One of the methods of reducing the data flow in packet transmission of speech is to exclude those speech packets that do not carry enough information. This leads to the use of the systems VoIP (Voice over IP) or voice transmission technologies by IP (Internet Protocol - IP) a new type of voice signal encoders - with a variable encoding speed. The functions of detecting packets with a small voice load are carried by nodes with the standard name Voice Activity Detector (VAD). The main difficulty in the synthesis of VAD for VoIP speech signal encoders is the correct detection of speech pauses against a background of sufficiently intense acoustic noise (office, street, car noise, etc.). However, the use of VAD makes it possible to significantly save the bandwidth of a distributed network.
moment T2 about its end. On the receiving side, during the pause, to improve the subjective perception
of speech, a comfortable noise can be submit into the phone [
The algorithm of such a VAD will be presented as follows. For each speech fragment, the following is calculated (1):
k m =∑ ( xi < xtv )
i=1 where k – the total number of readings in the fragment; xtv – threshold value (1)
The limit number of readings in a speech fragment is set – mtv < k , less than which is considered that the fragment does not contain speech load. If m < mtv , then a decision is made to discard (reset) this speech fragment.
3. There is a VAD method based on the power level calculated in each speech segment Pfr
[
Pmax σ x2
If the value is less than the specified threshold q < qtv , then, as in the previous case, this will determine the need to discard the packet.
1. The GSM standard adopts the VAD scheme with processing in the frequency domain
[
This operation is carried out by an inverse filter, the coefficients of which are set in relation to the action at the input of the circuit for predicting only background noise. If there is a speech signal and noise at the input, the inverse filter suppresses the noise components and reduces its intensity. The energy of the signal + noise at the output of the inverse filter is compared with the threshold. This threshold is higher than the energy level of the noise signal only. Exceeding the threshold is considered for presence of a signal. The coefficients of the inverse filter and the threshold level change over time depending on the current value of the noise level when only noise is applied to the input. Since these parameters (coefficients and threshold) are used by the VAD detector for speech detection, the VAD itself cannot make a decision at this stage of the analysis, since the threshold can change [6-9].
This decision is made by the secondary VAD based on the comparison of the bypass spectra in successive time periods. If they are similar or close for a relatively long time, then it is assumed that there is noise at the input of the detector, then the filter coefficients and the noise threshold can be changed, that is, adapted to the current level and spectral characteristics of the input noise.
VAD with processing in the spectral domain is successfully combined with the speech RPE/LTPLPC codec, since the envelope of the input action spectrum, necessary for the operation of the secondary VAD, is already determined in the process of LPC analysis. An obvious disadvantage of this scheme is the "relatively long period of time" during which a decision is made about the presence of noise or a signal.
The VAD detector must be sensitive and fast to avoid loss of word onsets when transitioning from a pause to an active speech fragment, at the same time the VAD detector must not be triggered by background noise.
The work of the VAD is to estimate the value of the parameter of the input signal (for example, level, power or energy) and if it exceeds a certain threshold, activate the transmission of the packet. This slightly increases the delay when processing the speech signal in the codec, but it can be minimized by using encoders that work with blocks of counts, for example, based on transformations.
In the coder analyzer with the language coding rate C (bit/c), the signal is segmented into separate fragments (as a rule, quasi-stationary sections) with duration Tfr from 2 to 50 ms, and for each input block consisting of N counts, is created accordingly an information frame of size (3): . ( ) = . ·С , (3)
Regardless of the details of the implementation, the main determinant in the evaluation of the encoder is the high quality of speech reproduction at a low speed of the output digital stream Сk with minimal requirements for the resources of the digital processing processor and minimal delay. 3.1. Determination of the main parameters of the telephone signal for the operation of the VAD system
A telephone conversation between two consumers is usually accompanied by a large number of emotional pauses, and taking into account the delays experienced by speech packets in VoIР networks (more than 50 ms), pauses are added to natural speech to wait for an answer.
F (x) = eax / 2, x ≥ 0; 1 − eax , x < 0, of the signal x(t) (5). where F(x) – the probability of instantaneous signal values; – the effective (root mean square) value = Р ⋅ 1 ℎ = 1 Т
Т ∫ 0 2( )
In this case, the value of the signal peak factor is equal to (7): where Ptf – the signal power averaged over the observation time = .
For a telephone signal (TF), it is possible to accept the exponential law of the distribution of voltage instantaneous values of speech signals with the parameter = 3√2/ then, for the value of the limit , the probability F of exceeding which is < 10−3, it is possible to write (6): = 2 where ≈ 5 .
In telephone networks, for the transmission of speech signals with high quality (with satisfactory
naturalness and intelligibility of syllables – 90% and phrases – 99%), you can limit yourself to the
frequency band of 0.3...3.4 kHz [
For this study, it can be assumed that the distribution of instantaneous values of speech signals + 10 20 2 = −12.7 + 0.115 ∙ 4.32 = −10.57
– the average power of the TF signal in the voice frequency Relation (8): = 10 .
= 10 = 20 ≈ 14 (4) (5) (6) (7) (8) (9) (10) (11) (12) is the dynamic level (volume) of the TF signal. In this expression is the power of the measured signal at the point of the tract where research is conducted. According to the ITU-T recommendations, volumes are measured with a special device (volume meter), which provides a quadratic law of summation of oscillations of different frequencies, has a logarithmic scale (in decibels) and a time constant (integration time) of = 200 volumes obeys the Gaussian law with an average value
. Statistical studies have established that the distribution of . = −12.7
and the mean square deviation = 4.3 . ( ) = 1 2 − − . 2 2 2 where ( )
– the volume distribution density; – the root mean square deviation.
The corresponding average power level in the voice frequency circuit can be found using the formula (10): . circuit without taking pauses into account.
The influence of pauses is taken into account using the activity coefficient Kа of the signal source. It is equal to the ratio of the time during which the signal level at its output exceeds the set threshold value (usually –40 dBmW) to the total conversation time. Statistics of the TF signal gives Kа ≥ 0.25. Then the average power of the TF signal, including pauses (11): . . ≈ .
+ 10, . . = 32 (−15 ) corresponding to the maximum power .
and the limit voltage here the second term of the right-hand side, equal to 10 µW, is introduced according to the ITU-T recommendations, as a correction for the increased power of the signals accompanying the TF conversation (staff conversations and service signals transmitted on the same channel).
Taking into account expression (6), it is possible to determine the maximum level of . . = . + − 10.57 + 14 = 3.43 (12): .
will be equal to 2220 μW.
For signals transmitted via digital channels, . = +3 is usually accepted, and for signals transmitted using analog transmission systems +3 . In the latter case, the maximum
The dynamic range of the TF signal, calculated as the ratio of the maximum power to the average power of the permissible fluctuation interference, is equal to (15): . ≥ 10 lg 88 ∙ 10−6 178000 ∙ 10−12
≈ 27 which slightly differs from the value found by formula (10). coefficient of the signal source. Then for the speech signal (16):
When evaluating the potential information volume, it is necessary to take into account the activity Here, = 3.4 . = 2 1 + 100.1 .
= 7.6 is the upper frequency of the tone frequency channel. power of .
, which corresponds to the minimum signal, will be lower than . by the value of the peak factor. Thus, the dynamic range of the . signal, taking into account formulas . = . − .
= 2 + 3.09 + 0.115 2 = 40
It was experimentally established [
the condition TRC> the tone frequency channel
. ≈ 27 a delay (for five packets) to start transmitting a sign of comfort noise.
In the VAD structure (Fig. 2), the noise level at the output of the encoder is equalized and its threshold is set, adapting the encoder in relation to the value of the input speech signal. Analogous to this solution, noise equalization can be replaced by maintaining the input signal level within certain limits, while simultaneously fixing the noise level. That is, such a scheme becomes adaptive to the signal level. The implementation of such a principle is simple, both in analog and digital form. Its structure corresponds to the circuit with automatic gain control (ARP), which means it has the ability to adapt almost instantly to the level of the input signal.
The very implementation of the proposed principle of VAD operation is faster and simpler than in the GSM system. The integrator present in the circuit must have a time constant τ = 0.5...1s = TRC and must be fulfilled. The noise level will be maintained at the protection level of
. At the same time, it is not necessary, as in GSM, to make
Considering the properties of the speech signal, the need to use VAD is obvious when creating and
researching models of speech signal encoders [
VAD modeling by signal level 1. Beforehand a calculation is made for 2. the speech fragment, that can be processed (by default is established. (expressed in the number of levels of the maximum signal count in all segments of
= 1); = 1, 2, 4, 8, 16, 32) in relation to = , (17) where
= 127 – the maximum number of levels; a comparison of all counts in the segments and determined their numbers < ∙ 0.8 readings of the segment are equal zero;
= ∑ ∈ ( < ), further, according to the set threshold for the number of readings that can be discarded in the segment (for example, 80%), which is smaller than the percentage, , (K=160 counts are selected in the research), then it is decided that all 3. after processing one segment, we move to the next (up to 2) until we check the entire file. if
(17)
The algorithm model implements the following sequence of operations: 1. estimation of the maximum signal power in the speech fragment that can be processed by the VAD detector; threshold – < 2. setting the threshold = 0.2, 0.3, 0.5, 0.75, 1, 2 in % of ; 3. performing a comparison of the signal powers in each segment of with the selected and if the condition is met, then this packet is discarded; 4. the operations of points 2, 3 are repeated for all segments of the speech fragment.
The speech segments processed according to the proposed VAD algorithms were transfer to the coder that uses the wavelet transformation, after which the reverse decoding and assembly of the segments was carried out. To assess the quality of speech signal reproduction, the objective assessment proposed in [19, 20] was used by comparing input and output speech readings and determining the root mean square deviation (SKD) between them. It is also possible to listen to the source file, that is, to subjectively evaluate the received speech signal by ear.
The results of modeling and research are presented in fig. 4-8. Shown: fig. 4 – signal form; Fig. 5 – (SKD) of the output signal; Fig. 6 – compression coefficient Kst; Fig. 7 – relevant parameters for VAD by power; Fig. 8 – comparison of VAD methods by level and power. (MSE) values were noted and Kst values were achieved with reasonably good legibility.
The proposed VAD operation algorithm is highly efficient and provides a sufficiently high compression ratio. When evaluating the quality of VAD work according to the SKD criterion and when listening to processed speech samples, it is possible to indicate the maximum value of rejected quantization levels and the value of the adaptive power threshold at which the number of zeroed segments practically does not affect the quality of the restored speech signal. Therefore, during the conducted research, slight deviations in the quality of speech signal reproduction are observed when discarding up to 8 quantization levels out of 128, or when a power threshold of 0.3% of the maximum signal level in the segment is set. Accordingly, the achieved compression ratio is 19.35 and 15.6 (bit rate 3.3 and 4.1 kbps).
The authors are appreciative of colleagues for their support and appropriate suggestions, which allowed to improve the materials of the article.
Conference on Electronics and Information Technologies (ELIT), 2021, pp. 95-98, doi: 10.1109/ELIT53502.2021.9501064. [6] J. Song et al., "Research on Digital Hearing Aid Speech Enhancement Algorithm," 2018 37th
Chinese Control Conference (CCC), 2018, pp. 4316-4320, doi: 10.23919/ChiCC.2018.8482732. [7] M. Shovheniuk, B. Kovalskiy, M. Semeniv, V. Semeniv, N. Zanko. Information Technology of Digital Images Processing with Saving of Material Resources. Proceedings of the 15th International Conference on ICT in Education, Research and Industrial Applications. Integration, Harmonization and Knowledge Transfer. Volume I: Main Conference. Kherson, Ukraine, June 1215, 2019. pp. 414-419. http://ceur-ws.org/Vol-2387/20190414.pdf. [8] O. Tymchenko, B. Havrysh, O. Khamula, B. Kovalskyi, S. Vasiuta and I. Lyakh, "Methods of Converting Weight Sequences in Digital Subtraction Filtration," 2019 IEEE 14th International Conference on Computer Sciences and Information Technologies (CSIT), 2019, pp. 32-36, doi: 10.1109/STC-CSIT.2019.8929750. [9] S. A. Wibowo and K. Usman, "Voice activity detection G729B improvement technique using KNearest Neighbor method," 2010 International Conference on Distributed Frameworks for Multimedia Applications, 2010, pp. 1-5. [10] W. Y. Yuan, Y. Zhou, Z. Y. Huang and H. Q. Liu, "A VAD-based switch fast LMS/Newton algorithm for acoustic echo cancellation," 2015 IEEE International Conference on Digital Signal Processing (DSP), 2015, pp. 967-970, doi: 10.1109/ICDSP.2015.7252021. [11] L. K. Hamaidi, M. Muma and A. M. Zoubir, "Multi-speaker voice activity detection by an improved multiplicative non-negative independent component analysis with sparseness constraints," 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 4611-4615, doi: 10.1109/ICASSP.2017.7953030. [12] M. Shahid, C. Beyan and V. Murino, "Voice Activity Detection by Upper Body Motion Analysis and Unsupervised Domain Adaptation," 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), 2019, pp. 1260-1269, doi: 10.1109/ICCVW.2019.00159. [13] B. Durnyak, O. Tymchenko, O. Tymchenko and B. Havrysh, "Applying the Neuronetchic Methodology to Text Images for Their Recognition," 2018 IEEE Second International Conference on Data Stream Mining & Processing (DSMP), 2018, pp. 584-589, doi: 10.1109/DSMP.2018.8478482. [14] S. Hizlisoy and Z. Tufekci, "Noise robust speech recognition using parallel model compensation and voice activity detection methods," 2016 5th International Conference on Electronic Devices, Systems and Applications (ICEDSA), 2016, pp. 1-4, doi: 10.1109/ICEDSA.2016.7818517. [15] F. Patrona, A. Iosifidis, A. Tefas, N. Nikolaidis and I. Pitas, "Visual Voice Activity Detection in the Wild," in IEEE Transactions on Multimedia, vol. 18, no. 6, pp. 967-977, June 2016, doi: 10.1109/TMM.2016.2535357. [16] M. Wake, M. Togami, K. Yoshii and T. Kawahara, "Integration of Semi-Blind Speech Source Separation and Voice Activity Detection for Flexible Spoken Dialogue," 2020 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2020, pp. 775-780. [17] M. Yang et al., "Nanowatt Acoustic Inference Sensing Exploiting Nonlinear Analog Feature Extraction," in IEEE Journal of Solid-State Circuits, vol. 56, no. 10, pp. 3123-3133, Oct. 2021, doi: 10.1109/JSSC.2021.3076344. [18] Y. Hou, Y. Deng, B. Zhu, Z. Ma and D. Botteldooren, "Rule-Embedded Network for Audio-Visual Voice Activity Detection in Live Musical Video Streams," ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 41654169, doi: 10.1109/ICASSP39728.2021.9413418. [19] A. Hassani, A. Bertrand and M. Moonen, "Real-time distributed speech enhancement with two collaborating microphone arrays," 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017, pp. 6586-6587, doi: 10.1109/ICASSP.2017.8005295. [20] O. Tymchenko, B. Havrysh, O. O. Tymchenko, O. Khamula, B. Kovalskyi and K. Havrysh, "Person Voice Recognition Methods," 2020 IEEE Third International Conference on Data Stream Mining & Processing (DSMP), 2020, pp. 287-290, doi: 10.1109/DSMP47368.2020.9204023.