An objective evaluation method for voice quality of voice communication system

Abstract

The present invention is applicable to the field of voice quality assessment and provides an objective evaluation method for the voice quality of a voice communication system, which comprises time alignment, signal framing and time alignment, 1 / 12 th Octave filtering, level calculation, auditory correlation calculation, feature statistics, deep learning, and evaluation value are performed. The algorithm is used to align the time and the three groups of signals are 1 / 12 th After octave filtering, the level value is calculated and the auditory correlation calculation is also required. The three groups of input signals are obtained for feature statistics, and deep learning is performed on them to obtain the evaluation value of the voice quality. The evaluation value is mapped to the MOS range to obtain the final calculation result. This objective evaluation method makes up for the shortcomings of the existing voice quality evaluation algorithm on the market. It can not only guide the efficient design and development of call products in the development stage, but also can quickly and accurately detect the call quality of the product in the product acceptance stage. At the same time, it has created a precedent for my country to independently realize the voice quality evaluation algorithm.

Description

Technical Field

[0001] The invention belongs to the field of voice quality assessment, and in particular relates to an objective evaluation method for the voice quality of a voice communication system. Background Art

[0002] With the rapid development of communication technology, the functions of mobile phones are also improving day by day, and they are gradually becoming intelligent, information-based, and network-based. At the same time, various APP software emerges in an endless stream, and have been deeply integrated into people's lives and become daily necessities. For Chinese people, you can go out without a wallet, a card, or even a key, but you can't go out without a mobile phone, which shows the importance of mobile phones in people's lives. However, no matter how mobile phones develop, the call function is its most basic and fundamental need. Therefore, with the development of current communication networks, people's requirements for the quality of mobile phone calls are getting higher and higher. For example, when you are on the subway after work or on a noisy street, you hope to be able to clearly hear what the other party is saying. This puts higher requirements on mobile phone manufacturers. While meeting the expansion of mobile phone functions, how can they occupy a greater advantage and market in the most basic call performance?

[0003] The voice quality test of the call system is not a new topic, but a mature detection method. The purpose of the test is to conduct objective test, and its results can truly reflect people's auditory perception, and then guide the design or manufacture of mobile phones. A PESQ method is currently used. This method compares the test signal with the reference signal, maps the comparison result to the MOS score range, and a voice quality evaluation method under background noise environment, which uses three groups of signals, namely voice signal, reference signal and terminal acquisition signal, to participate in the calculation and give an objective evaluation MOS score. However, the above schemes each have their own shortcomings. For example, the PESQ voice quality evaluation can evaluate the voice quality under background noise environment, but its application scope is limited and it is not suitable for today's technical environment. Although the voice quality evaluation method under background noise environment clarifies the voice quality evaluation under background noise, it is also applicable to broadband and ultra-wideband communication networks, but its standard technology and related parameter definitions are too vague and general, and there are some technical barriers. It is difficult for technical personnel in this industry to achieve accurate and objective voice evaluation based on this standard. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide an objective evaluation method for voice quality of a voice communication system, aiming to solve the problems existing in the above-mentioned background technology.

[0005] The embodiment of the present invention is implemented as follows: a method for objectively evaluating the voice quality of a voice communication system comprises the following steps:

[0006] Step 0: Time alignment: using the Clean Signa l as the reference signal, the Processed Signa l and the Unprocessed Signa l signals are time aligned using the cross-correlation algorithm;

[0007] Step 1: Signal framing and time alignment: the above three groups of signals are segmented, and the segmented signals are time-aligned again according to step 0;

[0008] Step 2, 1 / 12 th Octave filtering is used to divide the three groups of signals in each segment by 1 / 12 th Octave filtering is used to obtain a time domain signal with 127 frequency bands;

[0009] Step 3: Calculate the level. According to IEC61672-1-2013, calculate the level of each frequency band signal. The calculation formula is as follows:

[0010]

[0011] Where p is the time domain sound pressure signal, P0 = 2e-5, T is the integration time, i.e. the time constant, which is set to 5ms;

[0012] Step 4: auditory correlation calculation: the calculated level value data is subjected to auditory correlation calculation to obtain the auditory model results of each segment and each frequency band of the three groups of input signals;

[0013] Step 5: feature statistics. The specific method is as follows: according to the distribution of noise and speech segments in Clean Signa l, the result of step 4 is cut, the speech part and the noise part are retained respectively, and the mean and root mean square are calculated;

[0014] Step 6: deep learning, performing deep learning on the feature values ​​in the above feature statistics to obtain an evaluation value of the speech quality;

[0015] Step 7: Map the evaluation value to the MOS range to obtain the final calculation result.

[0016] Step 8. End.

[0017] Preferably, the auditory correlation calculation step in step 4 is as follows:

[0018] Step 9, the level value signals of the 127 frequency bands calculated in step 3;

[0019] Step 10: frequency band smoothing algorithm, frequency band smoothing is performed on the level value signal of each frequency band. The frequency band smoothing algorithm is to linearly average the predicted value 1 obtained after linear regression of the upper 8 frequency bands corresponding to a certain frequency band and the predicted value 2 obtained after linear regression of the corresponding lower 8 frequency bands, to obtain a signal after frequency band smoothing;

[0020] Step 11: Time linear regression algorithm: perform linear regression calculation in the time domain on the time axis for the signal after frequency band smoothing, and calculate the linear regression coefficient by taking the signal of 0-200ms;

[0021] Step 12: nonlinear transformation, converting the level value unit (Pa) into the nonlinear field unit cPa;

[0022] Step 13: End, the output result is the auditory model result of the input signal.

[0023] Preferably, the linear regression coefficient calculation formula in step 11 is as follows:

[0024] Assume that the sampling rate of the signal is f s , then the number of data points in the first 200ms is N = 0.2*f s According to the linear regression algorithm, the level value of point N+1 is estimated based on the level values ​​of points 1-N. The linear regression formula is as follows:

[0025] L1=α*1+b

[0026] L2=a*2+b

[0027]

[0028] L N =α*N+b

[0029] Linear regression is to calculate the values ​​of a and b. Solving the above formula, we have

[0030]

[0031] make have

[0032]

[0033] The final calculation is

[0034] L N+1 =α*(N+1)+b

[0035] Preferably, the specific conversion formula in step 12 is as follows:

[0036] Let x = p / p0, where p is the data output from step 10 and step 11. The data needs to be preprocessed. That is, if p < 0, then p = 0, p0 = 2 × 10 -5 ;

[0037] If x(i)≤5.66

[0038] cPa(i)=x(i)-α*x(i) 2

[0039] If x(i)>5.66

[0040] cPa(i)=b*x(i) 0.2 -c

[0041] Among them: α=0.0393, b=11.1161, c=11.32.

[0042] Preferably, the Clean Signal, Processed Signal and Unprocessed Signal are segmented with 4 seconds as one frame.

[0043] The objective evaluation method for the voice quality of a voice communication system provided by the embodiment of the present invention makes up for the shortcomings of the voice quality evaluation algorithm in the existing market. It can not only guide the efficient design and development of call products in the development stage, but also can quickly and accurately detect the call quality of the product in the product acceptance stage. At the same time, it has created a precedent for my country to independently realize the voice quality evaluation algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A voice quality test diagram of a voice communication system provided by an embodiment of the present invention for an objective evaluation method of voice quality of a voice communication system;

[0045] Figure 2 A flow chart of a speech quality evaluation algorithm of an objective evaluation method for speech quality of a speech communication system provided by an embodiment of the present invention;

[0046] Figure 3 A block diagram of an auditory correlation algorithm for an objective evaluation method of voice quality of a voice communication system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0048] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0049] An embodiment of the present invention provides an objective evaluation method for voice quality of a voice communication system, comprising the following steps:

[0050] Step 0: Time alignment: using the Clean Signa l as the reference signal, the Processed Signa l and the Unprocessed Signa l signals are time aligned using the cross-correlation algorithm;

[0051] Step 1: divide the signals into frames and time align them. The three groups of signals are segmented into 4 seconds per frame, and the segmented signals are time-aligned again according to step 0.

[0052] Step 2, 1 / 12 th Octave filtering is used to divide the three groups of signals in each segment by 1 / 12 th Octave filtering is used to obtain a time domain signal with 127 frequency bands;

[0053] Step 3: Calculate the level. According to IEC61672-1-2013, calculate the level of each frequency band signal. The calculation formula is as follows:

[0054]

[0055] Where p is the time domain sound pressure signal, P0 = 2e-5, T is the integration time, i.e. the time constant, which is set to 5ms;

[0056] Step 4: auditory correlation calculation: the calculated level value data is subjected to auditory correlation calculation to obtain the auditory model results of each segment and each frequency band of the three groups of input signals;

[0057] Step 5: feature statistics. The specific method is as follows: according to the distribution of noise and speech segments in Clean Signa l, the result of step 4 is cut, the speech part and the noise part are retained respectively, and the mean and root mean square are calculated;

[0058] Step 6: deep learning, performing deep learning on the feature values ​​in the above feature statistics to obtain an evaluation value of the speech quality;

[0059] Step 7: Map the evaluation value to the MOS range to obtain the final calculation result.

[0060] Step 8. End.

[0061] The objective evaluation method of the voice quality of a voice communication system provided in this application makes up for the shortcomings of the voice quality evaluation algorithms currently on the market. It can not only guide the efficient design and development of call products in the development stage, but also quickly and accurately detect the call quality of products in the product acceptance stage. At the same time, it has set a precedent for my country to independently implement a voice quality evaluation algorithm.

[0062] In one example of the present invention, the speech evaluation method can be applied to tests in multiple scenarios such as in a car and in an anechoic room. The player 1 is used to provide an audio signal to an artificial head 4 for playback and a speech quality evaluation algorithm 5 (denoted as Clean Signa l). The call terminal 2 collects and provides a processed noisy audio signal (denoted as Processed S i gna l). The reference microphone 3 collects and provides an unprocessed noisy audio signal (denoted as Unprocessed S i gna l). The above three groups of signals are input into the speech quality evaluation algorithm 5 for speech quality evaluation. In addition, the objective speech quality evaluation algorithm introduced in the present invention maps the calculation results to MOS scores with a value range of 0-5.

[0063] As a preferred embodiment of the present invention, the auditory correlation calculation step in step 4 is as follows:

[0064] Step 9, the level value signals of the 127 frequency bands calculated in step 3;

[0065] Step 10: frequency band smoothing algorithm, frequency band smoothing is performed on the level value signal of each frequency band. The frequency band smoothing algorithm is to linearly average the predicted value 1 obtained after linear regression of the upper 8 frequency bands corresponding to a certain frequency band and the predicted value 2 obtained after linear regression of the corresponding lower 8 frequency bands, to obtain a signal after frequency band smoothing;

[0066] Step 11: Time linear regression algorithm: perform linear regression calculation in the time domain on the time axis for the signal after frequency band smoothing, and calculate the linear regression coefficient by taking the signal of 0-200ms;

[0067] Step 12: nonlinear transformation, converting the level value unit (Pa) into the nonlinear field unit cPa;

[0068] Step 13: End, the output result is the auditory model result of the input signal.

[0069] In an example of the present invention, the linear regression method used in step 10 can refer to step 11, but is only implemented in the frequency domain.

[0070] As another preferred embodiment of the present invention, the linear regression coefficient calculation formula in step 11 is as follows:

[0071] According to the linear regression algorithm, the level value of point N+1 is estimated based on the level values ​​of points 1-N. The linear regression formula is as follows:

[0072] L1=a*1+b

[0073] L2=a*2+b

[0074]

[0075] L N =α*N+b

[0076] Linear regression is to calculate the values ​​of a and b. Solving the above formula, we have

[0077]

[0078] make have

[0079]

[0080] The final calculation is

[0081] L N+1 =a*(N+1)+b

[0082] In one embodiment of the present invention, it is assumed that the sampling rate of the signal is f s , then the number of data points in the first 200ms is N = 0.2*f s .

[0083] As another preferred embodiment of the present invention, the specific conversion formula in step 12 is as follows:

[0084] Let x = p / p0, where p is the data output from step 10 and step 11. The data needs to be preprocessed. That is, if p < 0, then p = 0, p0 = 2 × 10 -5 ;

[0085] If x(i)≤5.66

[0086] cPa(i)=x(i)-a*x(i) 2

[0087] If x(i)>5.66

[0088] cPa(i)=b*x(i) 0.2 -c

[0089] In one example of the present invention, a, b and c are 0.0393, 11.1161 and 11.32 respectively.

[0090] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

[0091] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. An objective evaluation method for voice quality of a voice communication system, characterized in that: The following steps are involved: Step 0: Time alignment: using Clean Signal as the reference signal, the cross-correlation algorithm is used to time align the Processed Signal and the Unprocessed Signal. Clean Signal is a pure signal. Processed Signal is a processed noisy audio signal collected and provided by the call terminal. Unprocessed Signal is an unprocessed noisy audio signal collected and provided by the reference microphone. Step 1: Signal framing and time alignment: the above three groups of signals are segmented, and the segmented signals are time-aligned again according to step 0; Step 2, 1 / 12 th Octave filtering is used to divide the three groups of signals in each segment by 1 / 12 th Octave filtering is used to obtain a time domain signal with 127 frequency bands; Step 3: Calculate the level. According to IEC61672-1-2013, calculate the level value of each frequency band signal. The calculation formula is as follows: Where p is the time domain sound pressure signal, P0 = 2e-5, T is the integration time, i.e. the time constant, which is set to 5ms; Step 4: auditory correlation calculation: the calculated level value data is subjected to auditory correlation calculation to obtain the auditory model results of each segment and each frequency band of the three groups of input signals; Step 5: Feature statistics. The specific method is as follows: according to the distribution of noise and speech segments in the Clean Signal, the result of Step 4 is cut, the speech part and the noise part are retained respectively, and the mean and root mean square are calculated; Step 6: deep learning, performing deep learning on the feature values ​​in the above feature statistics to obtain an evaluation value of the speech quality; Step 7: Evaluation value: Map the evaluation value to the MOS range to obtain the final calculation result. Step 8: End; The auditory correlation calculation steps in step 4 are as follows: Step 9, the level value signals of the 127 frequency bands calculated in step 3; Step 10: frequency band smoothing algorithm, frequency band smoothing is performed on the level value signal of each frequency band. The frequency band smoothing algorithm is to linearly average the predicted value 1 obtained after linear regression of the upper 8 frequency bands corresponding to a certain frequency band and the predicted value 2 obtained after linear regression of the corresponding lower 8 frequency bands, to obtain a signal after frequency band smoothing; Step 11: Time linear regression algorithm: perform linear regression calculation in the time domain on the time axis for the signal after frequency band smoothing, and calculate the linear regression coefficient by taking the signal of 0-200ms; Step 12: nonlinear transformation, converting the level value unit Pa into the nonlinear field unit cPa; Step 13: End, the output result is the auditory model result of the input signal.

2. The objective evaluation method of voice quality of a voice communication system according to claim 1, characterized in that: The linear regression coefficient calculation formula in step 11 is as follows: Assume that the sampling rate of the signal is f s , then the number of data points in the first 200ms is N = 0.2*f s According to the linear regression algorithm, the level value of point N+1 is estimated based on the level values ​​of points 1-N. The linear regression formula is as follows: Linear regression is to calculate the values ​​of a and b. Solving the above formula, we have make have The final calculation is L N+1 =a*(N+1)+b。 3. The objective evaluation method of voice quality of a voice communication system according to claim 1, characterized in that: The specific conversion formula in step 12 is as follows: Let x = p / p0, where p is the data output from step 10 and step 11. The data needs to be preprocessed. That is, if p < 0, then p = 0, p0 = 2 × 10 -5 ; If x(i)≤5.66 cPa(i)=x(i)-a*x(i) 2 If x(i)>5.66 cPa(i)=b*x(i) 0.2 -c Among them: a=0.0393, b=11.1161, c=11.

32.

4. The objective evaluation method of voice quality of a voice communication system according to claim 1, characterized in that: The CleanSignal, ProcessedSignal and UnprocessedSignal signals are segmented with 4 seconds as one frame.

Images