A method for modeling the annoyance of a superposition sound with respect to an audio injection

Abstract

The application provides a noise annoyance modeling method for an audio injection system, which comprises the following steps: firstly, according to a target sound, a corresponding control sound is selected to perform time domain superposition, so as to obtain a plurality of superimposed sound samples; then, a subjective evaluation experiment and data processing are performed on the obtained superimposed sound samples; then, psychoacoustic parameters of the target sound, the control sound and the superimposed sound are calculated, and the difference between the corresponding psychoacoustic parameters of the target sound and the control sound is calculated; the difference between the calculated psychoacoustic parameters of the target sound and the control sound is taken as an independent variable, and the corresponding superimposed sound annoyance score value is taken as a dependent variable; the correlation between the independent variable and the dependent variable, the correlation between the independent variables and the collinearity between the independent variables are calculated; according to the calculation results, the modeling independent variable is determined; the determined modeling independent variable and the superimposed sound annoyance score value are taken as the dependent variable, and a multiple linear regression is adopted to realize the audio injection superimposed sound annoyance modeling. The application improves the precision and parameter correlation of the model, and further improves the objective evaluation of the audio injection superimposed sound annoyance.

Description

Technical Field

[0001] This invention relates to the field of sound quality technology, and specifically to a noise annoyance modeling method for audio injection systems. Background Technology

[0002] The method of reducing the annoyance of original noise by injecting controlled sound into it is called audio injection. The original noise is also called the target sound (TS), the added sound is called the controlled sound (CS), and the mixed sound is called the combined noise (CN). Compared with traditional subtractive noise control strategies, audio injection has the advantages of simple implementation, low cost, and ease of engineering implementation, and has gradually become a cutting-edge topic in the field of noise control, attracting widespread attention and importance from researchers and industry.

[0003] The foundation of the audio injection method lies in establishing a mixed noise annoyance model to evaluate and predict the annoyance of superimposed sounds. Domestic and international scholars have conducted extensive research on modeling the annoyance of various similar mixed noises, proposing several models, among which the most representative include: the energy summation model, the independent effect model, the strongest component (or dominant source) model, and the vector summation model. All of these mixed noise annoyance models are based on energy-based calculations of the total annoyance, neglecting the influence of auditory effects. The simple energy summation model assumes that the relationship between the dose and annoyance of each individual noise component in the mixed noise is consistent; however, research results on the annoyance of different specific noises have proven this conclusion invalid, especially when the individual noises are dissimilar, further increasing the annoyance prediction error. The independent utility model considers the different relationships between noise dose and annoyance among different noises, but this model assumes that the effects of each noise's annoyance are independent and unrelated; however, the destructive and constructive effects of noise have already refuted this view. While the strongest component model outperforms the two models mentioned above in application, it is not suitable for predicting annoyance of superimposed audio-injected sounds. The principal component of superimposed audio-injected sounds is the target noise, which is equivalent to using the model to predict the annoyance level of the target noise. It cannot reflect the modulation effect and the degree of annoyance improvement produced by different modulated sounds. The vector superposition model uses a constant linear correction to modify the relationship between the annoyance level of a single noise and the total annoyance level. However, different combinations of modulated sounds and target sounds produce different annoyance suppression results. A single parameter cannot explain the annoyance relationship between different combinations.

[0004] These models are helpful for predicting the annoyance level of similar mixed noise, but in audio injection, a considerable portion of the superimposed noise is dissimilar mixed noise. The composition of this mixed noise is quite complex, with significant differences in sound pressure level and spectrum between the components. The mechanism of action is not simply the superposition and masking of sound energy, but also includes time-frequency compensation and information masking. Existing mixed noise annoyance level models have large prediction errors and are not suitable for the specific engineering environment of audio injection applications.

[0005] Unlike studies on the annoyance of traffic noise and other mixed noises, the mixed noise (i.e., superimposed sound) in audio injection is artificially synthesized, allowing us to obtain the specific parameters of each component of the mixed noise. Based on this, the annoyance of a certain type of noise can be evaluated by modeling the annoyance of superimposed sound using the audio injection method.

[0006] Audio injection is an "additive control" method that reduces annoyance by improving the sound quality of noise, thereby lowering its annoyance level. Annoyance level is a concept based on psychoacoustics, describing a person's perception of annoyance from sound. Annoyance level assessment is mainly divided into subjective and objective assessments, with subjective assessment currently being the primary method. While subjective assessment can directly reflect the listener's actual psychological perception of sound, it is limited by the specific noise environment and can only assess the annoyance level of a specific type of noise, resulting in a significant workload and cost. Therefore, objective assessment of annoyance level is currently the focus of research. Annoyance level models are established using objective parameters from psychoacoustics, such as loudness, sharpness, roughness, fluctuation intensity, and subjective duration, to evaluate the annoyance level of a particular type of noise. Unlike the annoyance level assessment of a single type of noise, the annoyance level assessment of audio injection evaluates the annoyance level of a mixture of target sound and control sound (superimposed sound). Table 1 shows an example of experimental data from an audio injection study.

[0007] Table 1 Comparison of parameters for target sound, controlled sound, and superimposed sound.

[0008]

[0009] However, experiments revealed that due to the significant difference in sound pressure levels between the controlled sound and the target sound, generally exceeding 10 dB, the superimposed sound after mixing is predominantly composed of the target sound, with acoustic parameters essentially identical to the target sound. Consequently, the acoustic parameters of the superimposed sound formed by different controlled sounds exhibit relatively small differences, with average sound pressure level differences within 0.8%, average loudness differences within 8.5%, and average sharpness differences within 5.5%. When evaluating mixed noise, the multiple linear regression model established using parameters such as loudness, sharpness, roughness, and fluctuation intensity of the superimposed sound as independent variables exhibits significant errors, and the linear regression relationship between the independent variables and the dependent variable is not close, meaning the correlation between the independent variable parameters and the dependent variable is weak. Summary of the Invention

[0010] To address the issue that existing methods for evaluating mixed noise using parameters such as loudness, sharpness, roughness, and fluctuation intensity of superimposed sound as independent variables in the established multiple linear regression model have large errors and weak linear regression relationships between independent and dependent variables, this invention proposes a more error-free superimposed sound annoyance modeling method for audio injection. This method uses the difference in loudness, sharpness, roughness, and fluctuation intensity between the controlled sound and the target sound as independent variables to establish a multiple linear regression model, improving the model's accuracy and parameter correlation, and further refining the objective evaluation of the annoyance of audio-injected superimposed sound.

[0011] The technical solution of this invention is as follows:

[0012] A method for modeling superimposed noise levels for audio injection includes the following steps:

[0013] Step 1: Based on the target sound, select the corresponding control sound for time-domain superposition to obtain several superimposed sound samples;

[0014] Step 2: Conduct a subjective evaluation experiment on annoyance level and process the data of the superimposed sound samples obtained in Step 1;

[0015] Step 3: Calculate the psychoacoustic parameters of the target sound, the controlled sound, and the superimposed sound, and calculate the difference in the corresponding psychoacoustic parameters between the target sound and the controlled sound;

[0016] Step 4: Using the difference in psychoacoustic parameters between the target sound and the controlled sound calculated in Step 3 as the independent variable and the corresponding superimposed sound annoyance score as the dependent variable, calculate the correlation between the independent and dependent variables, as well as the correlation and collinearity among the independent variables. Based on the calculation results, determine the modeling independent variables.

[0017] Step 5: Using the independent variables determined in Step 4, the noise annoyance score is superimposed as the dependent variable. Multiple linear regression is used to achieve audio injection superimposed noise annoyance modeling.

[0018] Furthermore, in step 1, target sound samples are first collected, and then sound analysis software is used to draw the time-domain diagram, spectrum diagram, and time-frequency diagram of the target sound samples to analyze the time-domain and frequency-domain characteristics of the target sound. Based on the time-domain and frequency-domain characteristics of the target sound, the corresponding control sound is selected and superimposed with the target sound in the time domain to obtain the superimposed sound sample.

[0019] Furthermore, in step 2, after conducting a subjective evaluation experiment on the annoyance level of the superimposed sound samples, the validity of the experimental data is tested: the range of the participants' scores is analyzed, and large amounts of data with excessively different scores are removed; misjudgment analysis is performed on the participants' data to determine the consistency of the evaluation results of the same participant in several evaluations; correlation analysis is performed to determine the correlation between the evaluation results of the same participant in several evaluations; finally, cluster analysis is performed on the experimental data to remove participants' data with excessively different scoring standards; after removing invalid data, the average annoyance level corresponding to each superimposed sound is calculated to obtain the final annoyance level score result.

[0020] Furthermore, in step 3, the psychoacoustic parameters include loudness, sharpness, roughness, and wave intensity.

[0021] Furthermore, in step 3, the loudness is calculated using the Zwicker model.

[0022] Furthermore, in step 4, the determined modeling independent variables are the loudness difference ΔN, sharpness difference ΔS, roughness difference ΔR, and wave intensity difference ΔFL between the target sound and the controlled sound.

[0023] Beneficial effects

[0024] This invention provides a novel method for modeling the annoyance of superimposed sound in audio injection. By starting with the difference between the parameters of the controlled sound and the superimposed sound, it more intuitively reflects the influence of their relative relationship on the annoyance of superimposed sound in audio injection, thus improving the model accuracy.

[0025] Furthermore, this invention advances theoretical research on audio injection methods. It intuitively reveals the numerical relationship between the difference in parameters between the control sound and the target sound and the annoyance level, which helps in studying the control mechanism of audio injection, thereby guiding the selection of the control sound and the optimization of the audio injection system.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0028] Figure 1 Noise spectrum diagram of a range hood;

[0029] Figure 2 : Correlation coefficient results of participant ratings;

[0030] Figure 3 Results of participant rating misjudgment analysis;

[0031] Figure 4 : Results of cluster analysis of participant ratings;

[0032] Figure 5 Line graph comparing experimental results with the superimposed sound parameter annoyance prediction model;

[0033] Figure 6 Line graph comparing experimental results with parameter difference annoyance prediction model;

[0034] Figure 7 Classical parametric modeling of residuals;

[0035] Figure 8 : Parameter difference modeling residuals;

[0036] Figure 9 : Flowchart of the present invention. Detailed Implementation

[0037] This invention proposes a method for modeling superimposed noise level of audio injection. It establishes a multiple linear regression model using the difference in loudness, sharpness, roughness, and fluctuation intensity between the controlled sound and the target sound as independent variables, thereby improving the accuracy and parameter correlation of the model and further refining the objective evaluation of superimposed noise level of audio injection.

[0038] To clearly describe this method and process, the relevant terminology and explanations will be provided first:

[0039] ① Audio Injection Method (AIM): This method involves injecting other sounds into the original noise to reduce its annoyance. The original noise is called the target sound, the added sound is called the modulating sound, and the mixed sound is called the superimposed sound.

[0040] ② Multiple Correlation Coefficient R and Coefficient of Determination R in a Multiple Regression Model 2 (Determinate Coefficient): The multiple correlation coefficient R, also known as the multivariate correlation coefficient, represents the coefficient of variation of all independent variables x in the model. i The degree of linear regression relationship between the variable y and the dependent variable y; the coefficient of determination R. 2 It equals the square of the multiple correlation coefficient and represents the proportion of the total change in the dependent variable y that can be explained by the independent variables in the regression model.

[0041] ③ Psychoacoustic parameters: loudness N, sharpness S, roughness R, and fluctuation strength FL; loudness difference ΔN, sharpness difference ΔS, roughness difference ΔR, and fluctuation strength difference ΔFL.

[0042] ④ Annoyance A: A parameter describing the degree of annoyance caused by noise. In this patent description, the annoyance score is set from 0 to 10, with an accuracy of 1. The lower the score, the less annoyance is felt, and the higher the score, the more annoyance is felt. The target noise score is set to 5.

[0043] This embodiment takes the audio injection system data of a range hood as an example, establishes an annoyance level model using the difference in psychoacoustic parameters between the target sound and the control sound, and compares it with traditional psychoacoustic parameter modeling. The entire modeling process is given as follows:

[0044] (1) Audio acquisition and synthesis of superimposed sound

[0045] According to the industry standard T-CAS 341-2019 "Test Method for Noise Quality of Range Hoods", noise samples of range hoods were recorded. Then, noise analysis software was used to draw time-domain diagrams, spectrum diagrams, and time-frequency diagrams of the sound samples. The time-domain and frequency-domain characteristics of the noise were analyzed. Appropriate control sounds and noise were selected for time-domain superposition to obtain mixed noise (superimposed sound) samples required for subjective evaluation experiments. A certain proportion of experimental sound samples were selected as a repeat control group.

[0046] In this embodiment, target noise is collected in a semi-anechoic chamber at a sampling frequency of 44100Hz. A manual head and a dual-channel SQuadriga III acquisition and playback system are used to record the noise of the range hood during normal operation, and its spectrum is analyzed. Figure 1 As shown.

[0047] The noise from range hoods is primarily broadband noise with some low-frequency line spectrum noise, resulting in a complex spectral composition. Traditional passive noise reduction methods are ineffective at reducing low frequencies, while active noise reduction measures are difficult and costly to implement. Therefore, audio injection is a good alternative. When adjusting the sound based on the spectral characteristics of the range hood noise, an audio signal with a wide spectral energy distribution and relatively uniform mid-to-high frequency energy distribution should be selected.

[0048] A total of 118 audio files of potentially applicable control sounds were obtained through methods such as artificial synthesis and online download. Each file was 20 seconds long, and its sound pressure level was adjusted. Three different sound pressure levels were retained for each control sound to minimize the impact of an unsuitable signal-to-noise ratio on the final annoyance score. The control sounds and target sounds were then time-domain superimposed, and the superimposed sounds were initially screened. Finally, 49 × 3 = 147 superimposed sounds were selected as the audio for the subjective evaluation experiment.

[0049] (2) Subjective evaluation experiment and data processing of superimposed noise annoyance

[0050] A subjective evaluation experimental protocol was developed, and participants were recruited to conduct the experiment. Participants' hearing and judgment abilities were confirmed to be normal. Prior to the experiment, participants received explanations and simple training to ensure its smooth execution. To eliminate environmental interference, the experiment was conducted in a noise-free laboratory using headphones, and scoring data was collected through questionnaires.

[0051] To ensure the accuracy of subsequent analysis results, the data must first be validated. This involves analyzing the range of participant ratings and removing data with excessively large discrepancies; performing misjudgment analysis to assess the consistency of multiple evaluations by the same participant; conducting correlation analysis to determine the correlation between multiple evaluations by the same participant; and finally, performing cluster analysis on the experimental data to remove data from participants with significantly different rating standards. After removing invalid data, the average annoyance level corresponding to each superimposed sound is calculated, yielding the final annoyance score.

[0052] In this embodiment, based on the annoyance level model to be established, the annoyance level score of each superimposed sound is obtained. We select the reference scoring method (selecting the noise of the range hood as the reference sound to evaluate the annoyance level, ensuring the real effectiveness of the controlled sound and the accuracy of the experimental results) for experimental research. The annoyance level score is an 11-level scoring scale of 0 to 10 points, the target sound score is 5, and the scoring step size is 1. The superimposed sound is scored. A score below 5 indicates that the annoyance level of the superimposed sound is lower than that of the target noise, that is, a control effect is produced. The lower the score, the lower the annoyance level of the superimposed sound, and vice versa, as shown in Table 2.

[0053] A total of 72 participants were recruited for the experiment, with a male-to-female ratio close to 1:1. To ensure that each participant heard the same sound under identical conditions, a subjective evaluation experiment was conducted via headphone playback. The evaluation results were unaffected by room acoustics or participant location, and the system effectively shielded against external noise interference with minimal signal distortion. Sound sample segments were generated by randomly sorting the sound samples using Matlab. These segments were then transmitted via computer to a binaural headphone equalizer (HEADlab-compatible binaural headphone equalizers labP2) using Artemis software, and subsequently played to the participants through dynamic high-fidelity stereo headphones (SENNHEISER HD600). During playback, the computer controlled the playback duration, intervals, and order of the sound samples. This sound playback system fully considered the binaural masking effect, maximizing the ideal listening experience.

[0054] Table 2. Annoyance Level 11 Evaluation Scale

[0055]

[0056] Correlation analysis is a quantitative analysis method used to analyze the statistical relationships between objective things, and it is the first step in analyzing subjective evaluation experimental data. Correlation coefficients are typically used to accurately reflect the strength of the linear correlation between variables numerically. Correlation analysis is used in data processing to determine the correlation between multiple evaluations of different sound samples by the same subject. The experimenters were divided into three groups, and the correlation analysis results of the subjects were analyzed as follows: Figure 2 As shown in the figure. In the experimental data of this embodiment, the correlation coefficient of the subject data scores is considered to be weak, and therefore the subject data with a correlation coefficient of less than 0.5 are excluded.

[0057] Misjudgment analysis is used to determine the consistency of evaluation results from multiple evaluations by the same subject. Misjudgments in the raw data are categorized into two types: "permissible misjudge" and "impermissible misjudge." Ideally, within the same laboratory, the p evaluation data obtained from p independent repeated evaluations of the same sound sample by the same evaluator using the same evaluation method within a short period should be completely consistent. However, considering the influence of various uncertainties in reality, under the above repeatability conditions, a certain degree of difference should be allowed between the results of multiple repeated evaluations of the same sound sample by the same evaluator, as long as the probability of this difference exceeding a certain value is lower than the significance level of the acceptability test.

[0058] On a continuous numerical scale, a difference of less than or equal to 1 between two consecutive evaluation scores is considered an acceptable misjudgment; otherwise, it is considered an unacceptable misjudgment. The misjudgment rate is obtained by counting the number of misjudgments made by the participants and then dividing by the total number of samples. The misjudgment analysis results are as follows: Figure 3 As shown. According to the misjudgment analysis, it is necessary to remove the subject data with a misjudgment rate higher than 0.3.

[0059] Cluster analysis is used to determine the consistency of ratings among different participants. Annoyance level is a subjective measure, and people's lifestyles and cultural backgrounds often influence their ratings; therefore, annoyance level ratings may exhibit significant individual differences. Clustering the participants' ratings yields the following results: Figure 4 As shown, most participants clustered into one group, while only a few participants moved away from the others, suggesting that they may have used different strategies when rating their level of annoyance.

[0060] (3) Calculation of psychoacoustic parameters

[0061] Psychoacoustic parameters, viewed and analyzed from a psychoacoustic perspective, reflect subjective auditory perception. Commonly used psychoacoustic parameters can be categorized into two types based on their model principles: loudness and other parameters based on loudness (e.g., sharpness, roughness, wave intensity, and tone modulation), the latter building upon the characteristic loudness of the former. In this embodiment, the loudness parameters and other parameters of the modulated sound, target sound, and superimposed sound are calculated using the Zwicker model and Matlab programming.

[0062] In this embodiment, the loudness, sharpness, roughness, and fluctuation intensity of the superimposed sound, the controlled sound, and the target sound are calculated. To ensure uniformity and universality of the calculation results, the calculation methods for all psychoacoustic parameters are as follows:

[0063] (a) Loudness

[0064] Loudness is a psychoacoustic parameter reflecting the subjective perception of sound intensity by the human ear. Commonly used loudness calculation models include the Stevens model, the Zwicker model, and the Moore model. The Zwicker model, considering factors such as the human ear transfer function, the auditory static threshold, and the sound field transfer function, proposes a loudness calculation method based on a critical frequency band excitation model. This method can calculate the loudness values ​​of steady-state or non-steady-state sound signals in reverberant and free sound fields, and has a wide range of applications. Therefore, this invention uses the Zwicker model to calculate loudness, and the subsequent sharpness calculation is also based on this model. The calculation formula is as follows:

[0065]

[0066] Where: N: loudness, unit is sone; n(z): characteristic loudness.

[0067] (b) Sharpness

[0068] Sharpness (S) is a psychoacoustic parameter that reflects the ratio of high-frequency to low-frequency sound components. The sharpness parameter calculation in this paper is based on the Zwicker model standard, and its formula is as follows:

[0069]

[0070] Where: S: sharpness, in acum; C1: calibration constant, usually taken as 0.11; n(z): characteristic loudness; z: psychoacoustic frequency, in bark; g(z): weighting curve.

[0071] In equation (2), the denominator is the loudness, and the weight curve is...

[0072] (c) Roughness

[0073] Roughness (R) reflects characteristics such as the magnitude of the modulation amplitude and the distribution of the modulation frequency of a signal. Its calculation formula is as follows:

[0074]

[0075] Where F: roughness, unit is asper; f mod Modulation frequency; ΔL E (z): Masking depth, calculated using the change in intensity of characteristic features across each frequency band. Replacement.

[0076] (d) Fluctuation intensity

[0077] Fluctuation strength (F) reflects the degree of loudness fluctuation perceived by the human ear, and its calculation formula is as follows:

[0078]

[0079] Where FL stands for wave intensity, measured in vacil.

[0080] The psychoacoustic parameters and differences of the superimposed sound, the controlled sound, and the target sound were calculated separately and are shown in Table 3.

[0081] Table 3 Psychoacoustic Parameter Scale

[0082]

[0083]

[0084] (4) Correlation analysis between independent and dependent variables

[0085] Organize the modeling parameters, calculate the correlation between loudness, sharpness, roughness, and fluctuation intensity as independent variables and annoyance level, and then calculate the correlation between loudness difference, sharpness difference, roughness difference, and fluctuation intensity difference as independent variables and annoyance level. Then calculate the cross-correlation between independent variables. If the correlation between independent variables is too strong or they are not relatively independent, they cannot be selected as modeling independent variables at the same time.

[0086] Correlation analysis (CA) is a quantitative analysis method for analyzing statistical relationships between objective phenomena, and it is the first step in analyzing subjectively evaluated experimental data. Correlation coefficients are typically used to accurately reflect the strength of the linear correlation between variables numerically. Commonly used correlation coefficients include Pearson's correlation coefficient, Spearman's correlation coefficient, and Kendall's correlation coefficient, with Pearson's correlation coefficient being the most frequently used. Therefore, this embodiment uses Pearson's correlation coefficient, abbreviated as r, and its mathematical definition is:

[0087]

[0088] In the formula, r: correlation coefficient; n: number of samples or data; x i : The value of variable x; y i : The value of variable y.

[0089] In this invention, the difference in psychoacoustic parameters between the target sound and the controlled sound is used as the independent variable, and the corresponding superimposed sound annoyance score is used as the dependent variable. The correlation between the independent and dependent variables is calculated. The value of r ranges from [-1, 1], where r = 1 indicates a perfectly positive correlation, r = -1 indicates a perfectly negative correlation, and r = 0 indicates no linear correlation. |r| > 0.8 indicates a strong linear relationship, and |r| < 0.3 indicates a very weak linear correlation.

[0090] Calculate the correlation and collinearity analysis between independent variables to determine whether there is a linear relationship between the selected independent variable data. If there is, they cannot be used for modeling at the same time. Calculate the linear relationship between the independent variable and the dependent variable to see if the independent variable has any influence on the value of the dependent variable.

[0091] (5) Audio injection superimposed noise level multiple linear regression modeling

[0092] This step is the core step of the invention: establishing a regression model to predict the annoyance level of superimposed sound noise. A multiple linear regression model is established with loudness difference, sharpness difference, roughness difference, and fluctuation intensity difference as independent variables and annoyance level as the dependent variable.

[0093] In regression analysis, a regression involving two or more independent variables is called multiple linear regression, used to examine the influence of multiple factors on the outcome. The influencing factors are called independent variables, and the affected outcome is called the dependent variable. Regression analysis aims to find the quantitative relationship between the dependent and independent variables, hoping to obtain the desired dependent variable by controlling for the independent variables. The basic model of multiple linear regression analysis is:

[0094] y = y′ + ε = b0 + b1x1 + ... + b px p +ε (6)

[0095] In the formula, y: observed value of the dependent variable; y′: estimated value of the dependent variable; ε: residual; b i Partial regression coefficient.

[0096] In equation (6), y′ is the estimated value of the dependent variable, representing the portion of the dependent variable determined by the independent variable; ε is the difference between the observed value and the estimated value of the dependent variable, representing the portion of the dependent variable not determined by the independent variable; b i This indicates the linear relationship between the estimated value of the dependent variable and the changes in the independent variable. The Ordinary Least Squares (OLS) method is typically used to determine the regression model; that is, the regression model with the smallest sum of squared errors among all regression models is selected as the desired regression model, and the coefficient of determination R is used. 2 To determine the quality of a model.

[0097] The following regression model is established to predict the annoyance level of superimposed sound noise. Using the annoyance level from a subjective evaluation experiment as the dependent variable, characteristic parameters of the sound samples are extracted to build the regression model. With the participants' annoyance score as the dependent variable, and selecting loudness difference, sharpness difference, roughness difference, and fluctuation intensity difference as independent variables, a multiple linear regression model of superimposed sound annoyance level is established. The data is fitted using the multiple linear regression method, and compared with a model constructed using the least squares method with loudness, sharpness, roughness, and fluctuation intensity as independent variables and annoyance score as the dependent variable.

[0098] Considering the limitations of the obtained sound samples and the requirement for the universality of the model, in practice, a portion of the data should be selected for modeling, with the remainder used for testing. In this embodiment, 60% of the annoyance scores from all sound samples are used for modeling, referred to as training data; the remaining 40% is used for testing, referred to as testing data. A superimposed sound annoyance model is established using classical parameters and parameter differences as independent variables, and the modeling results are as follows.

[0099] Table 4 Psychoacoustic Parameter Scale

[0100]

[0101]

[0102] As can be seen from the superimposed sound parameter model, the annoyance score is directly proportional to sharpness and roughness, and inversely proportional to loudness and fluctuation intensity. That is, the higher the sharpness and roughness of the superimposed sound, the lower the loudness and fluctuation intensity, the lower the annoyance, and the more comfortable the superimposed sound.

[0103] According to the parameter difference model, the greater the difference in loudness, sharpness, and roughness between the target sound and the control sound, the lower the annoyance level. In real engineering applications, the target sound is broadband noise, while the control sound is natural sound or light music with a more complex spectral composition. The sharpness and fluctuation intensity of the target sound are generally lower than those of the control sound, meaning that ΔS and ΔFL are often negative. Therefore, when selecting the control sound, the control sound with lower sharpness and higher fluctuation intensity can be given priority.

[0104] (6) Model error and R 2 calculate

[0105] The model's error is analyzed by calculating the errors for the training data, validation data, and all data, thus analyzing the model's accuracy and optimizing the model parameters accordingly. The subjective annoyance ratings are compared with the annoyance values ​​calculated using the multivariate model, and a line graph comparing the true and predicted values ​​is plotted. The relative and absolute errors are calculated, and the multiple correlation coefficient R0 and determination coefficient R0 of the multivariate regression model are also calculated. 2 .

[0106] To verify the applicability of the above model, the subjective annoyance rating results were compared with the annoyance values ​​calculated by the multivariate model, and a line graph comparing the true values ​​and predicted values ​​was plotted, as shown below. Figure 5 and Figure 6 As shown in the figure, the left side of the gray dashed line represents 60% of the training data and its prediction results, while the right side represents the reserved 40% of the test data and its prediction results. The figure shows that the trends of the experimental true values ​​and the model predictions are roughly the same. For most samples, the model's calculated values ​​are close to the experimental scores, indicating a good fit. The general prediction trends of both models are correct. For the prediction of some peak values, the parameter difference prediction model is more accurate, thus resulting in a smaller overall error.

[0107] Specific error data are shown in Table 5. Compared with the superimposed acoustic parameter annoyance model, the parameter difference annoyance model reduced the error by 2%, and the prediction error of the test group did not exceed 9%, indicating an improvement in overall model accuracy. This demonstrates that the parameter difference annoyance model has better predictive performance than the traditional superimposed acoustic parameter annoyance model.

[0108] Table 4 Comparison of Annoyance Level Model Errors

[0109]

[0110]

[0111] To make this invention more convincing, it is necessary not only to explain the model error but also to interpret the model's reliability, which can be achieved by calculating the coefficient of determination R. 2 The residual analysis of the models is used to illustrate this. The multiple correlation coefficient and determination coefficient of the two models are calculated separately and are shown in Table 6. R02 A larger value indicates a better fit, with a value of 1 indicating a perfect fit. It can be seen that the coefficient of determination of the parameter difference annoyance prediction model is higher than that of the superimposed sound parameter annoyance prediction model, indicating better linear fit, a greater correlation and influence of the independent variable on the dependent variable, and stronger reliability.

[0112] Table 6 Annoyance Level Model R and R 2 contrast

[0113] Model R <![CDATA[R 2 ]]> Superimposed sound parameter annoyance prediction model 0.4517 0.2040 Parameter difference annoyance prediction model 0.7980 0.6368

[0114] Calculate the residuals using the Durbin-Watson statistical method and plot a normal probability distribution (PP plot) to determine whether the residuals conform to a normal distribution. Figure 7 , Figure 8 It can be seen that the residuals of the two models basically conform to a normal distribution (similar to the sloping line), the residuals are relatively independent, and the models are usable.

[0115] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.

Claims

1. A method for modeling superimposed noise levels oriented towards audio injection, characterized in that: Includes the following steps: Step 1: Based on the target sound, select the corresponding control sound for time-domain superposition to obtain several superimposed sound samples; Step 2: Conduct a subjective evaluation experiment on annoyance level and process the data of the superimposed sound samples obtained in Step 1; Step 3: Calculate the psychoacoustic parameters of the target sound, the controlled sound, and the superimposed sound, and calculate the difference in the corresponding psychoacoustic parameters between the target sound and the controlled sound; Step 4: Using the difference in psychoacoustic parameters between the target sound and the controlled sound calculated in Step 3 as the independent variable and the corresponding superimposed sound annoyance score as the dependent variable, calculate the correlation between the independent and dependent variables, as well as the correlation and collinearity between the independent variables. Based on the calculation results, determine the modeling independent variables. Step 5: Using the independent variables determined in Step 4, the noise annoyance score is superimposed as the dependent variable. Multiple linear regression is used to achieve audio injection superimposed noise annoyance modeling.

2. The method for superimposed acoustic annoyance modeling oriented towards audio injection according to claim 1, characterized in that: In step 1, target sound samples are first collected, and then sound analysis software is used to draw the time-domain diagram, spectrum diagram and time-frequency diagram of the target sound samples. The time-domain and frequency-domain characteristics of the target sound are analyzed. Based on the time-domain and frequency-domain characteristics of the target sound, the corresponding control sound is selected and superimposed with the target sound in the time domain to obtain the superimposed sound sample.

3. The method for superimposed acoustic annoyance modeling oriented towards audio injection according to claim 1, characterized in that: In step 2, after conducting a subjective evaluation experiment on the annoyance level of the superimposed sound samples, the validity of the experimental data is tested: the range of the participants' scores is analyzed, and large amounts of data with excessively large differences in scores are removed; misjudgment analysis is performed on the participants' data to determine the consistency of the evaluation results of the same participant in several evaluations; correlation analysis is performed to determine the correlation between the evaluation results of the same participant in several evaluations; finally, cluster analysis is performed on the experimental data to remove participants' data with excessively large differences in scoring standards; after removing invalid data, the average annoyance level corresponding to each superimposed sound is calculated to obtain the final annoyance level score result.

4. The method for superimposed acoustic annoyance modeling oriented towards audio injection according to claim 1, characterized in that: In step 3, the psychoacoustic parameters include loudness, sharpness, roughness, and wave intensity.

5. The method for superimposed acoustic annoyance modeling oriented towards audio injection according to claim 4, characterized in that: In step 3, the loudness is calculated using the Zwicker model.

6. The method for superimposed acoustic annoyance modeling oriented towards audio injection according to claim 4, characterized in that: In step 4, the determined modeling independent variables are the loudness difference ΔN, sharpness difference ΔS, roughness difference ΔR, and wave intensity difference ΔFL between the target sound and the controlled sound.

Images