120 may be formed by the interior volume of the vehicle interior.

In figure 1, the signal of the speaker 130, or x [n], where the time indicator n is an integer, is fed to the speaker 100. The microphone 110 picks up noises, sounds and tones of the general sound environment 120 and generates a microphone signal 140 or y [n] . According to figure 1, the loudspeaker signal 130 and the microphone signal 140 in the form of input signals are supplied to the echo canceller 150, which at the output generates an echo-cleared signal 160, or e [n] from the microphone signal 140.

Thus, FIG. 1 illustrates the problem of the occurrence and control of acoustic echo in two-way communication systems. The signal from the far end of the telecommunication line, converted by the loudspeaker into sound, enters the microphone directly through 170 and along the reflected paths 180-1, 180-2, which are also called indirect channels. As a result, the microphone 110 not only perceives a voice that sounds locally at the transmitting end of the line, but also picks up an echo, which is immediately relayed back to the receiving end of the line.

In other words, the speaker signal x [n] is again mixed with the microphone signal y [n]. Ideally, using the echo cancellation unit 150, such an echo should be completely eliminated, while the outgoing voice signal at the near end of the telecommunication system should be skipped.

The standard way to combat the reflected signal is the parallel inclusion of an acoustic echo canceller (AED) into the echo propagation channel, as described in [1]. Such an acoustic echo canceller analyzes a digital replica of the echo signal, subtracting it subsequently from the measured or actual microphone signal. Standard approaches to the problem of acoustic echo removal are based on the conclusion that the passage of the echo signal can be modeled using a FIR filter (with a finite impulse response), after which the corresponding acoustic echo cancellers are used, which is also described in [C. Breining, P. Dreiseitel, E. Hänsler, A. Mader, B. Nitsch, H. Puder, T. Schertler, G. Schmidt, and J. Tilp. Acoustic echo control. IEEE Signal Processing Magazine, 16 (4): 42-69, July 1999]. Due to the fact that the path of echo formation is usually unknown and, moreover, may change during the working process, the linear filter of such an acoustic echo suppressor is usually implemented as adaptive. FIR filters with a duration not exceeding several hundred milliseconds, which corresponds to a sampling frequency, which, in turn, assumes a high level of computational complexity, are used to simulate typical echo formation paths.

The echo attenuation levels achieved with the practical use of traditional approaches are often insufficient for various reasons. Such reasons may be, in particular, too long reverberation (echo tail effect), which leads to errors in modeling the paths of the echo signal, nonlinear components of the echo signal, caused, for example, by vibrations or nonlinear deviations in inexpensive audio equipment, and convergence in the case of the high instability of the echo paths, as discussed in [A.N. Birkett and RAGoubran. Limitations of handsfree acoustic echo cancellers due to nonlinear loudspeaker distortion and enclosure vibration effects. In Proc. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, p.13 - 16, New Paltz, Oct. 1995]. Therefore, to eliminate the residual echo and its components, which were not eliminated using the acoustic echo canceller, echo cancellers are combined with non-linear post-processors, as described in [G. Schmidt and E. Hänsler. Acoustic echo and noise control: a practical approach. Hoboken: Wiley, 2004]. In most cases, the residual echo is quenched by a frequency-selective method, as described in [W.L. B. Jeannes, P. Scalart, G. Faucon, and C. Beaugeant. Combined noise and echo reduction in hands-free systems: a survey. IEEE Transactions on Speech and Audio Processing, 9 (8): 808-820, Nov. 2001]. In fact, all acoustic echo cancellers are equipped with postprocessors, because too often they do not completely suppress the echo so that it cannot be heard.

Recently in [C. Faller and J. Chen. Suppressing acoustic echo in a sampled auditory envelope space. IEEE Trans. on Speech and Audio Proc., 13 (5): 1.048-1.062, Sept. 20055, and C. Faller and C. Toumery. Estimating the delay and coloration effect of the acoustic echo path for low complexity echo suppression. In Proc. Intl. Works, on Acoust. Echo and Noise Control (IWAENC), Sept. 2005], several sub-band acoustic echo cancellation devices have been proposed that are similar to the aforementioned non-linear post-processors, but do not need an acoustic echo canceller and an assessment of the impulse response of the echo path. It is reported in publications that these systems are characterized by low computational complexity, reliability, and high duplex characteristics.

In development [C. Faller and C. Toumery. Estimating the delay and coloration effect of the acoustic echo path for low complexity echo suppression. In Proc. Intl. Works, on Acoust. Echo and Noise Control (IWAENC), Sept. 2005] acoustic echo cancellation devices for calculating the spectra of the loudspeaker and microphone signals, an algorithm using the window (short-term) Fourier transform (OPF) is proposed. To the corresponding loudspeaker signal, the delay value d is applied between two signals converted using an OPF, which is selected taking into account the greatest effect of the impulse response of the echo channel.

Then evaluate the filter analysis of the actual echo signal, which simulates the initial path of the echo. To obtain the amplitude spectral characteristic of the echo signal, a calculated delay value and a filter of preliminary analysis of the echo signal are applied to the spectrum of the loudspeaker signal. Using the obtained amplitude spectral characteristic of the echo signal, the actual echo cancellation filter is calculated and applied to the spectrum of the microphone signal for echo cancellation.

A disadvantage of the aforementioned acoustic echo cancellation systems is that they do not fully perform their functions with respect to mixed speaker signals containing both stationary and non-stationary components. This may be the case, for example, when speech at the far end of the line sounds in a high noise environment. In such a situation, the loudspeaker signal and the echo signal comprise a non-stationary component in the form of speech at the far end and a stationary component in the form of a noise background at the same far end of the communication line.

Existing acoustic echo cancellation systems define only one echo cancellation filter for the loudspeaker signal. It follows that these approaches do not take into account that echo components with different characteristics excite various types of distortions in the signals of the transmitting end, which subsequently require different types of processing.

Figure 2 shows the first embodiment of a device 200 for calculating the transmittance of an adaptive filter 210, in which, as in the accompanying method, the separation of stationary and non-stationary signals is used to improve echo cancellation and, therefore, to improve the quality of perceived sound. Thus, the technical solutions of the present invention involve the use of various methods of signal compensation depending on their statistical properties and features, which provides more efficient echo cancellation, less prone to artifacts.

A detailed description of the embodiments of the present invention shown in FIGS. 3-5 is preceded by a consideration of the block diagram of the device 200. Here it is necessary to pay attention to the fact that the illustrations and descriptions of the circuit diagrams of the devices proposed in the invention are simultaneously block diagrams of the corresponding methods. In other words, the basic block diagrams of the device structures under consideration correspond to the block diagrams of the algorithms of the proposed method, where the sequence of operations performed by the equipment components is displayed.

It should be borne in mind that in the framework of this description, for devices, blocks and circuits, identical or similar in their functions, the same or similar reference numbers are used. At the same time, devices, blocks and circuits denoted by matching or similar reference numbers contain identical or similar structural and functional characteristics. In other words, in the proposed description, the same reference numbers are used to refer to devices, blocks and circuits having the same or similar functions, purpose or design. This allows you to succinctly present the materials of the present invention, using the description of one technical solution to explain another implementation option, unless clearly defined otherwise.

In addition, when describing the invention, the numbers of generalizing references are used to repeatedly denote devices, blocks and circuits in the same figure. In particular, for two indirect paths 180-1, 180-2 in FIG. 1, different reference numbers are defined, but if indirect paths are referred to as such or if their general distinguishing features are considered, only generalizing reference number 180 is applied. This also contributes to brevity presentation and better understanding of the proposed description. The device 200 of FIG. 2 has an input terminal 220 for inputting a speaker signal not shown in FIG. 2. Through the input terminal 220, the signal is input to the time-frequency converter 230, shown in FIG. 2 by a dashed line as an option of the device 200. From the time-frequency converter 230, the signal can be supplied to the first filter of the analysis of the echo signal 240, which is also optional in the design of the device 200. The output of the preliminary analysis filter of the echo signal 240 is connected to the input 250a of the extractor 250, which, in turn, through the first output 250c and the second output 250d can be connected to an optionally installed second preliminary filter a an echo signal 260 through its first input 260a and a second input 260b. The presence of this pre-analysis filter of the echo signal is also optional and depends on the specific implementation. Say, the device 200 proposed in the present invention can be structurally solved with the introduction of the first 240 and second 260 filter preliminary analysis of the echo signal together or separately, or without any of them or both. Of course, a variant is possible where only one of the two filters for preliminary analysis of the echo signal 240, 260 is used. Technical solutions are also feasible involving other elements of the circuit.

With a second echo pre-analysis filter 260, its first output 260c and second output 260d are connected to the first input 270a and second input 270b of the transmittance calculator 270 of the adaptive filter 210. The calculator 270 is connected via an output 270d to the input of the adaptive filter 210.

A microphone signal may be provided to the other input of the adaptive filter 210 through an additional time-frequency converter 290 from the input terminal 280. The output of the adaptive filter 210 can be connected via an additional time-frequency converter 300 to the microphone signal output terminal 310. At the same time, the input terminal 280 is arbitrarily connected via the time-frequency converter 290 to the second input 250b of the extractor 250 and to the third input 270 from the calculator 270. Both inputs - 250b of the extractor 250 and 270c of the calculator 270 - play the role of auxiliary and can be entered independently from each other in various hardware versions of the present invention.

For example, the device 200 may be included in the echo cancellation unit 150 shown in FIG.

Before a more detailed consideration of the functions of the device 200 in the embodiment of FIG. 2, it should be noted that, in general, the design of the invention can be solved on the basis of both discrete and integrated, or other, more complex, circuits. In particular, the present invention can be integrated into data processing means such as processors, integrated systems (SOC = systems on a chip), application integrated circuits (ASICs) or other integrated circuits and specialized processors. In such configurations, identical elements of the data processing loops can be activated alternately for different devices. For example, the same logic gate of an arithmetic logic device (ALU) of a microprocessor can be used, firstly, to control the extractor 250, and secondly, the calculator 270. However, the devices can differ significantly from each other, as for example, in the aforementioned case, they require different control commands that together define each of the devices. Because of this, partial or complete overlap of circuitry elements implemented in various versions of devices is permissible.

Largely for this reason, here in the description, paired devices, blocks and circuits are understood as directly or indirectly interconnected. For example, if the implementation is based on data processing tools, interaction can be carried out through a memory cell containing an intermediate result in the form of a signal latched in it.

Moreover, however, the constructive solutions of the present invention are not limited to digital devices, although mainly digital means will be described hereinafter. The invention basically provides for the possibility of its analog and mixed analog-to-digital performance. In such configurations, an ADC or DAC (analog-to-digital or digital-to-analog converters) is additionally introduced to transform one type of signal into another.

Depending on the purpose of the device 200 shown in FIG. 2, the speaker signal supplied to input 220 can be converted into the frequency domain using a time-frequency converter 230, which is shown as an option. The time-frequency converter 230 provides an adequate spectral representation of data blocks (frames) from the time domain at the output. Depending on the particular implementation of the invention, a Fourier converter, a subband converter, or a KZF converter (based on the KZF = quadrature-mirror filter) may be introduced into the configuration of the time-frequency converter 230. Regardless of the particular application, the time-frequency converter 230 transforms the signal it receives (from the time domain) into a plurality of band signals. Each band signal has a characteristic frequency, which may be an average frequency, a lower cutoff frequency, or an upper cutoff frequency of a corresponding band. Depending on the features of the technical solution, different band signals can have more than one characteristic frequency or be characterized by other parameters.

The first pre-analysis filter of the echo signal 240 enables the modulation of the acoustic medium 120 (in FIG. 1), which would provide a signal at its output containing the calculated amplitude spectral characteristic corresponding to the signal that will be generated by superposing the speaker signal on the microphone signal. However, as already explained above, the first pre-filter analysis of the echo signal 240 is auxiliary and not required for installation.

Further, the loudspeaker signal or a signal derived from it, obtained by converting and filtering with the corresponding additional tools 230 and 240, is supplied to the first input of the extractor 250. The extractor 250 selects the stationary and non-stationary components from the speaker signal or from its derivative. In particular, the stationary component can be calculated by averaging the input signal, which is described below.

Depending on the technical solution chosen, the signal may be a calculated signal deviating from the “real” stationary component. Accordingly, the non-stationary component can be determined from the stationary component of the signal, for example, by using a gain filter, which is not shown in FIG.

In other implementations, extractor 250 may use a different stability assessment method.

The non-stationary component can also be determined, for example, by comparing the temporal change in the input signal. In addition, if embedding a voice codec in the configuration or in the medium of the device 200, one can resort to a prediction method using an extractor 250 to extract at least one of the two mentioned signals. A similar prediction method can be represented, for example, by an LPC codec error signal (LPC = linear prediction coding).

The extractor 250 has two outputs, indicated above, to which various signals can be supplied depending on the requirements. In particular, as a rule, at least a stationary or non-stationary component is fed to the first output of the extractor 250. The second of the two components or a signal containing information about the output signal at the first output is fed to the second output. This may be, presumably, the parameters for further processing by the calculator 270 of the corresponding signal, or it may be a simple control signal indicating which of the two components is transmitted.

According to the functions performed, an arbitrary second echo pre-analysis filter 260 typically corresponds to a first echo pre-analysis filter 240. Typically, a second echo pre-analysis filter 260, if it is designed to perform such an echo estimate, can evaluate the speaker at the input terminal 220, as a result, to obtain a signal that would correspond to the signal perceived by the microphone in the absence of other sources of noise. The first echo pre-analysis filter 240, like the second echo pre-analysis filter 260, may optionally include a delay device that takes into account the delay of the loudspeaker echo picked up by the microphone. In other words, filters 240, 260 can also be used to delay the signal from a loudspeaker, or a derivative of it, both with the help of an additionally mounted delay device, as well as through internal circuitry. In most cases, separation of the functions of echo estimation, on the one hand, and delay, on the other, is also possible if, for example, the first pre-analysis filter of the echo signal 240 is used only to delay the corresponding signal, and the second filter of the preliminary analysis of the echo signal 260 - for real echo analysis.

Then the signals from the second filter of the preliminary analysis of the echo 260 are fed to the calculator 270, which, in turn, calculates or determines the transmittance of the adaptive filter 210, based on the extracted stationary or non-stationary components. Depending on the particular application, the transmitter 270 may also access the microphone signal received at the input terminal 280, or the microphone signals converted to the frequency domain. This signal is also available for extractor 250, as described in more detail below.

Further, the adaptive filter 210, receiving filter coefficients from the calculator 270, corrects the spectrum of the microphone signal with the formation of an at least partially echo-compensated version of this signal, which is sent for further processing. Depending on the specifics of the technical design, the microphone signal that has undergone echo cancellation or the spectrum modification described above can be converted to the time domain using the time-frequency converter 300 or output directly to the output terminal 310. However, the need for inverse conversion to the time domain using the converter 300 may not be, if, for example, the microphone signal is encoded in the frequency or related region.

Here, before proceeding to a detailed discussion, accompanied by FIGS. 3A-5E, of the circuit elements of the device 200 of FIG. 2, it should be noted that the conversion of the speaker signal or its derivative can for the most part be performed in the frequency domain, and processing subject, respectively, to a single associated band signal, a plurality of band signals, a plurality of band signals or all band signals.

It should also be noted that individual devices and filters can work, for example, using energy indicators, which depends on the individual design solution. The energy indicator is the result of raising the actual value of the base to a power with an even indicator or the result of raising a module (absolute value) to a power with any indicator. For example, short-term spectra that are processed using separate filters or separate devices can use energy indicators, in particular, energy values formed as squares of modules of the corresponding spectral coefficients. Similarly, modular spectra, that is, the absolute values of the corresponding spectral coefficients, can be used with exponent 1. Formulating differently, quantities proportional to SzS ^m , where m is a positive, say, natural number, can be used as energy indicators, starting with any value of z that is a real or complex quantity. With z being the actual value, values proportional to z ^2m can be additionally used as energy indicators.

FIG. 3A is a schematic diagram of an extractor 250, implemented in accordance with the present invention as a possible component of the device 200. Extractor 250 has only one first input 250a, arbitrarily connected to the output of the first pre-analysis filter of the echo signal 240, shown in FIG. 2 as option. The extractor 250 in FIG. 3A does not have a second input (input 250b in FIG. 2).

An averager 320 is connected to the first input 250a of the extractor 250 to determine the average value of the signal at the input 250a. The term “signal” here means not only signals in the time domain (temporary signals), but also signals in the frequency domain, where the corresponding signals are the spectral representation of signals in the time domain. Similarly, signals can include and transmit information obtained from the above signals, such as the magnitude of the amplitude in the frequency domain (energy spectrum), energy values (squares of the amplitude), spectra and other derived values and indicators.

Inside the circuit of the extractor 250 in FIG. 3A, the signal received from the input 250a to the averager 320 is output from it in the form of a stationary signal component through the first output 250c. As shown in FIG. 2, a first output 250c is coupled to an arbitrary second echo pre-analysis filter 260 and / or to a calculator 270.

Within the extractor circuit 250, the signal received at the first input 250a then goes along with the stationary component of the signal from the averager 320 to the gain filter 330, which generates a non-stationary component of the signal and feeds it to the second output 250d. The gain filter 330 determines the non-stationary component of the signal based on the loudspeaker signal received or received from the first input 250a and based on the stationary component of the signal. In more detail, the functions of the averager 320 and the gain filter 330 will be discussed later in the context of the description of FIG. 2.

FIG. 3B shows a possible modification of the extractor 250 as part of the device 200. The extractor 250 in FIG. 3B differs from the extractor in FIG. 3A by the presence of a parameter calculator 340, the input of which is also connected to the first input 250a. From the output of the parameter calculator 340, the control parameters generated by it are output to the gain filter 330 for calculating the non-stationary signal component. Features of the operation are discussed below.

The optional extractor 250 shown in FIG. 3B has a second input 250b, already shown in FIG. 2, which may not be directly connected, on the one hand, to an additional input of the parameter calculator 340 and, on the other hand, to the input terminal 280 for a microphone signal, which is also shown in FIG. 2. In this case, an indirect connection can be established through a time-frequency converter 290. The operation features of a parameter calculator 340 are also discussed below.

On figs shows a possible option of embedding the extractor 250 in the configuration of the device 200 in figure 2. The layout of the extractor 250 in FIG. 3C is based on the construction of FIG. 3B with reference to the parameter calculator 340 as an option with the corresponding connections. Unlike the extractor 250 in FIG. 3B, the extractor 250 in FIG. 3C contains a distributor 350 having two inputs, one of which is connected to the output of the averager, and the second to the output of the gain filter 330. The distributor 350 receives a stationary component from the averager 320 and a non-stationary component from the gain filter 330.

Two outputs of the distributor 350 are connected to the first output 250c and to the second output 250d of the extractor 250. The distributor 350 determines which of the two components it receives will be transmitted through the output 250 s for further processing. Depending on the choice of one of the two components, the distributor 350 generates and outputs to the second output 250d of the extractor 250 a control signal, which may contain, for example, information about which of the two signal components arrived at the first output 250c, or the parameters necessary for further processing Components. Next, we will consider the parameters that may be contained in the output control signal.

Depending on the individual design solution, the distributor 350 may transmit the component signal to the first output 250c, which implies a higher volume level, higher energy level or a larger amount of energy compared to the other component signal. If necessary, different components can be output for different band signals.

Therefore, the extractor 250 in FIG. 3C differs from the versions shown in FIGS. 3A and 3B mainly in that it generates only one of two component signals at the first output 250c. As already explained in connection with FIG. 2, the extractor 250 in FIG. 3C generates at the output only a control signal that contains information about the component signal at the first output 250c.

On figa shows a first embodiment of a second filter preliminary analysis of the echo signal 260, presented in figure 2 also as an option. An arbitrarily set second pre-filter of the preliminary analysis of the echo signal 260 includes two filter links 360-1, 360-2, the input of each of which is separately connected to the inputs 260a, 260b, respectively. Both links of the filter 360-1, 360-2 are output separately to two outputs 260c, 260d, respectively, of the second filter of the preliminary analysis of the echo signal 260.

A second echo pre-analysis filter 260, shown as an option in FIG. 4A, can be used in conjunction with an extractor 250 shown in FIGS. 3A and 3B. More specifically, in FIG. 4A, a second echo pre-analysis filter 260 enables simultaneous processing of the stationary component of the signal using subfilter 360-1 and the non-stationary component of the signal using subfilter 360-2. The design of the two filter links 360-1, 360-2 may be the same or different depending on the identity or difference of the echo analysis filters used in modulating the acoustic environment 120 (in FIG. 1) for stationary and non-stationary signal components. Both links of the filter 360-1, 360-2 can be implemented on the same element base to perform latching or buffering of a single signal.

In the discussion of FIG. 2, it was already mentioned that there is a dual use of echo analysis filters 240, 260, similarly filter links 360 can be used, for example, only to provide delay. Naturally, the implementation of the second filter of the preliminary analysis of the echo signal 260 provides for the involvement of filtering links other than those described previously. For example, in the link of the filter 360 provides an additional input signal control action on the filtering process,

FIG. 4B illustrates an embodiment of the second filter for echo analysis 260, which differs from the version in FIG. 4A in that it implements only one filter link 360 mounted between the first input 260a and the first output 260c. In the technical embodiment shown in FIG. 4, the signal received at the second input 260b is supplied to the second output 260d.

Thus, the second pre-filter analysis of the echo signal 260 shown in fig.4b, can be functionally integrated, for example, in the design of the extractor 250, as shown in figs. In this case, the control signal, which contains information about the signal component at the first input 260a, is not modified by the echo analysis filter 260.

Naturally, an option is provided for integrating the pre-analysis filter of the echo signal in FIG. 4B with the extractor 250 in FIGS. 3a and 3B, when, for example, only one of the two component signals should be modified at the filtering stage 360. Here, a mirror version of the echo analysis filter 260 may be used, filtering the input signal at the second input 260b.

FIG. 5A shows a depictor 270 implemented in accordance with the present invention in combination with the device 200 of FIG. In this case, the calculator 270 also has a first input 270a and a second input 270b. The calculator 270 further includes first and second calculators of the filtering parameters 370-1, 370-2, the inputs of which are connected, respectively, to the inputs 270a, 270b. More specifically, the input of the filter computer 370-1 is connected to the first input 270a in order to receive, say, the stationary component of the signal. Accordingly, the second filter transmitter 370-2 is coupled to the second input 270b to receive the non-stationary component of the signal, for example, from the extractor 250, as shown in FIG. 3A or 3b. In the event that a second echo analysis filter 260 is mounted between the extractor 250 and calculator 270, a signal derived from the respective component signals is transmitted to both filter parameter calculators 370.

The outputs of both filter parameter calculators 370 are connected to a combiner 380, whose output, in turn, is connected to an output 270d. The calculator 270, shown in FIG. 5A as an arbitrary component, has a third input 270c, which is connected to both filter parameters calculators 370 inside the circuit of the calculator 270, and, as also shown in FIG. 2, is directly or indirectly connected to the microphone input terminal 280 signal.

Taking into account the operating mode of calculator 270, both filter characteristics calculators 370 are intended to calculate, based on the received component signals and, possibly, the microphone signal taken from the input terminal 280, the corresponding filter coefficients, which are later transmitted to combinator 380. Both filter parameter calculators 370 carry out the corresponding calculations based on the received derivatives of the corresponding component signals, which could be adjusted by the second pre-filter analysis of the echo signal 260. However, regardless of this, the calculators of the characteristics of the filter 370 are designed to calculate the transmittance of the first and second filters, respectively, based on the components of the signals received from the extractor 250.

Then, the transmittances of the first and second filters calculated in this way are combined using a combinator 380 into a set of filter coefficients, after which they are input to the adaptive filter 210 through the output 270d of the calculator 270. Such a combination can be performed using a series of different operations. The combination of the transmittance of the first and second filters depends on the individual technical solution of the filter, as well as, to a large extent, on the involved time-frequency converters 230, 290 and the frequency-time converter 300 associated with them, so that the transmittance of the adaptive filter 210 can be calculated. relevant examples are given.

On figv depicted the second calculator 270, similar to the calculator on figa. It differs from the calculator 270 in FIG. 5A in that the combinator 380 in FIG. 5A is replaced by a selector 390, which is designed to output to the output 270d a set of filter coefficients generated based on the transmittances of the first and second filters, which is based either on the transmittances of the first filter of the first filter transmitter 370-1 or at the transmittance of the second filter of the second filter transmitter 370-2. Formulated differently, the selector 390 is designed to determine the transmittance of the adaptive filter 210 based on either a stationary or non-stationary component signal.

With this architecture, the selector 390 can implement more complex mathematical dependencies based on the corresponding set of filter coefficients of the calculator 370. However, it differs from the combinator 380 as part of the calculator 270 in FIG. 5A in that it takes into account the combination of two sets of filter coefficients provided by the filter parameter calculator 370

FIG. 5C shows yet another embodiment of the calculator 270, which differs from the calculator 270 in FIG. 5A in that the filter parameter calculator 370 is only connected to the first input 270a therein. Additionally, the filter parameter calculator 370 in the structure of the calculator 270 in FIG. 5C is connected to the second input 270b to obtain through it the parameters necessary to determine the filtering coefficients. In addition, as an option, the filter parameter calculator 370 in FIG. 5C can be connected to a third input 270 s to enable calculation of filter coefficients taking into account the microphone signal.

Thus, the calculator 270 in FIG. 5C can work in conjunction with the extractor 250 shown in FIG. 3C and the second echo pre-filter 260 shown in FIG. 4B. Through the second output 250d of the extractor 250, the parameters necessary for calculating the corresponding filter characteristics calculator 370 are transmitted to it directly through the second input 270b of the calculator 270. For this, the second input 270b is connected to the parameter input terminal of the filter characteristics calculator 370, through which auxiliary parameters can be entered for computing filter coefficients.

Due to the fact that only a single filter parameter calculator 370 is included in the loop of the calculator 270, the combinator and selector may not be used.

On fig.5D shows the version of the calculator 270, in configuration and functions performed similar to the analogue in figs. In contrast to the calculator 270 in FIG. 5C, the design of the calculator in FIG. 5D is supplemented by a parameter determiner 400 mounted between the second input 270b and the corresponding connector for inputting the parameters of the filter characteristics calculator 370.

In contrast to the calculator 270 in FIG. 5C, the calculator 270 in FIG. 5D can function as part of the device 200, where the extractor 250 outputs a control signal through the second output 250d that contains information about which of the two component signals it transmits through the corresponding first output 250s If the parameters required by the filter parameter calculator 370 for calculating the filter coefficients for the two signal components are different, or if the signals generated on the basis of the difference are different, the corresponding parameters can be calculated taking into account the component signal transmitted by the parameter determiner 400 while using the calculator 270, as shown in fig.5D. By virtue of this, the parameter determiner 400 may be implemented, for example, in the form of a memory unit or computing device. Implementation as a means of storing information is possible in the form of read-only memory (ROM / ROM), non-volatile memory (NVM / ENZU) or random access memory (RAM / RAM).

FIG. 5E shows another embodiment of a calculator 270, which incorporates two filter parameter calculators 370-1 and 370-2, whose operation depends on a component signal, on the basis of which the transmittances of the adaptive filter 210 should be calculated. Here, the inputs of both filter parameter calculators 370 are connected to the first input 270a. In addition, each of the two filter parameter calculators 370 can be arbitrarily connected to the third input 270c, and also to the input of the distributor 410, the output of which is coupled to the output of the calculator 270. The distribution unit 410 has an additional input for the control signal connected to the second input 270b calculator 270.

Thus, the calculator 270 in FIG. 5E allows the first set of filter coefficients to be calculated using the filter parameter calculator 370-1 and the second set of filter coefficients using the second filter parameter calculator 370-2 based on the signals received at the first input 270a. The choice of one of the two filter coefficients calculated by the filter parameter calculator 370, which ultimately must be sent to the output 270d, depends on the control signal received at the input for the commands of the setting action of the distributor 410 through the second input 270b. Depending on the control signal received at the input for the control signals, a distributor 410 connects one of the two inputs to the output 270d.

Thus, the calculator 270 in FIG. 5E can act, for example, in combination with an extractor 250, as shown in FIG. 3C, where a control command containing data of the component signal transmitted through the first output 250c is supplied through the second output 250d. Therefore, the design of calculator 270 shown in FIG. 5C can be applied, for example, in cases where the initial data for the two component signals calculated by the filter parameter calculator 370 have such differences that they cannot be effectively converted by parameters of simple changes.

It would be logical to note here that the types of extractor 250 shown in FIGS. 3A-3C, the filter options for preliminary analysis of the echo signal 260 shown in FIGS. 4A-4B, and the modifications to the calculator 270 shown in FIGS. 5A-5E may be mutually integrated in accordance with a specific application. Suppose that a filter parameter calculator 370 selector device with subsequent output to output 270d performs further manipulations, say calculations, based on filter coefficients, then it can be used, for example, in the model of calculator in Fig. 5E instead of distributor 410.

The design solutions of the invention presented above in the description and in FIG. 2 to 5 are the latest developments providing separate suppression of stationary and non-stationary components of the acoustic echo. This is achieved by evaluating the echo separately from the unstable and stable components of the loudspeaker signal. Further, the design according to this development provides for calculating the characteristics of two respective echo cancellation filters for both types of signals. Echo cancellation filters can be individually optimized to maximize echo cancellation and minimize artifacts and signal distortion at the transmitting end.

The further process is structured as follows. First, a speaker signal is modeled. Then, in accordance with this model, stationary and non-stationary components are separated, which can be done based on the assessment of stationary components. After that, the power spectra of the stationary and non-stationary components of the echo are evaluated using filters of preliminary analysis of the echo signal. Because of this, in some implementations of the present invention provides for the calculation of the parameters of two echo canceller filters. Subsequently, the process of separation of stationary and non-stationary components can be adjusted based on the practice of using echo cancellation filters.

Speaking about signal modeling, it should be taken into account that the estimation of the spectrum of the introduced echo or the power density spectrum of the echo signal using the preliminary echo analysis filter in practice, as a rule, is not very accurate, since only a part of the true length of the echo path is available for consideration. In order to prevent residual echo due to a high degree of error, echo cancellation filters are set to an aggressive echo cancellation mode, in which the residual echo is completely removed. This tuning is achieved by overestimating the spectrum of the power density of the echo signal and smoothing over time, which helps to maintain a low gain filter.

When the loudspeaker signal contains stationary noise, the echo suppressor attempts to block the echo. The use of the above filters of aggressive echo cancellation often leads to jamming not only stationary echo noises, but also to weaken stationary noise and speech at the near end.

It is proposed here to mitigate the problem by means of two different ways of suppressing the echo of stationary and non-stationary signals, as illustrated in FIG. 6.

6 is a schematic block diagram of a device 200 in accordance with the present invention, which includes a loudspeaker 100 and a microphone 110. A signal x [n] is supplied to the loudspeaker 100, which also goes to the extractor 250. Another name for the extractor 250 is the discriminator of the stable state (stability selector). As explained in connection with FIG. 2, the extractor 250 is provided with two outputs connected to the calculator 270. In addition, the signal y [n] of the microphone 110 is supplied to the calculator 270.

In the context of FIG. 5A, a calculator 270 has already been shown, which includes a first filter parameter calculator 370-1 for the stationary signal component and a second filter parameter calculator 370-2 for the non-stationary signal component, which are generated at the output of the extractor 250. Additionally, both parameter calculators Filtration 370 receive a microphone signal.

Based on the received signals, the filter parameter calculator 370 calculates the filter coefficients H _w and H _s , which are transmitted to the combinator 380. For this purpose, the output of each of the two filter parameter calculators 370 is connected to the combinator 380. The combinator 380, in turn, outputs to the adaptive filter 210 filtering coefficients calculated or specified based on two sets of filtering coefficients H _w and H _s .

In order to obtain a signal with a blocked echo e [n] from the microphone signal y [n], the microphone signal is additionally directly input to the adaptive filter 210. At the output of the adaptive filter 210, a signal with a blocked echo e [n] is generated.

Thus, adaptive filter 210 performs echo cancellation, in which two filter parameter calculators 370 are involved, each of which calculates echo cancellation filters in the form of respective filter transmittances, which are then integrated by combiner 380 into an effective echo cancellation filter.

Turning to Fig.6, it should be borne in mind that the schematic diagram presented on it is a simplified block diagram, where, for example, implemented time-frequency converter or filters for preliminary analysis of the echo signal are not shown.

Suppression of unsteady (speech) echo should be performed aggressively to avoid residual echo, which is an annoying factor. However, stationary echoes, which may result from stationary interference in the loudspeaker signals, are usually suppressed less aggressively in order to prevent the occurrence of artifacts such as tonal distortions. In order to generate a proper model, the signal x [n] coming from the speaker can be broken down into components according to the expression

where x _s [n] models the non-stationary component of the speech signal, ax _w [n] simulates stationary interference. Discrete time is indicated by a variable exponent n.

First, on both sides of the model, expressed by equation (1), a short-term Fourier transform is performed taking into account the separation of stationary and non-stationary components, from which we obtain

where m, the frequency indicator, and k, the time indicator of the data block, are integers. In equation (2), the non-stationary and stationary components of the power spectral density of the loudspeaker signal | X [k, m] | ² expressed as

| X _s [k, m] | ² and | X _w [k, m] | ² .

It is logical to assume that x _s [n] and x _w [n] do not correlate and have an average value approaching zero. It follows that | X [k, m] | ² is derived from

составляющей сигнала громкоговорителя x_s[n] может быть восстановлена путем вычитания оценочного спектра мощности стационарной компоненты сигнала из спектра мощности сигнала громкоговорителя |X[k,m]|² согласно уравнениюBecause of this, the instantaneous spectral power density of non-stationary

the loudspeaker signal component x _s [n] can be restored by subtracting the estimated power spectrum of the stationary signal component from the power spectrum of the loudspeaker signal | X [k, m] | ² according to the equation

оценивается посредством фильтрации спектральной плотности мощности сигнала громкоговорителя |X[k,m]|², следуя выражениюIn fact

estimated by filtering the spectral power density of the speaker signal | X [k, m] | ² , following the expression

The filter F _x [k, m], also called the gain filter, can be written in its original form according to [W.Etter and GSMoschytz. Noise reduction by noise-adaptive spectral magnitude expansion. J. Audio Eng. Soc., 42: 341-349, May 1994] as follows:

where γ _x is the exponent, and β _x is the control parameter or the parameter for adjusting the intensity of suppression of the components of the stationary signal in case the expected value of this characteristic was underestimated or overestimated. The separation of stationary and non-stationary components will be demonstrated in the context of Fig. 8 at a frequency index of 1 kHz.

Equations (5) and (6) describe the functionality of the gain filter 330 in the extractor 250, presented in the context of figure 2-6.

времени. В каждом блоке данных (фрейме) k энергетический спектр шума корректируется путем однополюсного усреднения при двух постоянных времени с целью дискриминации речи и шума. Короткий период атаки показывает, что текущий блок данных содержит шум.Estimation of stationary interference can be performed by adjusting the expected short-term noise power spectral density

time. In each data block (frame) k, the noise energy spectrum is corrected by unipolar averaging at two time constants in order to discriminate between speech and noise. A short attack period indicates that the current data block contains noise.

A long damping time constant indicates that the current data block contains speech.

In practice, this is implemented in accordance with

where µ ₁ is the attack time constant and µ ₂ is the decay time constant. It should be clarified that μ ₁ and μ ₂ in equation (7) are dimensionless parameters for which the condition μ ₁ <μ ₂ is valid. However, taking into account the sampling rate, these parameters can be interpreted and perceived as, for example, the aforementioned time constants. As subsequently confirms the proportionality relation (16), the actual time constants and these parameters are inversely proportional to each other. The attack time constant µ ₁ can, for example, be 10,000 ms = 10 seconds, while the attenuation time constant, taking into account the sampling period, can be 10 ms.

When implementing the present invention, shown in Fig.2-6, the functions described by equation (7), performs averager 320 as part of the extractor 250.

When determining the power of the echo signal, the evaluation of the echo spectrum can be performed by applying the filter of preliminary analysis of the echo signal G [k, m] to the time-delayed power spectrum of the loudspeaker signal according to the equation:

where | Ŷ [k, m] | ² expresses an estimate of the spectral density of an echo power in a microphone signal. It follows from the applied equation (3) that the echo generated from the non-stationary component of the loudspeaker signal is determined from

and the echo formed from the stationary component of the loudspeaker signal is determined from

Depending on one of the constructive solutions of the present invention, the variants of which are shown, for example, in the context of FIGS. 2-5, the working functions described by equation (8) can be performed, say, by the first filter of preliminary analysis of the echo signal 240. The technical capabilities expressed by equations (9) and (10) may be possessed by a second filter for preliminary analysis of the echo signal 260, which contains two filter links 360-1 and 360-2.

As already mentioned above, the delay of signals by the value of d used in equations (8) - (10) can also be realized by filters for preliminary analysis of the echo signal 240, 260. Instead, the delay can be provided by a time-frequency converter 230, if this function is not assigned to another target standalone device.

и

.To compensate for the echo, the interacting echo cancellation filters H _s [k, m] and H _w [k, m] are calculated and used to eliminate the echo of the microphone signal based on estimates of non-stationary and stationary echo signals

and

.

A structure including a calculator 270, shown in particular in FIG. 5A, may in practice be implemented in accordance with

In this case, the multiplication of the respective transmission coefficients of the echo canceller filters H _s [k, m] and H _w [k, m] is performed by a combinator 380, replacing the serial connection of the respective echo cancellation filters. The multiplication of the desired filtering coefficients in the frequency domain corresponds to the convolution of the corresponding impulse characteristics in the time domain.

By implementing echo cancellation filters by factoring according to equation (II), different gain factors can be introduced as parameters of the various echo components. The components of a non-stationary echo cancellation filter can be calculated, for example, by following the equation:

and the components of the stationary echo cancellation filter can be calculated in accordance with

The calculated parameters β _s , γ _s , β _w and γ _w can be used to control the planned mode of operation of each of the echo canceling filters. Depending on the specifics of the implementation of the proposed development, the named calculated indicators can be selected and fixed, predefined in any required adaptable, programmable or modifiable form. The standard set of exponential parameters is, in particular, γ _s = γ _w = 2.

The so-called redundancy ratios β _s and β _w are used to control the degree of aggressiveness of echo jamming. Say, the intensity of some echo cancellation filter can be enhanced by increasing the redundancy coefficient of the estimate. Therefore, usually the compensation filter for the stationary echo signal H _w [k, m] is selected taking into account the parameter β _w = 2, due to which a moderate attenuation of the echo is achieved.

On the other hand, an echo cancellation filter responsible for suppressing the non-stationary components of an echo implies an aggressive suppression mode to effectively attenuate the noise components of speech in the echo signal. Due to this, the estimated redundancy coefficient of the estimate β _s often exceeds β _w , and, accordingly, β _s > β _w . For example, when choosing β _w = 2, β _s can have values in the range 20> β _s > 2 = β _w , (suppose β _s = 4). In most cases, β _w and β _s are of the same order of magnitude.

The threshold values L _s and L _w set the maximum decay limit of the echo in decibels (dB). The nominal value for the stationary echo cancellation filter is L _w = -10 dB or -15 dB, which optimally limits the attenuation of stationary echoes, reducing the likelihood of random artifacts. In the presence of an unsteady voice signal at the far end of the communication line, the interfering echo components must be completely eliminated, which is achieved by setting the limit value L _{s of} about -60 dB for the unsteady component signal.

The design capabilities described by equations (12) and (13) can be implemented in the filtering parameter calculator 370 included in the structures described using FIGS. 2-6.

Sometimes the practical purpose of the proposed development does not require direct echo cancellation using echo cancellation filters, as equations (12) and (13) describe, but rather, compensation of the echo signal based on a comparable time-smoothed interpretation. Like the calculated parameters discussed above, the parameters of temporary smoothing usually require manual tuning and optimization of the suppression of separately non-stationary and stationary echoes. This improves the quality of perceived sound, since the specifics of suppressing stationary noise components differ from the features of compensation for non-stationary components of a voice signal.

These functionalities can be implemented, for example, by the filter parameters calculator 370 itself or by any device on the supply side from them, for example, combinator 380, selector 390, or distributor 410. If necessary, such time smoothing can be performed directly using adaptive filter 210.

Speaking about improving sound quality, it should be remembered, for example, that smoothing should be enhanced when the stationary components of the signal are suppressed in order to avoid the so-called tonal distortions, which is described in [O. Capp'e. Elimination of the musical noise phenomenon with the ephrain and malah noise suppressor. IEEE Trans. Speech and Audio Processing, 2 (2): 345-349, April 1994.]. At the same time, during attenuation, low throughput characteristics of non-stationary echo compensation filters should be ensured in order to maintain a sufficiently high degree of attenuation of the echo signal, since the reflection of the echo path contributes to this attenuation. However, this should not reduce the adaptability of the adaptive filter 210 in cases of rapid changes in echo levels. The discussion of quality aspects clearly shows the need for an individual approach to debugging and optimization of echo cancellation filters defined by equations (12) and (13).

Figure 7 presents an expanded, more complete, basic diagram of the process or a block diagram of the attenuation algorithm of the acoustic echo, which is discussed further. The design solution shown in Fig. 7 is in many respects similar to the embodiment in Fig. 2. Here, a time-frequency converter 230 is also introduced into the device 200 in the configuration of a window Fourier converter (OPF) receiving a speaker signal x [n].

At the top of FIG. 7, an example is a curve 420 of the speaker signal x [n] as a function of time passing over time values n.

In addition to converting the signal x [n] from the time domain to the frequency domain, the time-to-frequency converter 230 performs a delay of the amount d mentioned above. As a result, the spectrum X [kd, m] is formed at the output of the time-frequency converter 230, which is, as a rule, complex-valued. This spectrum X [kd, m] is transmitted to an extractor 250, indicated in FIG. 7 by the letters SD [DS] (stationary discrimination [= stability discriminator]). With regard to the technical solutions of the invention shown in FIGS. 2-5, it has already been explained that the extractor 250 is also capable of generating, as shown in FIG. 7, a stationary component signal X _w [k, m,] and a non-stationary component signal in the frequency domain X _s [k, m]. These component signals are fed to a calculator 270.

In addition, the design of the device 200 of FIG. 7 includes a time-frequency converter 290, also implemented in the form of a window Fourier converter (OPF). In the upper part of Fig. 7, it is shown as an example that the time-frequency converter 290 receives the microphone signal y [n] displayed by curve 430 at the input. The time-frequency converter 290 also converts the microphone signal Y [k, m] into the corresponding frequency representation, while the indicator k still indicates a data block, and the indicator m indicates a frequency band or a frequency value or a transform coefficient. The indicated spectrum Y [k, m] is also usually complex-valued.

However, unlike the time-frequency converter 230, the time-frequency converter 290 does not perform an additional delay function. Basically, due to the fact that this is not necessary, since the speed of propagation of sound waves (speed of sound) is noticeably lower than the speed of transmission of electrical signals in circuits and circuit elements, which causes the signal y [n] picked up by the microphone to be delayed relative to the correlated loudspeaker signal x [n].

For clarity, this aspect is highlighted by the first curly bracket 440 as a fragment of the speaker signal x [n] when comparing the graph curves 420 and 430 at the top of FIG. In the same part of FIG. 7, in the microphone signal y [n], the portion corresponding to the fragment of the speaker signal x [n] indicated by the bracket 440 is shown by the second curly bracket 450. Thus, the speaker signal x [n] and the microphone signal y [n ] are offset relative to each other by the amount of delay d, which in Fig.7 is indicated by arrow 460.

. Эти коэффициенты фильтрации также передаются на вычислитель 270.In accordance with the algorithm of Fig. 7, the spectra of the loudspeaker and microphone signals are sent to the preliminary analysis filter of the echo signal 470, which on the basis of the received signals determines the parameters of the preliminary analysis filter of the echo signal

. These filter coefficients are also transmitted to a calculator 270.

The calculator 270 in Fig. 7, in turn, consists of two calculators of the filtering parameters of the stationary and non-stationary components of the signal, 370-1 and 370-2, each of which receives at its input a spectrum of the microphone signal and

. Таким образом, функциональные возможности, заложенные в обоих вычислителях параметров фильтрации 370, предусматривают не только расчет параметров фильтрации, о чем говорилось в описании к фиг.2-5, но и выполнение операций второго фильтра предварительного анализа эхо-сигнала 260.transmission coefficients of the filter of the preliminary analysis of the echo signal

. Thus, the functionality embedded in both calculators of the filtering parameters 370, provide not only the calculation of the filtering parameters, as described in the description to figure 2-5, but also the execution of the second filter preliminary analysis of the echo signal 260.

Both filter parameter calculators 370, indicated in FIG. 7 as ERF (echo cancellation filter), as in FIG. 5A, are connected to combiner 380, which is indicated in FIG. 7 as FC (combination of filter parameters). Combinator 380 combines filter coefficients obtained from both filter parameter calculators to produce transmittances of adaptive filter 210.

Further, as already mentioned in the description of FIGS. 2 and 5, the combinator 380 is connected to an adaptive filter 210, which is indicated in FIG. 7 as SM (spectral modification). Adaptive filter 210 performs spectrum adjustment based on the received spectral representation Y [k, m] of the microphone signal y [n] to attenuate or suppress the echo component of the microphone signal.

Finally, the adaptive filter 210 is coupled to a time-frequency converter 300, which is an inverse Fourier window converter (OOPF). This transducer generates at the output a signal e [n] in the time domain, cleared of echo. Embodiments of the present invention in the form of appropriate methods or devices 200, as shown, for example, in FIG. 7, can eliminate artifacts introduced when the spectrum is modified by the adaptive filter 210. Defining otherwise, the proposed structural solutions in accordance with the invention provide adaptive power control. If there is speech only at the far end, the echo cancellation process should proceed in an aggressive enough mode to prevent the passage of any signal, since in such a situation it may not be desirable to separate non-stationary and stationary signals and component signals. For this reason, when recognizing such a situation, it may be necessary to adapt the control parameter β _x from equation (6), which regulates or at least affects the amplitude of the stationary component signal subtracted from the loudspeaker signal.

To distinguish a situation in which the output signal of the speaker contains only speech at the far end of the communication line, two different parameters are calculated. The first of these is the value of the so-called prediction gain (prediction gain) corresponding to full-band averaging of the coherence functions between the loudspeaker channel and the microphone channel. As a second parameter, an indicator of the speech activity of the loudspeaker channel is used, which can be obtained, for example, by comparing the speaker signal levels different in time or selected from special parameters of the voice codec used in voice transmission. The word "codec" is composed of two abbreviated English words - an encoder and a decoder, and such codecs can be built, for example, on the basis of LPC (linear prediction coding) or CELP (linear prediction with multi-code control).

The prediction gain or echo prediction gain ω [k] describes the level of similarity between the microphone signal and the delayed speaker signal. The prediction gain ω [k] is calculated based on the quadratic coherence function between the delayed power spectrum of the speaker signal | X _d [k, m] | ² and the power spectrum of the microphone signal in accordance with the equation

where E {...} stands for mathematical expectation. The named expected mathematical value can be obtained using a short-term estimate of the coherence function Г _d [k, m] by calculating or approximating the expected value according to the equation

Index α determines the degree of smoothing of the estimate over time. This indicator is related to the time constant, since equality (15) approximately corresponds to exponential attenuation.

The time constant Tα of the exponential decay in seconds is approximately

where f _s denotes the sampling rate. In other words, the proportionality relation (16) shows how coefficients that are actually dimensionless (here, α), related to the sampling frequency f _s , can be represented as a time constant (here, T _α ).

After that, the gain from the prediction ω [k] is calculated as the average value of the coherence functions Г _d [k, m] at the frequencies indicated by the indices m = 0, ..., M-1, in accordance with

where M shows the number of frequency bands. An echo gain close to 1 indicates that the microphone signal can be (almost) completely predicted based on the delayed speaker signal. As a result, the probability that the microphone signal contains only speech at the receiving end tends to 1. In this case, the setting parameter β _x can be adjusted depending on the value of the gain from the prediction ω. A high prediction gain is an indication of the presence of a voice signal only at the far end, and echo attenuation should be aggressive enough to eliminate all (echo) signals. Thus, the interference is kept within the non-stationary path and is eliminated at the lowest limit value L _s in decibels (dB), since the setting parameter is chosen as β _x = β _w = 0. A low prediction gain indicates a possible presence of a speech signal at both the near and far ends, and echo cancellation should be less aggressive to avoid artifacts. In this case, the interference is compensated through the stationary path and eliminated at the limiting value of L _w in decibels (dB). Here, the parameter β _x = β _w is applied.

In this case, however, it should be noted that the gain from the prediction can also be high if the loudspeaker signal contains only the noise picked up by the microphone in the absence of a voice signal. To avoid choosing an overestimated value of the control parameter β _x , which can lead to excessive suppression, a second control parameter is used - an indicator of speech activity in the loudspeaker channel. Because of this, the above rules for calculating the control parameter β _x as a function of enhancing the prediction ω in practice are applied only if there is speech activity in the loudspeaker channel.

In the embodiment of FIG. 7, the technological operations described by equations (14) through (17) are performed by a calculator 270, including two filter parameter calculators 370, and a combinator 380. Various design options shown in FIGS. 2-5 provide for use by the calculator 270 of not only the microphone signal received by it through the input 280, which is shown as an option in FIG. 2, but also of the unmodified loudspeaker signal received through the input 220.

Next, the principle of signal processing using structures implemented in accordance with the present invention and shown in figures 8 to 10 will be graphically illustrated in detail.

On Fig shows the process of separation or extraction of stationary and non-stationary components of the loudspeaker signal. Section (a) of FIG. 8 is a graph of the spectral power density of a speaker signal at a frequency of 1 kHz as a function of time in the range of about 5 to 7.5 seconds. The abscissa of the graph in section (c) of figure 8 applies to all three sections (a) - (c). Graph (b) gives a graph of the power spectral density of the non-stationary component, and graph (c) shows the corresponding power spectral density of the stationary component of the signal.

The non-stationary component or the conjugated non-stationary component signal displayed on the graph 8 (b) contains the upper values corresponding to each case of the increase in the power spectral density on the graph 8 (a). It should be noted that in the intervals between these ranges the unsteady component almost completely disappears.

In contrast to the non-stationary component in Fig. 8 (b), the stationary component in Fig. 8 (c), determined using floating recursive averaging according to equation (7), graphically displays obviously lower amplitudes and, due to floating averaging, is an obviously more gentle curve. In particular, the stationary component in Fig. 8 (c) and / or the corresponding stationary component signal have an exponential or similar exponential decrease over a period of about 6.4 seconds, as mentioned in connection with the proportionality relation (16). Such a decline is a consequence of the absence in the power spectrum in this range of Fig. 8 (a) of large quantities that correspond to voice signals. Spectral components that go beyond the stationary component, respectively, are removed.

Based on the data of FIG. 8, FIG. 9 displays respective echo cancellation filters. More specifically, Fig. 9 shows two correlating curves of two interconnected echo cancellation filters H _s and H _w at a frequency of 1 kHz, calculated on the basis of equations (12) and (13). So, FIG. 9 (a) shows the function of the echo cancellation filter H _s calculated according to equation (12) for the non-stationary component at a frequency of 1 kHz. Section (b) gives a graph of the function of the corresponding echo canceller filter H _w for the stationary component, constructed in accordance with equation (13).

Figure 10 graphically represents identical parameters in an extended time scale displayed on the abscissa of the graph 10 (c) and relating simultaneously to figures 10 (a) and 10 (b). Figure 10 shows the dependence on the value of the control parameter β _x - in graph 10 (a) - the process of separation of stationary / non-stationary components and - in graph 10 (b) - gain from the prediction ω against the background of the speech activity of the loudspeaker channel displayed in graph 10 (from).

Determining more precisely, figure 10 clearly demonstrates the interdependence between the control parameter β _x and two other control parameters - ω and speech activity, introduced and described earlier. The first third of the process modeled in FIG. 10 proceeds under conditions characterized by the presence of a speech signal only at the far end and a high gain from the prediction. In this case, the control parameter β _x is set to β _x = β _w = 0, which corresponds to the aggressive mode of suppressing the non-stationary component and the complete suppression of the stationary component.

The second third of the model displays a situation characterized by the presence of a speech signal only at the far end and the possibility of its recognition with a low prediction gain ω and the absence of speech activity in the speaker signal. In this case, the value of the control parameter β _x is set so as to ensure that all stationary components pass along the stationary path and eliminate them at a low level of aggressiveness, preventing artifacts from being introduced. The last third of the model reproduces a dialogue mode in which the control parameter β _x varies between low values when there is speech activity in the speaker channel and higher values when speech activity is not recognized.

In the embodiments of the invention described above, including FIG. 6, which is also a general block diagram of the corresponding algorithm, the separate suppression of the stationary and non-stationary components of the echo is carried out not by separating the corresponding loudspeaker signals, but by evaluating the echo signal as a whole.

In all considered versions of the invention, the evaluation of the spectral density of the power of the echo signal was performed by applying the filter of preliminary analysis of the echo signal G [k, m] or G [k, m] ² to the delayed version of the power spectrum of the signal of the speaker according to equation (8), where

an estimate of the echo power spectrum contained in the microphone signal. As a result of dividing the power spectrum of the loudspeaker signal in accordance with equation (3) into stationary | X _w [k, m] | ² and non-stationary | X _s [k, m] | ^{The 2} components of the echo arising from the non-stationary component of the loudspeaker signal are calculated using equation (10), and the echo signal arising from the stationary components of the loudspeaker signal is obtained from equation (9).

Using estimates of non-stationary

and stationary

echo signals, the corresponding echo cancellation filters H _s [k, m] and H _w [k, m] can be calculated. These echo cancellation filters are then combined and used to suppress the echo in the microphone signal in accordance with the equation

where H [k, m] is derived from

One possible way to combine different echo cancellation filters H _s [k, m] and H _w [k, m] is to use their output in accordance with equation (11), which is similar to connecting two filters in series.

Another possible way is to use an appropriate minimum of echo cancellation filters according to

where the function min (...) represents the minimum of the corresponding quantities. In other words, in this case, the correspondence (...) = min (...) is applied.

As explained above, these calculations can be performed, in particular, by combinator 380, and in addition, by selector 390 or distributor 410. In addition, individual devices are able to perform more complex operations of combining and calculating individual blocking filters, based, for example, on linear constructions or nonlinear equations. Also, it is possible to combine not only band signals, but also groups of band signals or the entire set of band signals.

By combining echo cancellation filters for different echo components, different gain factors can be introduced. The non-stationary echo cancellation filter is calculated according to equation (12), and the stationary echo cancellation filter is calculated according to equation (13).

Often, in practice, echo cancellation is not based on the direct application of echo cancellation filters in accordance with equations (12) and (13), but on the basis of the corresponding time-smoothed versions. Like the calculation parameters described above, the temporal smoothing parameters can also be manually adjusted separately to suppress non-stationary and stationary echo signals. Thus, the quality of the perceived sound can be improved, since the requirements for suppressing stationary noise components are different from the requirements for compensating for non-stationary components of a voice signal.

Say, it is well known that to suppress the stationary components of a signal, more intensive smoothing is required to avoid the so-called musical tones. On the other hand, the anti-aliasing given to the non-stationary echo compensation filters should maintain lower values to sufficiently suppress the echoes introduced by the long components of the echo path, or the long tails of the echo path. However, this should not interfere with the passage of rapidly changing echo levels. The discussion of the quality of perception clearly shows the need for individual adaptation and optimization of two types of echo cancellation filters in accordance with equations (12) and (13).

The following approach to the implementation of the proposed method and / or device consists in the separate application of echo cancellation filters to the stationary and non-stationary components of the signal.

Figure 11 is a schematic block diagram of a device 200 including an adaptive filter 210. The uniformity of the structures allows further reference to the structural solutions presented in figures 2-5, 6 and 7.

The device 200 according to the invention includes a speaker 100 or a terminal for connecting a speaker 100, or an input for a speaker signal x [n]. The time-frequency converter (RFP) 230, designated as a DFT (discrete Fourier transform) in the diagram, converts the speaker signal x [n] into a transform in the form X [k, m]. Then, the loudspeaker signal is supplied to a delay device 480, which generates a delayed signal X [k-d (k, m), m] with a delay value d (k, m).

From the delay device 480, the delayed signal is transmitted to the first filter of the preliminary analysis of the echo signal 240, which is based on filtering coefficients

G [k, m] generates an estimated echo signal

Echo Evaluation

sent to the extractor 250, which, based on the spectral coefficients of such an estimated echo signal, generates non-stationary and stationary power spectra of this signal as (derivatives) of the loudspeaker signal.

After that signals

и

and

output from the extractor 250 to the calculator 270.

The signal y [n] of the microphone 110 is input to a time frequency converter (RFP) 290, abbreviated as DFT, which converts it from the time domain to the spectral representation Y [k, m]. The converted signal enters the energy level calculator 490, which, taking into account the spectral components of the microphone signal, calculates their spectral power density by squaring the (absolute) value of each indicator. The power spectrum obtained in this way is also input into a calculator 270, which, in parallel with the power spectra described above, calculates two echo cancellation filters H _s [k, m] and H _w [k, m], the transmittances of the current adaptive filter H [k, m] and transmits them to adaptive filter 210.

The adaptive filter 210 is simultaneously coupled to the output of the time-frequency converter 290 and, therefore, also receives the spectral components Y [k, m] of the microphone signal y [n], from which it generates a signal with a blocked echo in the frequency domain H [k, m], given filtering coefficients H [k, m]. Then this signal with a compensated echo is transmitted to a time-frequency converter (FWM) 300, indicated in the diagram as an DFT (inverse DFT), which finally converts this signal back to the time domain.

In order to establish the delay value d (k, m) for the delay device 480 and to determine the filter coefficients of the preliminary analysis of the echo signal 240, the representations in the spectral region of the speaker signal X [k, m] and the microphone signal Y [k, m] are input into the corresponding an energy calculator 500, 510, each of which is connected to the output of the time-frequency converter 230, 290, respectively. The energy level calculator 500 is connected to the output of the time-frequency converter 230, the energy level calculator 510 is connected to the output of the time-frequency converter 300.

Each of the two energy value calculators 500 and 510 calculates, like the energy level calculator 490, the power spectral densities, squaring the values of the corresponding spectral components, and outputs the next calculator 520. Based on the values entered into it, the calculator 520 estimates the delay value d ( k, m) and transmission coefficients G [k, m] of the preliminary analysis filter of the echo signal 240. In parallel, the calculator 520 is coupled to a delay device 480 and a filter of the preliminary analysis of the echo signal 240, which are sent via Scientists named the figures.

As follows from the solution in FIG. 11, the corresponding component signals

и

),(

and

),

thus can be divided based on the evaluation of the spectrum of the echo

which is done in accordance with the equation

The specified calculation is performed by the filter preliminary analysis of the echo signal 240.

The determination of the two echo cancellation filters H _s [k, m] and H _w [k, m] according to equations (12) and (13) remains unchanged. The same applies to the definition of the combined echo canceller filter H [k, m]. Therefore, the additional method and associated device 200 shown in FIG. 11 are based on the conclusion that the stationary and non-stationary echo components of the predicted echoes are uncorrelated, so that

из спектральной плотности мощности оценочного эхо-сигнала.Then, the estimated power spectra of the stationary components of the echo can be determined by subtracting the estimates of the stationary component of the echo

from the power spectral density of the estimated echo signal.

In this way,

In practice, the signal | Ŷ _s [k, m] | ² is estimated by filtering the power spectrum of the echo signal calculated by following

Since the used gain filter F _y [k, m] or the square of its value F _y [k, m] ² is determined by analogy with the gain filter F _x [k, m] or F _x [k, m] ² , this component is detailed here not construed. The extractor 250 also performs similar functions using the received signals.

уже известен. Безусловно, подобный способ применим также, когда известен только сигнал расчетной мощности эхо-сигнала

, оцененного с применением уравнения (8). Подобный вариант рассмотрен подробно при описании технического решения, отображенного на фиг.12.It should be noted here that the constructive solution shown in FIG. 11 relates to the case where the estimated echo spectrum

already known. Of course, this method is also applicable when only the signal of the estimated power of the echo signal is known.

estimated using equation (8). A similar option is considered in detail in the description of the technical solution shown in Fig. 12.

Тем не менее, способ на фиг.12 отличается тем, что при нем аттенюация акустического эха построена на разделении стационарных и нестационарных составляющих эха на базе оценки спектральной плотности мощности эхо-сигнала

The flowchart of FIG. 12 illustrates an approach similar to that presented in FIG. 11 for an acoustic echo attenuation algorithm for separating stationary and non-stationary echo components based on the expected echo spectrum.

However, the method of FIG. 12 is characterized in that, with it, attenuation of the acoustic echo is based on the separation of stationary and non-stationary components of the echo based on an estimate of the spectral power density of the echo signal

From the following description, it is obvious that the implementations in figures 11 and 12 are similar to each other not only in their functions, but also in design.

More specifically, a significant difference between the version of FIG. 12 and FIG. 11 is that the energy level calculator 500, which receives and processes the loudspeaker signal x [n] converted to the frequency domain, is mounted not strictly in front of the calculator 520, but is connected directly to the time-frequency output Converter 230 in the configuration of the DFT. With this arrangement, both the calculator 520 and the delay device 480 to the preliminary filter analysis of the echo signal 240 and the extractor 250 no longer receive the actual spectral components, and their power spectra.

Otherwise, the two varieties of the same design in FIGS. 11 and 12 differ only in that similar calculations are performed in them by different elements and devices in slightly different ways. In particular, the extractor 250 does not calculate the energy indices of the individual components of the spectrum, since this was previously done by the energy magnitude calculator 500.

FIG. 13 shows an embodiment of the invention in which, for example, more than one loudspeaker signal or more than one microphone signal is supplied to the device 200. Formulating otherwise, FIG. 13 shows an implementation of a multi-channel device.

Previously, structural solutions of the present invention with separate channels or with a single channel for transmitting only one loudspeaker signal and one microphone signal have been described and discussed, however, this invention is not limited to a single-channel execution, which will be discussed later. The previous design options by analogy can be applied in multichannel acoustic echo jamming systems.

Since the version of the device 200 in FIG. 13 is generally similar to the design in FIG. 2, below, when describing operation modes, connections and other aspects, reference will be made to the description of FIGS.

The multichannel embodiment of the device 200 shown in FIG. 13 has an unlimited number of inputs 220-1, 220-2, ... through which a plurality of speaker signals can be fed to it. Accordingly, the device 200 may include an option in the form of the required number of time-frequency converters 230-1, 230-2, ... to translate or convert the speaker signals from the time domain to the frequency domain, as discussed in detail in connection with FIG. 2.

All time-frequency converters 230 are coupled to the corresponding number of inputs of the grouping device 530, which combines the input loudspeaker signals into a derivative loudspeaker signal, which is then sent to the first pre-analysis filter of the echo signal 240 or to the extractor 250, depending on whether an arbitrary first preliminary analysis filter is mounted echo signal 240. In the context of FIG. 2, it has already been said that the extractor 250 can be connected to an arbitrary second echo-pre-filter feeder 260 or directly with calculator 270. The named calculator generates calculated filtration coefficients at the output.

In contrast to the variant in FIG. 2, the multi-channel device 200 in FIG. 13 also contains in its design a grouping device 540, the input channels of which are connected to the corresponding number of input terminals 280-1, 280-2, ... for microphone signals, possibly via time-frequency converters 290-1, 290-2, ..., also an option. The grouping device 540, like the grouping device 530, generates, based on the received microphone signals in time or frequency representation, a derivative - effective or common - microphone signal, which can optionally be transmitted to the extractor 250 or calculator 270.

Further, the multi-channel device 200, as shown in FIG. 13, comprises adaptive filters 210-1, 210-2, ... for each microphone signal or each microphone input terminal 280, and the connection of adaptive filters 210-1, 210-2 , ... to the corresponding inputs 280-1, 280-2 ... arbitrarily possible through time-frequency converters 290-1, 290-2, ... Also, adaptive filters 210-1, 210-2, ... are connected to the corresponding output terminals 310-1, 310-2 ..., if necessary, through the appropriate number of time-frequency converters 300-1, 300-2, ... At the output of adaptive filters 210, the signals, cleared of the echo or spectrally modified, are sent to output terminals 310 from device 200.

All adaptive filters 210-1, 210-2, ... in parallel are connected to the output of the calculator 270, from which they receive filtering coefficients. In other words, in the embodiment of FIG. 13, the entire set of microphone signals is filtered, from a functional point of view, by the same adaptive filter, that is, based on the same filtering coefficients, in order to obtain spectrally modified or echo-compensated interpretations of the corresponding microphone signals .

Therefore, if x ₁ [n] are speaker signals 1, where 1 is an integer ranging from 0 to L - 1 and where L is the number of speakers or speaker signals, then the same model can be introduced by analogy with equation (1 ) in accordance with

where x _{s, 1} [n] modulates the component of non-stationary speech, ax _{w, 1} [n] modulates the component of stationary noise contained in the signal of loudspeaker 1. In accordance with equation (2), the OPF representation of equation (25) is derived from

Then, using the grouping device 530, which can be seen in FIG. 13, the total group power spectrum of all speaker channels obtained by combining the individual spectra of the speaker signals in accordance with

where L denotes the number of speaker channels. After that, non-stationary and stationary signal components are separated according to equations (5) and (7) taking into account the total or grouped power spectral density, following equation (27).

By analogy with this, the total or combined power spectrum of the microphone channels is calculated in accordance with

where Y _p [k, m] determines the microphone signal 110 p, and P displays the number of microphones. The exponent p is also an integer ranging from 0 to P - 1. In the version of FIG. 13, this calculation can be performed by the second grouping device 540.

To determine the two echo cancellation filters in accordance with equations (12) and (13), the spectra (powers) of the loudspeaker | X [k, m] | ² in accordance with equation (27) and the spectrum (power) of the microphone | Y [k, m] | ² in accordance with equation (28), as already described above. The setting of the control parameter β _x according to equations (14) to (17), described above in the context of monitoring work processes, can also be performed on the basis of general or group spectra according to equations (27) and (28).

Actually, the echo cancellation as part of the spectrum modification is then performed individually for each microphone signal, but using one echo cancellation filter 210 for each microphone channel, following the equation

at p = 0, 1, ..., P-1.

But, as mentioned above, echo cancellation filters 210 can be implemented differently, say, in accordance with equation (19).

It should be noted here that with the multi-channel design of the device 200, as, for example, in FIG. 13, the number of speaker signals L and the number of microphone signals P can be the same or different. In principle, the number of input signals from the speaker and microphone can be any. Moreover, it is not necessary to use both grouping devices 530, 540 for a plurality of speaker and microphone input signals. The invention allows only a plurality of loudspeaker signals to be input using a grouping device 530 without using a grouping device 540 for multiple microphone signals. Such a system is applicable when a single microphone signal from one subscriber at the far end of the communication line arrives at several loudspeakers, for example, during dispatch communication with cars.

Naturally, there is no need to use multichannel grouping device 530 to input one loudspeaker signal, for example, central in a conference communication system, where each of the many participants in the dialogue has a personal microphone. In this situation, the introduction of the 540 grouper is recommended.

It should be added that designs with groupers 530 and 540 can be designed for a larger number of loudspeaker or microphone signals than are received at a particular time. Naturally, in the device 200 may be provided with a larger number of inputs 220, 280 than is used in practice. In such cases, the circuits preceding the scheme, for example, arbitrary time-frequency converters 230, 290 or grouping units 530, 540, are able to independently determine the number of working channels and select the corresponding indicators L and P. Naturally, input of indicators of the number of channels and the expected number of microphone and speaker signals is also provided from the outside.

In addition, the constructive solution presented in Fig. 13, of course, can work with single signals of the loudspeaker and microphone, the corresponding indicators L and P are introduced into the grouping device 530. In principle, equations (27) and (28) are applicable at P = 1 and / or L = 1. Thus, the constructive solution shown in FIG. 13 is a top-down compatible extension of the implementation version of FIG. 2.

Frequency resolution is recommended in the form of a derivative of OPF. The uniformity of OPF in spectral resolution does not correlate very well with the physiology of human hearing. Therefore, it is preferable to rearrange the uniformly distributed coefficients | X [k, m] | ² and | Y [k, m] | ² into the order of disjoint sectors or groups, as shown in [C. Faller and F. Baumgarte. Binaural Cue Coding - Part II: Schemes and applications. IEEE Trans. on Speech and Audio Proc., 11 (6): 520-531, Nov. 2003], containing frequency bands correlated in frequency resolution with the human auditory system, as presented, in particular, in [10].

The sampling frequency of 16 kHz during the short-term Fourier transform normally corresponds to the length of the DFT block of 512 samples and 15 groups or segments, each of which has a passband that approximately corresponds to the double width of the equivalent rectangular passband (ERB / EPT), as described in [B .R. Glasberg and B.C. J. Moore. Derivation of auditory filter shapes from notched-noise data. Hear. Res., 47: 103-138, 1990]. The bandwidths correspond to segments, as shown in FIG.

On Fig shows how the coefficients of the uniform spectrum of OPF can be grouped or decomposed in order to simulate the uneven frequency resolution of the human auditory system. As can be seen in FIG. 14, the frequency axis extends from 0 Hz to about 8000 Hz, which corresponds to an effective bandwidth based on a sampling frequency of 16 kHz.

Gain filters are only calculated for the center frequency of each group. Additionally, this reduces computational complexity compared to the full spectral resolution of uniform OPF. Before applying the gain filter of the last segment or group to a uniform signal of the OPF spectrum, the latter is interpolated by Hann interpolator filters.

FIG. 15 (a) shows Hann interpolation filters useful for smoothing frequency-dependent gain filters. On Fig (b) in a solid line shows the coefficients of the gain filters, interpolated from the values of the gain filters in individual segments, marked, in turn, by thick dots.

Image (a) in Fig. 15 represents Hann filters in detail, image (b) gives an example of gain filter values before and after interpolation. The points in FIG. 15b indicate the values before interpolation, while the solid line corresponds to the values obtained by interpolation. Smoothing the frequency gain filters results in a smoothed version of the spectrum as a function of frequency and thus compensates for musical tones and other artifacts.

The previous description of the options for constructive solutions showed that this invention is implemented by introducing into the proposed design various functional units that perform a certain sequence of operations that make up a given algorithm, summarized below. The implementation of the invention includes the following procedure: receiving at least one loudspeaker signal, receiving at least one microphone signal, converting the loudspeaker signal and the microphone signal into short-term spectra, calculating the power spectral density of the loudspeaker and microphone signals, highlighting or decomposition of the power spectral density into stationary and non-stationary components, calculation of an echo cancellation amplification filter using stationary loudspeaker power spectra, calculating an echo cancellation gain filter using an unsteady loudspeaker power spectrum, applying an amplification filter to the microphone spectrum to suppress the echo signal, inverting the echo-compensated microphone spectrum to the time domain.

Depending on the conditions, the method constituting the present invention can be implemented both in hardware and in software. The invention can be implemented on any digital storage device, in particular a floppy disk, CD or DVD, carrying electronically readable control signals that can interact with a programmable computer system so that the inventive method can be implemented. The implementation of the present invention, basically, is a software or computer program, or software product with program code stored on a machine-readable medium, designed to implement the proposed method, provided that the program is executed on a computer or microprocessor. In other words, the present invention can be implemented as a computer program or software, or a program having program code, for implementing the method of the invention when executing a program using a processor. The processor may be a circuitry element of a computer, a chip card (smart card), an integrated SOC system (SOC = system on a chip), an application integrated circuit (ASIC), or some other integrated circuit (IC).

Claims (24)

1. The device (200) for calculating the transmittance of the adaptive filter (210) of the microphone microphone signal to suppress the echo excited by the loudspeaker signal, including a filter for preliminary analysis of the echo signal (240) to evaluate the spectrum of the echo component or the spectral density of the echo power in the microphone signal; an extractor (250) for extracting the stationary component of the signal and the non-stationary component of the signal (1) from the speaker signal or (2) from a signal derived from the speaker signal, calculated based on the estimated spectrum of the echo component or spectral density of the echo power in the microphone signal; and a computer (270) for calculating the transmittance of the adaptive filter (210) - (1) based on the stationary component of the signal and the non-stationary component of the signal extracted from the speaker signal, and on the basis of the estimated spectrum of the echo component or the energy spectrum of the echo in the microphone signal, or - (2) based on the stationary component of the signal and the non-stationary component of the signal derived from the derived signal. 2. The device (200) according to claim 1, in which the extractor (250) extracts the stationary component of the signal based on averaging the energy indicator of the band-pass signal of the speaker or a signal derived from it. 3. The device (200) according to claim 2, where the extractor (250) performs averaging by finding the floating average value of the current data block on which the band signal is based, and the values of at least one data block that precedes the current time data block. 4. The device (200) according to claim 2, in which the extractor (250) performs averaging by determining a floating average based on a combination of the details of the calculation and based on a comparison of the energy indicator of the current data unit with the energy indicator of the previous data unit or comparison with the value of the previous averaging. 5. The device (200) according to claim 3, in which the extractor (250) performs recursive floating averaging by summing the energy index of the current data block with the result of the previous averaging depending on the addition parameter, the addition parameter being smaller when the energy indicator of the current data block greater than the value of the previously derived averaging, and a larger value when the power indicator of the current data block is less than the value of the previously determined averaging. 6. The device (200) according to claim 1, in which the extractor (250) selects the non-stationary component of the signal based on the strip signal of the speaker or a derivative of it. 7. The device (200) according to claim 1, in which the extractor (250) selects the non-stationary signal component based on the stationary component and the gain filter. 8. The device (200) according to claim 7, in which the extractor (250) is characterized in that the gain filter depends on a variable or non-variable control parameter. 9. The device (200) according to claim 8, in which the extractor (250) sets the control parameter of the gain filter based on the coherence function based on the loudspeaker signal or its derivative, and on the microphone signal or its derivative. 10. The device (200) according to claim 9, in which the extractor (250) sets the control parameter based on the average value of the coherence function over the set of loudspeaker band signals or for a signal derived from them and for the set of band signals of a microphone or signal derived from them. 11. The device (200) according to claim 1, including a grouping device (540), designed to combine multiple microphone signals with the receipt of the total microphone signal or a signal derived from it. 12. The device (200) according to claim 1, in which the extractor (250) generates the stationary signal component and the non-stationary signal component at the output, also comprising the transmitter (270) calculates the transmittance of the first filter based on the stationary signal component and calculates the coefficients transmittance of the second filter based on the non-stationary component of the signal and, in addition, determines the filtering coefficients based on the transmittance of the first and second filters. 13. The device (200) according to item 12, in which the extractor (250) calculates the filtering coefficients so that they correspond to the serial connection of the first filter, which corresponds to the transmittance of the first filter, with the second filter, which corresponds to the transmittance of the second filter. 14. The device (200) according to claim 1, in which the extractor (250) generates a stationary component signal and an unsteady component signal at the output, also comprising a calculator (270) calculates the transmittance of the first filter based on the stationary component signal and transmittance a second filter based on a non-stationary component signal and, in addition, determines the filtering coefficients based on the transmittance of either the first or second filter. 15. The device (200) according to claim 14, in which the calculator (270) sets the filtering coefficients based on those transmittances of the first or second filter that correspond to a higher attenuation level. 16. The device (200) according to claim 1, in which the extractor (250) provides a stationary component signal or a non-stationary component signal, as well as a control information signal containing the parameters of the output component signal, in addition, in which the calculator (270) calculates filtering coefficients based on the output signal of the extractor (250), the signal derived from it, and the data contained in the control information signal. 17. The device (200) according to clause 16, in which the extractor (250) generates at the output a stationary component signal or an unsteady component signal for transmission to a computer (270) depending on the ratio of the energy indicators of the stationary component signal and the unsteady component signal. 18. The device (200) according to claim 1, in which the extractor (250) selects a stationary component signal or an unsteady component signal as expected signals. 19. The device (200) according to claim 1, in which the extractor (250) generates a stationary component signal and an unsteady component signal at the output. 20. The device (200) according to claim 1, including, in addition to the above, an adaptive filter (210), designed to filter the microphone signal using the specified transmittance. 21. The device (200) according to claim 1, including a grouping device (540) and a number of adaptive filters (210), in order to filter at least two of the many microphone signals based on identical filtering coefficients received from the computer (270 ) 22. The device (200) according to claim 1, including a grouping device (530) for combining a plurality of speaker signals and generating an aggregate speaker signal or a signal derived from it. 23. The method of calculating the transmittance of the adaptive filter (210) of the microphone signal, including: estimating the spectrum of the echo component or the spectral density of the echo power in the microphone signal; extracting the stationary component of the signal or the non-stationary component of the signal (1) from the speaker signal or (2) from the signal derived from the speaker signal based on the estimated spectrum of the echo component or spectral density of the echo power in the microphone signal; and calculating the transmittances of the adaptive filter,
- (1) based on the stationary component of the signal or the non-stationary component of the signal derived from the loudspeaker signal, and on the basis of the estimated spectrum of the echo component or the energy spectrum of the echo in the microphone signal; or
- (2) based on the stationary component of the signal and the non-stationary component of the signal derived from the derived signal. 24. Machine-readable medium containing program code for implementing the method according to item 23 using processor technology.

Images