RU2312405C2 - Method for realizing machine estimation of quality of sound signals - Google Patents

Abstract

FIELD: analysis of sound signal quality, possible use for estimating quality of speech transferred through radio communication channels.

SUBSTANCE: in accordance to the method for machine estimation of sound signal quality, the signal is divided onto critical bands and spectral energy values are computed for critical bands, values of spectral likeness of active phase of fragments are determined, and quality of tested sound signal is determined by means of weighted linear combination of aforementioned quality values for each phase. The difference of the method is that selected fragments of active and inactive phase of both signals are synchronized, inactive phase spectrums are determined for each fragment, resulting spectrums of active and inactive phase of fragments are divided onto additional sets of bands, for each one of which spectral energy values are computed, resulting spectral energies of active and inactive fragment phases are compared in couples, to determine spectral likeness coefficients, resulting likeness coefficient for each phase is determined as an average value of likeness coefficients for all sets of bands, which is the estimate of quality of each phase.

EFFECT: ensured universality and optimized quality of estimation process depending on purposes of estimation.

5 cl, 13 dwg, 6 tbl

Description

The invention relates to the analysis of the quality of sound signals and can be used to assess the quality of speech transmitted over radio channels, telephony and communication channels, as well as to assess the quality of sound reproduced by various audio equipment, including those that underwent any compression / restoration procedures using various vocoders and assessment of the acoustic quality of the premises.

Evaluation of the quality of sound signals is becoming increasingly important with the growth in the spread and use of mobile communications, synthetic telephony systems, various portable sound recording and reproducing devices. The desire to create a method that ensures the objectivity of the assessment (i.e., independence from the assessment of a particular person) and the possibility of its automatic implementation, it is clear - an objective assessment is necessary both for comparing product samples of competitors, and for optimizing your own parameters.

One of the main indicators of compression, transmission and playback of audio information is the quality of the restored, received or reproduced sound.

The quantitative measurement of sound quality has its own specific features associated with the fact that, in the end, the receiver of the sound signal is always a person, and he is the source of most sound signals. Accordingly, the quality of sound signals is determined not only by the technical characteristics of sound processing and transmission systems, but also by the properties of the speech apparatus and hearing of people, which change over time and from person to person.

Distinguish between subjective and objective methods of measuring speech quality. Subjective methods are methods in which a person’s hearing is an integral part of a measuring complex. Accordingly, objective methods exclude the participation of human hearing from the measurement process.

The most common subjective method for assessing the quality of speech (not necessarily speech, although usually it is speech) is the MOS score (mean opinion score) - an assessment on a five-point scale.

The MOS score is determined by processing the ratings given by the groups of auditors to several audio signals played by various audio systems. Each auditor evaluates each signal. Then the results are averaged.

The process of organizing and conducting subjective examinations is a rather complicated, lengthy and costly procedure, therefore, for many years, work has been ongoing on the search for objective methods for assessing intelligibility, which allow us to obtain quick and automated assessments that are in good agreement with subjective examinations.

Various valuation methods are known, some of which are given below:

AI (Articulation Index) - The idea is that the entire frequency range of the speech signal is divided into 20 bands, within which the signal-to-noise ratio is determined. The bandwidth is chosen so that the contribution of each band to speech perception is the same. In each band, the signal-to-noise ratio is calculated. The articulation index is taken equal to the weighted sum of the values in the bands.

The articulation index is bad in that it, although it is focused on a speech signal, does not take into account the properties of hearing and speech formation.

SII (Speech Intelligibility Index) - Speech intelligibility index - development of the AI method. The speech intelligibility index is included in the American standard ANSI S3.5-1997 and offers four measurement procedures for different groups of bands: critical bands (21 bands), one-third octave bands (18 bands), equal critical bands (17 bands) and octave bands ( 6 bands). In each of the bands, the signal-to-noise ratio is calculated and the total coefficient SII is calculated, lying in the range from 0 to 1.

The speech intelligibility index takes into account only the properties of hearing and does not take into account the properties of speech formation.

STI (Speech Transmission Index) - speech transmission index - A speech signal can be approximately considered as a broadband signal modulated by a low-frequency signal. The modulation frequency is determined by the rate of articulation. Reducing the depth of modulation likens a speech signal to a noise signal and reduces its intelligibility. Accordingly, a decrease in intelligibility can be estimated by a decrease in the modulation depth.

The entire speech range is divided into seven octave bands, an octave noise signal is fed to the input of the test system. The distribution of the intensity of the test signal coincides with the distribution of intensities of the speech signal. The frequencies of the modulating signal vary from 0.5 to 12.5 Hz with a third-octave interval (a total of 14 frequencies).

The STI measurement method is recorded in the international standard IEC 268-16.

RATSI / STIPA (Rapid Speech Transmission Index) is a fast speech transmission index. The STI method requires a large number of measurements and calculations. A simplified method was developed, providing for measurements in only two bands at five modulation frequencies, and reduced the number of measurements and calculations. For good readability, RASTI values should be at least 0.6.

The speech transmission index, as well as the fast index, imitates the process of speech formation using a noise model, however, such an account of the properties of speech formation and hearing is far from optimal.

C50 - clarity coefficient - determines the clarity or clarity of sound and is calculated as the ratio of near and far echo. The method is based on the fact that the echo reduces the intelligibility of the signal. The ratio of near and far echo is measured at several frequency bands. The near echo (up to 33 ms) is considered a useful signal, and the distant echo (more than 33 ms) is considered an interfering signal.

The clarity factor takes into account only one type of possible distortion and it is advisable to use it as one of the assessments of speech quality.

There is a method of evaluating speech intelligibility obtained through the paths of intercoms of personal respiratory protection by applying a voice message transducer to an electric signal and a complex of recording and processing equipment to obtain the amplitude-frequency dependence of the voice message, determining formants of equal intelligibility, their sensation level, calculation the probability of receiving formants, the magnitude of which assesses speech intelligibility, characterized in that the voice message converter in the ele a ctric signal is connected to the input of the sound adapter of a personal computer with a digitization board, information is transferred from an analog form to digital, digital information is processed and the output characteristics required for evaluating intelligibility are evaluated (application No. 2002133196).

The disadvantage of this method is that it does not fully take into account the properties of speech formation. The presence of formants is characteristic only for vowels and voiced consonants. In addition, this method is applicable only to assess speech intelligibility as a measure of the quality of a speech audio signal, but it is not applicable to audio signals in general.

The closest technical solution to the claimed one is a method for performing a machine estimation of the transmission quality of audio signals, in particular speech signals, in which the spectra of the transmitted signal of the source and the received signal are determined in one frequency range, the value of spectral similarity that corresponds to the quality of transmission is determined, while the covariance of the signal spectra the source and the received signal are divided by the product of the standard deviation of both spectra (RF Patent No. 2232434).

In addition, the spectral similarity values are weighted by a coefficient that depends on the ratio of the energies of the spectra of the receive signal to the source signal, which provides regulation of the interference signal, since the higher the energy of the received signal, the greater the similarity value decreases.

Prior to signal processing, the active and inactive phases are isolated from the source signal and the received signal, while the signal fragments whose energy exceeds a predetermined threshold are correlated with the active phases, and the remaining fragments are qualified as pauses. Pauses and interferences in pauses are also separated and accounted for to a lesser extent than the active phases of the signals.

Based on this, the value of spectral similarity is determined only for fragments of the received signal and the source signal related to the active phase, and for inactive phases, a quality function is applied, depending on the maximum and average energy in the interval of pauses, which decreases degressively.

Before conversion to the frequency domain of the active phase signals, temporary masking is performed, for which they are divided into temporary data blocks in such a way that successive data blocks are overlapped by a substantial part up to 50%, and before temporary masking, the components of the spectra are compressed by raising to the power c an indicator less than 1.

The obtained spectra of the source and the received signal are divided into critical bands (according to the Zwicker model) and calculates similarity coefficients for them. Before determining the similarity value, the spectra are respectively convolved using a frequency-asymmetric blur function, and before the convolution, the components of the spectra are expanded using exponentiation with an exponent greater than 1.

The transmission quality is calculated by a weighted linear combination of the similarity value of the active phase and the quality value of the inactive phase.

The main disadvantages of the prototype include:

- in practice, only the active phases of the initial and received (tested) signals are processed, which reduces the objectivity of the assessment;

- this method does not take into account the properties of speech formation, because critical Zwicker bands used by the inventors reflect only hearing properties;

- the method takes into account the perception of the inactive phase only in terms of volume, which also reduces the accuracy of the assessment.

The objective of the invention is to develop a method for obtaining an objective assessment of the quality of the audio signal, which can be used in these areas of application of the invention.

The technical result is achieved due to the fact that in the known method of machine estimation of the quality of sound signals, in which fragments of the active and inactive phases are extracted from the initial signal and the test signal, the spectrum of the active phase is determined, the spectral energy values at critical bands and similarity values are calculated, and the quality the tested sound signal is determined by means of a weighted linear combination of the obtained values for each phase, changes are made, namely:

- the selected fragments of the active and inactive phases are synchronized in time;

- additionally determine the spectra of fragments of the inactive phase;

- the obtained spectra of fragments of both phases are divided into additional sets of bands for which the values of spectral energy are calculated;

- fragments are compared;

- the resulting similarity coefficient for each phase is determined as the average value of the similarity coefficients of the sets of bands for all fragments.

Then, taking into account the results obtained, the quality of the tested sound signal is evaluated.

Besides:

- as an initial signal, you can use either an arbitrary audio signal or a specialized set of signals;

- the spectra of fragments of the active and inactive phases are determined using discrete cosine transform;

- as additional sets of bands can be used logarithmic, resonator and various known critical bands;

- the number and composition of sets of bands can vary in various combinations to determine the similarity coefficient of each phase.

The essence of the invention is illustrated using figures 1-3, figures 4-8 explain an example implementation, and figures 9-13 - possible ways of using:

Figure 1 - enlarged algorithm for assessing the quality of the audio signal;

Figure 2 - algorithm for comparing signal fragments in bands;

Figure 3 - General algorithm for synchronizing the source and test signals;

4 is a VAD emission filtering algorithm;

Figure 5 - algorithm of the synchronization unit (start);

6 - algorithm of the synchronization unit (continued);

7 - the algorithm of the synchronization unit (continued);

Fig - the algorithm of the synchronization unit (end);

Fig.9 is an example of assessing the quality of sound transmitted through a telephone network;

Figure 10 is an example of evaluating the quality of sound transmission over VoIP;

11 is an example of assessing the quality of sound transmission in cellular and satellite networks;

Fig - an example of the use of quality ratings by a group of developers of sound processing systems (s);

Fig - an example of assessing the sound quality of the premises.

The need to develop new methods and improve existing ones is caused by the desire to increase the proximity of objective and subjective quality assessments, the need to take into account the properties of hearing and speech formation.

The use of an arbitrary or specialized signal as the initial signal depends on the purpose of the assessment (determining speech intelligibility, sound reproduction quality, evaluating the quality of speech received through intercom channels, etc.) and makes it possible to increase its objectivity.

Almost any sound signal can be divided into active and inactive phases. The first corresponds to active sound processes, the second to low-level background noise. The simplest method of separation is separation by signal energy level, but this approach does not have high accuracy. In the proposed method, the well-known VAD algorithm, fixed in G.723 recommendation (as an element of the vocoder of the same name), is used to separate the signal into active and inactive phases.

The source and test audio signals are analyzed and divided into active and inactive phases (figure 1). Then the fragments of the active and inactive phases are synchronized (fragments of the same type are combined in time) and analyzed by different blocks according to one algorithm. The synchronization algorithm is described below.

Separate comparison of combined pairs of fragments of the active and inactive phases allows one to increase the accuracy of the resulting estimate.

For each fragment, the integrated spectrum is determined using discrete cosine transform (DCT), which has some advantage over the fast Fourier transform (FFT) to achieve a technical result.

The integration of the spectrum is carried out according to the formula (1):

where j = 0 ... N / 2-1 are the indices of the spectral energy,

i is the number of the integration step;

N is the number of signal samples used in the calculation of the spectrum;

- получаемое усредненное значение спектра;

- the resulting average value of the spectrum;

- усредненное значение спектра на прошлом шаге;

- the average value of the spectrum at the last step;

Sp _{i, j} is the spectrum value obtained using DCT.

When calculating the integral spectrum, the window overlap is N / 2 samples; the well-known window function of Hamming or Blackman-Harris is superimposed on each window.

For all selected sets of bands, the spectral energy levels in the bands are determined. There are known groups of critical bands defined by different authors, based on various models of sound perception and speech formation.

The human hearing system is a non-linear system, which leads to the appearance of a phenomenon called masking. Masking occurs when listening to a message against a background of clutter, or masking sounds.

As a result of the study of masking harmonic signals with narrow-band noise, Zwicker determined that the entire spectrum of audible frequencies can be divided into frequency groups or bands emitted by the human ear. Before Zwicker, a similar conclusion was made by Fletcher, who called the selected frequency groups critical hearing bands.

The critical bands defined by Fletcher and Zwicker differ because the first determined the bands using noise masking, and the second from the ratios of perceived loudness.

Sapozhkov defined the critical band as “a strip of the frequency range of speech, which is perceived as a whole.” In his early studies, he even talked about the possibility of replacing the audio signal on the strip with an equivalent tone signal, but this assumption did not stand up to experimental verification. The critical bands defined by Sapozhkov differ from the bands defined by Fletcher and Zwicker, because Sapozhkov proceeded from the properties of a speech signal.

Pokrovsky also determined critical bands based on the properties of the speech signal. The bands defined by Pokrovsky provide an equal probability of hit by formants.

The value of the spectral energy in the bands can be used for various purposes, one of which is the assessment of the quality of the audio signal. However, the use of critical bands of only one author (in the prototype, for example, Zwicker critical bands are used), does not allow to obtain a fairly objective assessment, because reflect only one aspect of either perception or speech formation. In the present invention, the spectral energy can be determined on various critical bands, as well as on the logarithmic and resonator bands, which allows you to take into account more features of hearing and speech formation.

Taking into account the fact that the bands defined by Pokrovsky and Sapozhkov are better suited for speech signals, and not for sound signals in general, can improve the accuracy of the assessment, depending on its purpose. Table 1 shows the critical bands for various authors.

The following notation is used:

Fc is the center frequency of the band;

L is the bandwidth.

Table 1 Critical bands identified by different authors. No. Zwicker Pokrovsky Fletcher Sapozhkov Fc L Fc L Fc L Fc L one 51 80 260 320 200 53 200 60 2 150 one hundred 495 150 300 fifty 300 60 3 250 one hundred 640 140 400 fifty 500 60 four 350 one hundred 787 155 500 fifty 800 70 5 450 110 947 165 600 53 1000 80 6 570 120 1125 190 700 54 1500 one hundred 7 700 140 1315 190 800 58 2000 130 8 840 150 1505 190 900 60 3000 200 9 1000 160 1690 180 1000 63 5000 300 10 1170 190 1870 180 1250 71 8000 600 eleven 1370 210 2050 180 1500 80 12 1600 240 2230 180 1750 87 13 1850 280 2435 230 2000 98 fourteen 2150 320 2725 350 2500 120 fifteen 2500 380 3100 400 3000 141 16 2900 450 3480 360 4000 200 17 3400 550 3855 390 5000 276 eighteen 4000 700 4530 960 6000 370 19 4800 900 6130 2240 7000 480 twenty 5800 1100 8625 2750 8000 590 21 7000 1300 22 8500 1800 23 10500 2500 24 13500 3500

Additionally, it is proposed to use logarithmic bands or bands of equal volume. The idea is simple - loudness is proportional to 10 logarithms of energy. To determine the boundaries of the logarithmic bands, a phonetically representative text is used (a well-known text developed at the Department of Phonetics of St. Petersburg State University), read by announcers of different sex and age.

The voice path is a complex speaker system. The acoustics of the vocal tract are non-stationary and non-linear. With the movement of articulatory organs, the shape and volume of the upper resonator change, resulting in a speech function. The height of the voice is determined by the number of vibrations of the vocal cords per second, as well as the length of the ligaments, the strength of their tension and the position of the epiglottis. The strength of sound is determined by the closing force of the vocal cords and the force of exhalation. The timbre changes depending on the position of the larynx and epiglottis.

Due to the anatomical features of the structure of the speech apparatus and the ability to use resonators, some people get amplification or weakening of the harmonic components of sounds. The main effect on phonation is exerted by the upper resonator and pharynx. Also, the resonator function, which consists in enhancing the tones of the voice and giving it an individual timbre, is carried out by the nasal cavity and paranasal sinuses.

Resonator bands characteristic of various speech sounds were identified by V.N. Sorokin (table 2). Consideration of resonator bands is useful in determining the quality of speech audio (especially speech) signals. Resonator bands can be used to determine the playback quality of individual sounds.

The indices at the center frequencies and bandwidths are given according to Sorokin. F _x corresponds to Fc, a L _x -L.

table 2 Resonator bands No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ one "BUT" 273.5 72,4 574.6 78.1 994.1 48.3 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2404.8 77.7 2711.4 102.5 3796.5 145.6 4735.3 221.8 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 2 "ABOUT" 287.6 72,4 497.1 100.9 914.2 47.1 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2316.4 67.9 2635.1 87.6 4030.9 142.3 4728.3 189.5 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 3 "U" 296.8 72,4 408.6 149.2 858.0 41.9 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2042.8 54,2 2761.3 71.2 3612.3 92.4 4434.3 122.7 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ four "AND" 287.7 72,4 393.5 54.9 2272.1 66.1 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 3094.6 77.6 4003.6 83.7 5047,3 117.0 6103.5 133.6 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 5 "Y" 302.6 72,4 485.7 85.5 1378.4 47.0 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 1847.7 46.3 2574.5 63.3 3732.5 97.7 4421.9 124.8 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 6 "E" 279.0 72,4 490.9 73.1 1353.0 41,4 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2235.0 60.8 2775.0 78.5 3575.7 109,4 4226.4 141.3 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 7 "FROM" 325,4 72,4 482.7 72.7 1619.4 45.7 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2861.0 72.7 4029.8 106.3 4,406.1 115.9 5290.6 153.9 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 8 "Sh" 335.1 72,4 473.4 97.5 1439.9 53.7 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2101.6 57.1 2528.8 62.8 3159.8 72.9 4516.78 117.3 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 9 "X" 349.9 72,4 543.8 91.9 1459.7 54.8 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2035.0 53.5 2915.1 78.5 3699.1 93.5 4540.6 120.5 No. Sound F _p L _p F ₁ L ₁ F ₂ L ₂ 10 "F" 274.9 72,4 338.9 83,2 1,024.6 37,4 F ₃ L ₃ F ₄ L ₄ F ₅ L ₅ F ₆ L ₆ 2110.2 43,2 2694.5 53.5 3872.9 78.0 4798,0 104.9

Additionally, the “importance factors” of the bands can be determined based on the assumption that the lower the integrated energy in the band, the higher the importance of the band for speech perception. Accordingly, to assess the quality of sound signals in general, it is advisable to consider the bands as equally important, and when evaluating the quality of speech signals transmitted along the paths of intercoms, take into account the importance factors.

The borders of the bands (starting and ending indices) are determined by the following formulas:

where nSpecLen is the number of points in the spectrum (N / 2);

SampleRate - signal sampling rate;

n is the number of the strip.

The energy in the bands is defined as:

- значения интегрального спектра (

равно

, полученному на последнем окне фрагмента).Where

are the values of the integrated spectrum (

equally

obtained on the last window of the fragment).

The strip comparison algorithm (for one set) is shown in FIG. 2. The initial quality score is assumed to be 100%. Further, it decreases in proportion to the difference in energy in the bands. Quality assessments are determined for each set of bands. Quality assessment for all sets of bands is defined as the average value of individual ratings by the formula:

where Nk is the number of used strip tables;

k is the number of the current table;

dQ _k is the estimate obtained for the k-th strip table;

- интегральная оценка по всем таблицам.

- integral score for all tables.

Quality assessment for each phase is determined as the average for all pairs of fragments:

- получаемое интегральное значение коэффициента потери качества;Where

- the resulting integral value of the quality loss coefficient;

- интегральное значение коэффициента качества на предыдущем шаге;

- the integral value of the quality factor in the previous step;

- значение коэффициента качества на паре фрагментов с номером t;

- the value of the quality factor on a pair of fragments with number t;

- значение коэффициента качества на первой паре фрагментов;

- the value of the quality factor on the first pair of fragments;

t is the number of a pair of fragments.

(Active)) и неактивной (

(Pause)) фаз:The resulting quality score for the entire signal (dQGlobal) is defined as the sum of the weighted quality ratings of the active (

(Active)) and inactive (

(Pause)) phases:

The general signal synchronization algorithm is shown in FIG. 3. Signal segments (pDATA) equal in duration to the VAD frame and signs of VAD activity on pDATA segments are input to the signal synchronizer. There are two inputs: for the reference (or source) signal and for the test signal.

Before synchronization, a filtering of emissions of signs of VAD activity is carried out, which consists in the fact that the sign of activity in short sections (with a duration less than a threshold) is equated with signs of activity of the surrounding signal.

After the filter, the state signs and signal frames are sent to synchronization blocks combining fragments of the active signal and pauses. The modules use common data: buffer of the active reference signal (EBuffer1), buffer of the active test signal (TBuffer1), pause buffer of the reference signal (EBuffer0), buffer of pause of the test signal (TBuffer0), sign of readiness of the active signal buffers and pauses (dReady [0 .. .1]), a synchronization error counter (dErrorCounter) is also provided.

At the output of the synchronizer, a pair of buffers with an active signal or a pair of buffers with pauses is obtained. Both synchronization blocks can trigger the appearance of a pair of synchronized buffers.

Synchronized buffers, depending on the sign of activity, are sent to the unit for comparing active fragments or pauses (Fig. 1).

At present, testing of the proposed method continues with respect to assessing the quality of telephone channels and IP-telephony. The search for optimal synchronization algorithms is carried out and the relationship between quality assessment and syllabic intelligibility is specified.

The following is a description of the implementation of the method. Implementation of the proposed method for evaluating the quality of sound signals is carried out on a personal computer using software developed by the inventors. The method is implemented as a program for assessing the quality of vocoders and comparing external source and test signals.

As external signals, arbitrary signals recorded with a sampling frequency of 8 kHz and a sampling capacity of 16 bits can be used. It is assumed that the test signal is obtained from the original signal as a result of any transformations (for example, compression / restoration, transmission via communication channels, filtering).

Additionally, a phonetically representative text read by several speakers of different sexes and ages can be used as the source of the external signal.

As internal source signals (signals to which the user of the program does not have access) are used signals generated in accordance with the noise model (description of the generator is given below) and signals generated based on the statistical model.

Internal signals are fed to the input of the implementation of the compression / restoration system of audio data, implemented as a DLL with a specified interface. It is allowed to use DLLs developed both by the authors of the proposed method and by third-party developers. The signal processed by the methods contained in the DLL is considered to be tested and undergoes the quality assessment procedure described above.

4 shows a VAD emission filtering algorithm. PDATA signal segments and signs of VAD-dVAD activity are used as initial data. Table 3 shows the names of the variables, their purpose and initial values. In addition to the variables in the algorithm, three constants were used: the threshold for straightening pauses to the active state (dBound [0] = 6), the threshold for straightening pauses in the active state (dBound [1] = 4) and the length of the delay line (dDLSize = max (dBound []) +1).

The used values of the constants are determined experimentally (for the case of evaluating the quality of signals that have passed the compression / restoration procedure) and can be changed during implementation for better synchronization of specific signals.

Table 3 Variables Used by VAD Emission Filter Variable Appointment N / a dVAD The value of the activity indicator received at the input of the algorithm - pDATA Array of signal samples with a length equal to the VAD frame - dState Sign of site activity (previous value of activity sign) -one dSLen The number of consecutive frames with the same sign of activity 0 dNDLFrames The total number of frames received at the input of the algorithm 0 DelayLine [] Delay line. Saves activity signs and sample arrays -

The algorithm checks the sign of activity of the current signal block. If the sign of activity coincides with the current received state, then the incoming frame is simply added to the delay line, and the first element of the delay line is issued to the input of the synchronizer block.

If the sign of activity does not coincide with the current accepted state, then a check is made for the arrival of the first frame of the signal. The first frame is simply placed in the delay line, and its sign of activity is taken as the current state.

If there is a change in the activity of the received signal during the filtering process, then the number of signal frames received in the previous state is checked. If the number of frames is less than the set threshold, then their attribute of activity is changed to the opposite, if not, then the current state simply changes and the counter of frames received in the current state is reset. After all operations to change the state, the frame is placed in the delay line.

The operation of the algorithm is completed by obtaining the sign of the end of the signal. In this case, the entire accumulated signal is given to the input of the synchronization block, if, of course, there is one, and only then is a sign of the end of the signal.

To synchronize the signals, a pair of synchronization blocks is used, working with several common variables described above. The operation algorithm of the synchronization unit is presented in figures 5-8.

- тестируемого, и наоборот - в блоке 1 XBuffer0 - буфер пауз тестируемого сигнала, a

- эталонного.Synchronization block 0 processes the reference signal (figure 5), and block 1 is the test one. The block algorithms are identical, the blocks use cross references to buffers, i.e. in block 0, XBuffer0 is the pause buffer of the reference signal, a

- tested, and vice versa - in block 1 XBuffer0 - buffer of pauses of the tested signal, a

- reference.

- тестируемого, и наоборот - в блоке 1 XBuffer1 - буфер активного тестируемого сигнала, а

- эталонного.Similarly, in block 0, XBuffer1 is the buffer of the active reference signal, a

- tested, and vice versa - in block 1 XBuffer1 - buffer of the active test signal, and

- reference.

Upon receipt of the sign of the end of the signal, the algorithm completes its work. The stop branch is shown in Fig. 8.

Depending on the sign of VAD activity, the signal is placed either in the pause buffer or in the active signal buffer. If the buffer size exceeds the threshold value, then synchronized buffers are issued to the comparison module. Branches issuing synchronization by the size of the buffer are presented in Fig.7.

After placing the signal in the buffer, the current state of signal activity is checked. If it is the same, then a transition to the beginning is made and new data is expected. When the state changes, it is checked if this was the first piece of data? If “yes”, then its state of activity is accepted and the transition to the beginning is carried out.

If not, then the sign of signal availability in this state increases, after which it is checked whether both signals are ready, i.e. sections of the active signal or pause are synchronized. If there are synchronized signal fragments, go to the branch shown in Fig.6. If not, then go to the beginning of the algorithm.

The current state determines whether synchronization was found for pauses or for the active signal. We check the synchronization result for an error by comparing with zero the size of the buffers (ours and the buffer from the parallel block) of the signal. If at least one of them is equal to zero, then a synchronization error has occurred.

If everything is in order, the synchronized buffers are issued to the input of the comparison module. If not, then the error counter is increased, the buffers are discarded, the state of activity changes and there is a return to waiting for a new piece of data.

Before you give the buffer to exceed the size of the segment, we check the size of the parallel buffer (Fig.7). If the parallel block buffer is empty, the buffers are discarded and the counter of synchronization errors increases. If data is present in both buffers, synchronized fragments are given to the signal comparison module.

Before finishing work, it is checked: is there data in the pause buffers and active signal buffers. If there is, then we give the corresponding synchronized pairs (or a pair) of signals to the comparison module. After that, we give the comparison module a sign of the end of the signal.

Next, the integrated spectra of the selected and combined fragments are calculated in accordance with the description of the method given above. To calculate the spectra, a 1024-point discrete cosine transform is used, which provides sufficient accuracy for determining the band boundaries for an 8 kHz signal.

Table 4 presents the coefficients of importance of the individual bands, determined in accordance with the description of the method. Coefficients are defined for signals recorded with a sampling frequency of 8 kHz.

Table 4 Critical Band Importance Factors No. Zwicker Pokrovsky Fletcher Sapozhkov Vc _log Vc _line Vc _log Vc _line Vc _log Vc _line Vc _log Vc _line one .112257 .022757 .023221 .000234 .060620 .000554 .119224 .002201 2 .071777 .001918 .052324 .001002 .062399 .000774 .122034 .002950 3 .063108 .000816 .059593 .002275 .056358 .001181 .150998 .008126 four .066354 .001426 .057859 .004045 .058028 .001615 .136877 .027655 5 .063906 .001986 .061305 .009510 .061681 .003002 .141754 .081240 6 .063221 .003309 .059533 .019082 .064525 .004624 .124286 .165172 7 .056019 .005001 .061430 .029982 .067569 .006987 .123389 .462036 8 .057442 .009524 .064073 .032110 .072189 .012508 .073987 .249604 9 .061323 .023545 .068123 .037674 .068211 .020201 10 .055594 .037177 .074750 .066339 .068900 .045774 eleven .052207 .048333 .082703 .121272 .062134 .041917 12 .046928 .043929 .086423 .153918 .059783 .055413 13 .043235 .066483 .067701 .114197 .063507 .122423 fourteen .043545 .132619 .042399 .063481 .049434 .112761 fifteen .037087 .111828 .035929 .044686 .041380 .072089 16 .029355 .072354 .045791 .100261 .081286 .497994 17 .026423 .087377 .055098 .199775 eighteen .049202 .329361

Also, in accordance with the description of the method, the logarithmic bands and their importance coefficients are determined, which are valid for the 8 kHz signal (Table 5).

Table 5 Logarithmic bands and their band importance factors No. Fc L Vc _line No. Fc L Vc _line No. Fc L Vc _line one 74 149 .005170 8 1035 180 .013173 fifteen 2551 242 .090052 2 207 117 .000426 9 1219 188 .022271 16 2797 250 .081422 3 324 117 .000556 10 1410 195 .029766 17 3035 227 .069182 four 445 125 .000925 eleven 1609 203 .027939 eighteen 3273 250 .079362 5 574 133 .001577 12 1816 211 .042986 19 3539 281 .142375 6 715 148 .002717 13 2047 250 .079971 twenty 3836 313 .204682 7 867 156 .005893 fourteen 2301 258 .099413

In accordance with the description of the method, the integrated spectra of the fragments, the energy in the bands, the quality estimates for each pair of fragments, and the integral, resulting quality assessment are calculated. This implementation uses all sets of bands.

Then, for convenience of comparison with subjective assessments of MOS, the objective assessment in percent is converted into points by dividing by 20.

The signal generator corresponding to the noise model of speech generation works as follows: white noise is generated. Critical bands defined by Pokrovsky or Sapozhkov are cut out of it (Table 1). Each band is modulated by the frequencies listed below. Modulation frequencies are applied sequentially on the number of samples indicated for each frequency (the number in parentheses). After all modulation frequencies have been tried, a pause of 8000 samples (1 second) is made and the transition to the next band is carried out.

The following modulation frequencies are used: 0.63 (40,000), 0.84 (40,000), 1.05 (40,000), 1.26 (40,000), 1.68 (40,000), 2.10 (20,000), 2.52 (20,000), 3.36 (20,000), 4.20 (20,000), 5.04 (20000), 6.72 (10000), 8.40 (10000), 10.08 (10000), 13.44 (10000).

The statistical model generates an audio signal based on knowledge of the sound composition of the Russian language, frequency of sounds, statistical information on the physical characteristics of sounds, statistical data on the composition of the population, voice samples of several speakers. The model generates the original sound signal as a sequence of samples of the speaker’s voices taken in random order, in an amount proportional to their frequency.

As a result of testing the proposed method, quality assessments of several standard vocoders were obtained. Table 6 shows the quality assessment of several standard vocoders obtained on various test signals, the proposed method using the described implementation. For comparison, the table shows the MOS estimates.

Table 6 Assessing the sound quality of vocoders Codec MOS Noise model Statistical model FPT Minimum Abbreviated Full - Vc - Fc - Vc - Vc - Vc A-law 4.10 4.79 4.73 4.78 4.78 4.78 4.78 4.79 4.80 4.80 4.84 Mu-law 4.10 4.79 4.84 4.77 4.77 4.77 4.78 4.78 4.79 4.79 4.82 G.723.6.3 3.90 4.25 4.48 4.21 4.29 4.22 4.33 4.15 4.04 4.08 3.95 GSM.6.10 3.70 3.20 1.99 3.01 1.65 3.04 1.78 4.22 3.66 4.01 3.21 G.723.5.3 3.65 4.23 4.44 4.18 4.27 4.19 4.32 4.14 4.04 4.06 3.93

In the column with the “-” sign, the estimates for making the bands are equally probable, and in the column “Vc”, the estimates obtained taking into account the importance factors are given.

The proposed method for evaluating audio signals has several advantages over the known methods of measuring quality, namely:

- has versatility, because allows you to judge the quality of signals of various origins that have gone through various processing procedures;

- the quality assessment process can be optimized depending on the objectives of obtaining the assessment:

- speed (for example, it is possible to quickly get a rough estimate);

- according to the type of signal (the use of different bands for speech signals and sound signals in general);

- the obtained estimate correlates well with the MOS estimates;

- quality estimates obtained for speech signals can be converted into values of different types of intelligibility.

The following is a brief description of several possible applications of the proposed method for evaluating sound quality.

Figure 9 presents a diagram of the application of the proposed method for assessing the quality of sound transmission through a public telephone network. This scheme is valid for both local and long-distance / international calls.

The sound quality assessment server generates an initial signal (or selects among pre-prepared ones) and transmits it to one of the subscribers participating in the testing.

The subscriber who received the signal establishes a regular telephone connection with the second subscriber, and reproduces the original signal. The second subscriber records the received audio signal and transmits it to the server for evaluating sound quality.

The sound quality assessment server compares the source and test signals in accordance with the proposed method and provides an estimate of the quality of sound transmitted through the telephone network. The resulting assessment can be used to improve the quality of customer service, make decisions about the need to replace or configure equipment (both on the subscriber side and on the station side), for advertising purposes, etc.

Similarly, there is an assessment of sound quality over the IP network, presented in figure 10. The difference from the previous application is in the method of transmitting the source and test audio signals from the server for evaluating sound quality to subscribers, and in the method of transmitting data between subscribers.

In addition, the obtained quality assessments can be used to select the codecs used in VoIP communications and to select the operators providing IP-telephony services.

Similarly, the proposed method can be used to evaluate cellular and satellite communications (see Fig.11). The obtained estimates can be used by subscribers to select telecom operators and phone models, and by operators to optimize the placement of base stations.

On Fig presents the process of using the proposed method for assessing sound quality by developers and testers of systems and algorithms (methods) for compressing audio data. Each version of the codec (or codec with a set of parameters) requires evaluation and comparison with analogues. Each developer can turn to the database of sound samples, compress and restore the signal and get an objective assessment of the quality of the codec.

Such a system will allow you to control the process of developing codecs and optimizing their parameters, in addition, the end user will be able to get the optimal algorithm, and not just working.

On Fig presents the process of evaluating the sound quality of the premises. In this case, the source signal is received from a microphone located opposite the speaker, and test signals from microphones located in different parts of the room, at the locations of listeners and sound-reproducing equipment.

The resulting estimates can be used to optimize the location of sound-reproducing equipment, furniture and seats.

After testing, the proposed method for 2005-06 will be widely used in various fields of technology.

Claims (5)

1. A method for the implementation of a machine-based evaluation of the quality of sound signals, in which the active and inactive phases are isolated from fragments of the initial and test signals, the spectrum of the active phase is determined, it is divided into critical bands and the spectral energy values are calculated on the critical bands, the spectral similarity values of the active phase of the fragments are determined and the quality of the tested audio signal is determined by a weighted linear combination of the obtained quality values for each phase, characterized in that These fragments of the active and inactive phase of both signals are synchronized, the spectra of the inactive phase for each of the fragments are determined, the obtained spectra of the active and inactive phase of the fragments are divided into sets of bands, including additional sets of critical, as well as logarithmic and resonator bands, for each of which the values of the spectral energy, compare the pairwise obtained spectral energies of the active and inactive phases of the fragments, to determine the coefficients of spectral similarity, the resulting coefficient the similarity coefficient for each phase is determined as the average value of similarity coefficients for all sets of bands, which is an estimate of the quality of each phase. 2. The method according to claim 1, characterized in that as an initial signal, you can use either an arbitrary audio signal or a specialized set of signals. 3. The method according to claim 1, characterized in that the spectra of fragments of the active and inactive phases are determined using a discrete cosine transform. 4. The method according to claim 1, characterized in that the spectral energy of the active and inactive phases of each fragment is calculated taking into account the importance factors of each band included in the set. 5. The method according to any one of claim 1 or 4, characterized in that as the sets of bands use a different combination of logarithmic, resonator and known critical bands.

Images