AI audio emotion test method, voice evaluation system and storage medium

Abstract

The application belongs to the technical field of audio signal processing and AI emotion recognition, and discloses an AI audio emotion test method, a voice evaluation system and a storage medium. The application obtains a standardized digital audio stream by collecting and converting target human voice simulation audio signals, respectively relies on architectural acoustics modeling and human voice sound rhythm modeling, corresponds to solve spatial reverberation attenuation characteristics and human voice sound disturbance characteristics, combines multi-dimensional basic emotion parameter calculation basic emotion characteristic fusion value, and relies on dynamic weight factor and dimensionless coefficient to fuse multiple features to obtain audio emotion evaluation confidence. The application adopts double modeling combination and four-order step-by-step serial dimensionless operation architecture, can simultaneously process spatial acoustic interference and human voice sound disturbance interference, solves the problem of audio emotion evaluation data deviation in a strong interference scene in the prior art, and effectively improves the standardization and accuracy of audio emotion evaluation in a complex interference environment.

Description

Technical Field

[0001] This invention belongs to the field of audio signal processing and AI emotion recognition technology, and discloses an AI audio emotion testing method, a voice evaluation system and a storage medium. Background Technology

[0002] Current AI audio emotion assessment technologies largely rely on the basic spectral and energy characteristics of human voice audio to perform emotion recognition calculations. They are designed with fixed computational logic only for pure human voice scenarios, failing to consider the spatial acoustic reverberation interference and vocal rhythm disturbances commonly present in real-world applications. In complex and highly interfering environments such as indoor spaces and noisy public places, spatial reverberation alters the amplitude and frequency transmission characteristics of audio signals. Fluctuations in speech rate and changes in vocal muscle state generate additional audio disturbances. These two types of interference superimpose on the original audio emotion features, making it impossible for traditional single-dimensional emotion feature calculation methods to isolate interfering data. Traditional technologies lack a two-dimensional interference correction mechanism considering both acoustic space and vocal production, making it impossible to distinguish between valid emotion features and interference features. Ultimately, this results in biased audio emotion assessment results, failing to meet the high-precision assessment requirements in highly interfering scenarios.

[0003] There is an urgent need in the industry for an audio emotion assessment technology that can simultaneously isolate spatial acoustic interference and human voice disturbances and achieve multi-feature linkage correction. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based audio emotion testing method, executed by a computer device's processor, comprising the following steps:

[0005] S1. Collect the analog audio signal corresponding to the target human voice and convert the analog audio signal into a standardized digital audio stream;

[0006] S2. Based on architectural acoustic modeling, the standardized digital audio stream is analyzed, spatial acoustic parameters are extracted, and spatial reverberation attenuation characteristics are calculated.

[0007] S3. Based on the human voice rhythm modeling method, the standardized digital audio stream is analyzed, the human voice speech rate parameter is extracted, and the human voice perturbation feature is calculated by combining the spatial reverberation attenuation feature.

[0008] S4. Extract the time-domain energy parameters, frequency-domain spectrum parameters, and Mel spectrum basic emotional parameters corresponding to the standardized digital audio stream, and calculate the basic emotional feature fusion value by combining the human voice perturbation features.

[0009] S5. Call the preset dynamic weighting factor and dimensional normalization coefficient, and fuse the spatial reverberation attenuation feature, the human voice perturbation feature and the basic emotional feature fusion value to calculate the audio emotional assessment confidence level.

[0010] This method adopts a cross-disciplinary approach combining architectural acoustic modeling and human vocal rhythm modeling, and completes audio emotion assessment data processing through a fourth-order step-by-step serial substitution dimensionless computational architecture.

[0011] Preferably, the acquisition and conversion of the standardized digital audio stream in S1 includes noise reduction sampling, signal amplitude calibration and timing regularization processing of the original analog audio signal, outputting a digital audio timing signal with continuous timing and uniform amplitude, and synchronously distributing the digital audio timing signal to the corresponding parameter parsing and calculation unit.

[0012] More preferably, the spatial acoustic parameters in S2 include the instantaneous audio sound pressure amplitude, the spatial reference sound pressure amplitude, the spatial reverberation cutoff frequency, the effective sampling duration of a single audio segment, and the spatial acoustic damping coefficient. The instantaneous audio sound pressure amplitude, the spatial reference sound pressure amplitude, the spatial reverberation cutoff frequency, the effective sampling duration of a single audio segment, and the spatial acoustic damping coefficient are all time-domain acoustic physical parameters extracted in real time from the standardized digital audio stream.

[0013] In a further preferred embodiment, the speech rate parameter in S3 includes the real-time instantaneous speech rate and the standard reference speech rate corresponding to the language. The calculation process of the speech perturbation features introduces spatial acoustic interference parameters and speech muscle linkage correction parameters to complete the joint calculation of environmental interference data and speech movement interference data.

[0014] In a further preferred embodiment, this method is configured with uniquely identified operational parameters, each with a fixed physical meaning and fixed dimensions. All operational parameters follow dimensionless homogeneous operational rules and are used for step-by-step calculation of multi-dimensional audio features and unified calculation of emotion assessment results.

[0015] In a further preferred embodiment, this method is configured with a fixed fourth-order operational timing logic to sequentially complete spatial reverberation attenuation feature calculation, human voice perturbation feature calculation, basic emotional feature fusion calculation, and emotional assessment confidence comprehensive calculation. The output data of the preceding calculation is used as the input data of the subsequent calculation, forming a unidirectional serial data operation link.

[0016] In a further preferred embodiment, the basic emotional feature fusion operation of this method uses the hyperbolic tangent function to perform numerical transformation processing on the human voice perturbation feature data, and the transformed values ​​participate in the fusion calculation of time-domain energy parameters and frequency-domain spectrum parameters.

[0017] In a further preferred embodiment, after obtaining the audio emotion assessment confidence score through S5 computation, the standardized digital audio stream, all computational parameters, intermediate calculation results of the fourth-order computation, and the audio emotion assessment confidence score are written into the embedded storage medium, while the audio emotion assessment confidence score data is output to the outside.

[0018] A voice evaluation system, applied to any of the AI ​​audio emotion evaluation methods described above, includes an audio signal acquisition module, a spatial acoustic feature analysis module, a human voice production feature analysis module, an AI emotion basic feature extraction module, a multi-feature operation module, and a data storage and result output module.

[0019] The audio signal acquisition module is used to acquire the target human voice audio analog signal in the environment, complete the conversion between analog and digital signals, output a standardized digital audio stream and distribute it synchronously to the other modules.

[0020] The spatial acoustic feature analysis module is electrically connected to the audio signal acquisition module and is used to receive standardized digital audio streams, extract spatial acoustic parameters, and calculate reverberation attenuation feature output values.

[0021] The human voice production feature analysis module is electrically connected to the audio signal acquisition module and the spatial acoustic feature analysis module. It is used to receive the standardized digital audio stream and the reverberation attenuation feature output value, and calculate the human voice production disturbance coefficient output value.

[0022] The AI ​​emotion basic feature extraction module is electrically connected to the audio signal acquisition module and is used to receive standardized digital audio streams and extract basic emotion parameters corresponding to time-domain energy, frequency-domain spectrum and Mel spectrum.

[0023] The multi-feature operation module is electrically connected to the spatial acoustic feature analysis module, the human voice production feature analysis module, and the AI ​​emotion basic feature extraction module, respectively, and is used to execute the fourth-order step-by-step serial substitution operation logic to complete the multi-feature data operation and audio emotion assessment confidence calculation.

[0024] The data storage and result output module is electrically connected to the multi-feature operation module and is used to store algorithm programs, model parameters and evaluation data, and output audio emotion evaluation result data.

[0025] A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the AI ​​audio emotion testing method described above.

[0026] The technical effects include:

[0027] This invention addresses the core deficiency of existing technologies in simultaneously separating the dual interferences of spatial reverberation and human vocalization. It leverages a dual-modeling architecture of architectural acoustics and vocal rhythm, coupled with fourth-order dimensionless homogeneous serial computation logic, to achieve layered quantification correction of environmental interference and vocal disturbances. This invention standardizes the dimensions of all computational parameters, avoiding data disorder issues associated with multi-parameter calculations. Through a step-by-step feature correction and fusion mechanism, it accurately selects effective emotional features, effectively solving the problem of distorted audio emotion assessment results in highly interfering scenarios, and significantly improving the accuracy and standardization of AI audio emotion assessment in complex environments. Attached Figure Description

[0028] Figure 1 This is a flowchart of the AI ​​audio emotion testing method in this application;

[0029] Figure 2 This is a connection diagram of the AI ​​audio emotion testing and voice evaluation system of this application. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0031] Please see Figures 1-2 This embodiment provides an AI audio emotion assessment method for strong interference scenarios, which is executed by the processor of a computer device. This method combines architectural acoustic modeling and human vocal rhythm modeling across disciplines, and completes audio emotion assessment data processing through a fourth-order step-by-step serial substitution dimensionless computing architecture.

[0032] Step S1 involves acquiring the analog audio signal corresponding to the target person's voice and converting it into a standardized digital audio stream. A microphone deployed at the testing site captures the analog audio signal emitted by the target person. This analog audio signal is a continuous-time signal propagating through the air medium from the vibration of the vocal cords. The microphone's transduction principle is based on sound pressure driving diaphragm vibration, which in turn causes a coil to move in a magnetic field, generating an induced electromotive force, or utilizing the capacitance change of an electret film to output a voltage proportional to the sound pressure. The analog audio signal is input to a signal conditioning circuit, which internally includes a preamplifier, an anti-aliasing filter, and an analog-to-digital converter. The preamplifier is configured as an instrumentation amplifier topology with a common-mode rejection ratio (CMRR) of at least 90 dB in the power frequency band. It linearly boosts the microvolt to millivolt-level analog audio signal output by the microphone to a voltage amplitude range recognizable by the analog-to-digital converter, typically a peak-to-peak value of 5 volts. The anti-aliasing filter is an eighth-order Butterworth low-pass filter with passband ripple controlled to within 0.1 dB and stopband attenuation not less than 48 dB per octave. It performs low-pass filtering on the amplified signal, filtering out frequency components higher than half the sampling frequency based on the Nyquist sampling theorem, thus physically blocking spectral aliasing. The analog-to-digital converter (ADC) adopts a successive approximation architecture with a 16-bit resolution, a sampling frequency of 48 kHz, and an integral nonlinearity error not exceeding ±1 least significant bit. It discretizes the filtered signal in time and amplitude to generate the original digital audio sequence. This original digital audio sequence is written to the processor's internal cache via a serial peripheral interface or internal bus. The cache is a static random access memory (SRAM) configured to hold at least 500 milliseconds of the original digital audio sequence. The processor loads audio preprocessing instructions, which perform three processes on the original digital audio sequence: noise reduction sampling, signal amplitude calibration, and timing warping. In the noise reduction sampling stage, the processor calls the spectral subtraction noise cancellation algorithm. A silent segment of at least 100 milliseconds is reserved at the beginning or end of the original digital audio sequence. The noise power spectrum is extracted from this silent segment, and the squared mean of the noise amplitude spectrum is calculated using a 2048-point Fast Fourier Transform. The noise power spectrum is then subtracted from the overall signal power spectrum. The original phase of the subtracted amplitude spectrum is preserved, and the noise-reduced time-domain signal is reconstructed using an inverse Fourier transform. In the signal amplitude calibration stage, the processor iterates through all sampling points of the noise-reduced sequence, retrieves the maximum absolute amplitude, and uses the ratio of the reference amplitude to the maximum absolute amplitude as the calibration gain. The reference amplitude is set to 80% of the full-scale amplitude. This gain is multiplied point-by-point into each sample value in the sequence to unify the overall signal amplitude within the standard range defined by the reference amplitude.In the timing warping process, the processor invokes a dynamic time warping algorithm, using equally spaced timestamps generated by a standard clock source (a temperature-compensated crystal oscillator with a frequency stability within ±1 / 1,000,000) as a reference. This corrects for sampling point timestamp offsets caused by microphone clock jitter or transmission delays. The warped timestamp sequence corresponds one-to-one with the sampling point sequence, with the difference between adjacent timestamps remaining constant as the reciprocal of the sampling period. This results in a digital audio timing signal with continuous timing and uniform amplitude. This digital audio timing signal is the standardized digital audio stream. The processor synchronously distributes this standardized digital audio stream to the corresponding storage areas of subsequent parameter parsing and processing units via a data bus. The distribution process is implemented through a direct memory access controller (DMI). When the standardized digital audio stream is generated, the DMI triggers a burst transmission, transferring a complete frame of standardized digital audio stream data to the dual-port random access memory (DRAM) of each processing unit. Each processing unit can then read a copy of the data from its respective storage area in parallel.

[0033] After acquiring the standardized digital audio stream, step S2 is executed. Based on architectural acoustic modeling, the standardized digital audio stream is analyzed to extract spatial acoustic parameters and calculate the spatial reverberation attenuation characteristics. The spatial acoustic parameters include the instantaneous audio sound pressure amplitude, the spatial reference sound pressure amplitude, the spatial reverberation cutoff frequency, the effective sampling duration of a single audio segment, and the spatial acoustic damping coefficient. These five types of parameters are all time-domain acoustic physical parameters extracted from the standardized digital audio stream. The processor extracts an audio frame from the standardized digital audio stream, with a duration equal to the effective sampling duration of a single audio segment. The sound pressure amplitude is extracted point-by-point from this audio frame. The signal voltage value at the current sampling point is taken and divided by the microphone sensitivity coefficient, which is calibrated at 10 mV / Pascal, to obtain the instantaneous audio sound pressure amplitude corresponding to the current sampling point, with dimensions in M·L. -1 ·T -2 The spatial reference sound pressure amplitude is obtained based on the environmental conditions at the testing site: when the target person is silent, the steady-state background sound signal continuously collected by the microphone is converted into a steady-state sound pressure amplitude sequence through the same conversion process. The sequence length is not less than 1 second. The long-time root mean square value of this sequence is taken as the spatial reference sound pressure amplitude. The long-time root mean square value is calculated by taking the square root of the mean of the squares of each sampling point, with the dimension M·L. -1 ·T -2The extraction of the spatial reverberation cutoff frequency is implemented as follows: An audio frame is passed through a group of bandpass filters with center frequencies increasing by one-third octaves. The bandpass filter group covers a frequency band from 100 Hz to 10000 Hz, and each bandpass filter is a fourth-order infinite impulse response filter, obtaining the time-domain response of each frequency band. A Schroeder inverse integral curve is calculated at the tail of each frequency band response. The starting point of the tail is taken as the moment when the peak value of the response envelope of that frequency band drops by 20 dB. The formula for calculating the Schroeder inverse integral is to integrate point by point from the end of the response to the beginning and take the square value. The time required for the curve to drop to 30 dB below the steady-state sound pressure level is taken as the reverberation time of that frequency band. The reciprocal of the reverberation time of each frequency band is taken as the frequency band reverberation decay rate. A preset threshold for the frequency band reverberation decay rate is set to 0.5 sec. The center frequency of the frequency band where the reverberation decay rate suddenly drops to the preset threshold is sequentially retrieved from the low-frequency band to the high-frequency band. This center frequency is the spatial reverberation cutoff frequency, with dimensions in T. -1 The effective sampling duration for a single audio segment is set to 30 milliseconds to accommodate the short-term stability of the audio signal. The spatial acoustic damping coefficient is a dimensionless parameter, obtained through a lookup table: Acoustic tests are conducted beforehand in a reverberation chamber on various common building envelope materials to establish a "material type - acoustic damping coefficient" mapping table. The test method involves placing an omnidirectional sound source and a pickup in the center of the reverberation chamber and measuring the change in reverberation time before and after different materials are laid. The acoustic damping coefficient is defined as the ratio of the change in reverberation time to the original reverberation time. Based on the material type of the building envelope at the test site, the spatial acoustic damping coefficient is set to 0.6 for concrete walls, 0.3 for glass curtain walls, and 0.8 for wood panels.

[0034] Based on the extracted parameters, the processor executes the reverberation attenuation feature calculation formula to solve for the spatial reverberation attenuation feature. The formula expression is:

[0035]

[0036] In the formula, The instantaneous sound pressure level of the audio signal has the following dimensions: ; The spatial reference sound pressure amplitude, with dimensions of ; The spatial reverberation cutoff frequency has the following dimensions: ; The effective sampling duration of a single audio segment, with dimensions of ; is the spatial acoustic damping coefficient, which is dimensionless; The output spatial reverberation attenuation characteristic value is dimensionless. The derivation of this formula is based on the acoustic attenuation propagation law: when an audio signal propagates in a closed space, the sound wave is reflected, absorbed, and transmitted through the spatial boundary. Each reflection is accompanied by energy loss, and the sound pressure level gradually decreases over time. The sound pressure attenuation ratio is positively correlated with the ratio of the instantaneous sound pressure to the reference sound pressure, and also follows the exponential decay law; that is, the time-domain attenuation envelope of the sound pressure can be expressed as the initial sound pressure multiplied by a natural exponential decay term. This step calculates the relative intensity of the current sound pressure level with respect to the reference sound pressure level, eliminating sound field amplitude reference deviation. This is applicable when human voice signals are present within the audio frame. Significantly greater than This value is greater than 1; when the audio frame is in a silent segment. Approaching The value of this term approaches 1. The term is the exponential decay factor, established based on the time-domain exponential decay model of the acoustic signal, where... and The product of these factors determines the cumulative effect of the decay rate. The spatial reverberation cutoff frequency represents the dominant frequency of the spatial reverberation effect. The higher this frequency, the more concentrated the reverberation energy is in the high-frequency range. The effective sampling duration of a single audio segment represents the length of the time window for reverberation; the larger the product of the two, the greater the reverberation effect. The smaller the value, the stronger the corresponding reverberation decay. As a correction factor for the overall attenuation coefficient, and The terms are multiplied to ensure that the final result reflects the overall scaling effect of spatial damping on the attenuation process. The value varies between 0 and positive infinity. A larger value indicates weaker spatial reverberation attenuation interference, meaning the reverberation energy decays faster in that spatial environment; a smaller value indicates stronger spatial reverberation attenuation interference. Three calculation examples are given below. Example 1: Pascal, Pascal, The reciprocal of Hertz Second, ,but , , Example 2: Pascal, Pascal, The reciprocal of Hertz Second, ,but , , This value is smaller than that in Example 1, indicating stronger spatial reverberation attenuation interference. Example 3: Pascal, Pascal, The reciprocal of Hertz Second, ,but , , This value is larger than the previous two examples, indicating weaker spatial reverberation attenuation interference. The three sets of examples above cover different... , , , Value combinations The output range reflects the formula's ability to distinguish between different spatial environment interference intensities.

[0037] After obtaining the spatial reverberation attenuation features, step S3 is executed. Based on the human voice rhythm modeling method, the standardized digital audio stream is analyzed to extract the human voice speech rate parameters. These parameters, combined with the spatial reverberation attenuation features, are used to calculate the human voice perturbation features. The human voice speech rate parameters include the real-time instantaneous speech rate and the standard reference speech rate corresponding to the language. The extraction method for the real-time instantaneous speech rate is as follows: Endpoint detection is performed on the standardized digital audio stream. The endpoint detection algorithm is based on a dual-threshold decision of short-time energy and short-time zero-crossing rate. The short-time energy threshold is three times the root mean square value of the background noise, and the short-time zero-crossing rate threshold is 100 times per second, segmenting syllable boundaries. The number of syllables passing through per unit time is counted, with a unit time window of 200 milliseconds, resulting in the real-time instantaneous speech rate in syllables per second, with the dimension T. -1 The standard benchmark speech rate for a specific language is a pre-statistically calculated constant: For a given language, speech rate data from at least 100 native speakers in natural conversation are collected. Each native speaker provides at least 3 minutes of continuous speech samples. The ratio of the total number of syllables to the total duration of all samples is calculated, and the statistical mean is taken as the standard benchmark speech rate for that language, with the dimension T. -1For Mandarin Chinese, the standard reference speech rate is 4.0 syllables per second; for English, it is 4.5 syllables per second; and for Japanese, it is 5.0 syllables per second. The calculation process for human voice perturbation characteristics also incorporates spatial acoustic interference parameters and human voice muscle-based vocalization linkage correction parameters, completing the joint calculation of environmental interference data and human voice movement interference data. The human vocal muscle activation linkage correction parameter is a dimensionless parameter, obtained through experimental calibration: In an anechoic chamber environment, with a background noise level not exceeding 10 dB and a free sound field range not less than 100 Hz, audio samples of test subjects reading standard texts at a standard speaking speed and in a neutral emotional state were collected. The standard texts were balanced corpora containing all vowel and consonant phonemes. The baseline value of the human vocal muscle activation linkage correction parameter was calibrated to 1.0 under these conditions. By changing the test instructions, the test subjects read at three different vocal intensities: soft voice, normal voice, and high voice. The soft voice corresponds to a sound pressure level of approximately 50 dB, and the normal voice corresponds to a sound pressure level of approximately 65 dB. A high-pitched voice corresponds to a sound pressure level of approximately 80 dB. The deviation ratio of speech rate from the baseline speech rate under different vocal intensities was measured. This deviation ratio was used as the basis for determining the correction factor, and a "vocal intensity - correction factor" mapping table was constructed. The mapping table shows that a soft voice corresponds to 0.8, a normal voice corresponds to 1.0, and a high-pitched voice corresponds to 1.2. In actual evaluation, the equivalent sound pressure level was calculated based on the short-time energy of the current audio frame to determine the vocal intensity level. An equivalent sound pressure level in the range of 40 to 55 dB was determined to be a soft voice, 55 to 70 dB was determined to be a normal voice, and above 70 dB was determined to be a high-pitched voice. The corresponding human vocal muscle linkage correction parameter values ​​were obtained by referring to the table.

[0038] Based on the above parameters, the processor executes the vocal perturbation coefficient calculation formula to solve for the vocal perturbation characteristics. The formula expression is:

[0039]

[0040] In the formula, For real-time instantaneous speech rate, the dimension is . ; The standard reference speaking speed for the corresponding language, with dimensions of ; It is a dimensionless correction factor for the coordination of vocal muscles in vocal production. The spatial reverberation attenuation characteristic value output by the first-level formula is dimensionless. This is the dimensionless coefficient representing the output vocal perturbation coefficient. The formula uses a binary superposition structure to quantify both the vocal perturbation itself and the spatial coupling perturbation with the vocal sound. The first term... Quantification of human voice self-perturbation term: This represents the proportion by which the real-time speech rate deviates from the reference speech rate. The ratio is 1 when the actual speech rate matches the reference rate, greater than 1 when the actual speech rate is faster than the reference rate, and less than 1 when it is slower than the reference rate. The squaring operation non-linearly amplifies the degree of deviation. When the value is 1, the square value is still 1. When the value is 1.3, the square value is 1.69. When the value is 0.7, the square value is 0.49, which means that the greater the deviation, the faster the contribution of the disturbance term increases; The factor corrects this deviation at the physiological level, eliminating non-emotional fluctuations in speech rate caused by tension or relaxation of vocal muscles. When the vocal force is strong, the muscles tend to be tense, and the speech rate may increase. A value greater than 1 amplifies this deviation; when vocal intensity is low, muscles tend to relax, and the speech rate may slow down. A value less than 1 reduces the deviation. (Second item) For the spatial and human voice coupling perturbation term: It is a spatial reverberation attenuation characteristic value, representing the intensity of interference from the spatial environment. The smaller the value, the stronger the spatial reverberation interference; It is a square root mapping of the speech rate variation. The square root operation makes the contribution of speech rate variation to the coupling perturbation term smoother than the square operation. This term characterizes the amplitude of the superimposed perturbation caused by the change in spatial reverberation interference with the rhythm of human voice production. The two terms use different powers. By participating in the computation, the perturbation term itself becomes more sensitive to changes in speech rate, while the coupled perturbation term responds relatively smoothly to changes in speech rate, thus distinguishing the different impact characteristics of the two types of perturbations at the computational level. Three computational examples are given below. Example 1: syllables per second, syllables per second, , , Then the first item The second item , Example 2: syllables per second, syllables per second, , , Then the first item The second item , Example 3: syllables per second, syllables per second, , , Then the first item The second item , Three examples The values ​​are 1.450, 0.6787, and 2.872, covering different ranges from low to high disturbances, as shown in Example 2. The value is very small; the contribution of the coupling perturbation term is weak, and the self-perturbation term dominates; in Example 3... The magnitude is large, the contribution of the coupled perturbation term increases, and the overall perturbation increases significantly, reflecting the formula's ability to jointly respond to different human speech rates and spatial environmental interference.

[0041] After obtaining the vocal perturbation features, step S4 is executed to extract the temporal energy parameters, frequency spectrum parameters, and Mel-spectrum basic emotional parameters corresponding to the standardized digital audio stream. These are then combined with the vocal perturbation features to calculate the fusion value of the basic emotional features. The temporal energy parameters are extracted as follows: the sum of the squares of the amplitudes at each sampling point of a frame of the standardized digital audio stream is calculated, divided by the number of sampling points within the frame, to obtain the mean temporal energy value of that frame. Dimensions are The frequency domain spectral parameters are extracted as follows: After applying a Hamming window to the frame of data, a 2048-point Fast Fourier Transform is performed. The main lobe width of the Hamming window is twice the frequency resolution, and the first side lobe attenuates to -41 dB, resulting in the frequency domain amplitude spectrum. The sum of the squares of the amplitudes of each frequency component within the amplitude spectrum is then calculated and divided by the number of frequency components to obtain the mean energy of the frequency domain spectrum. Dimensions are The extraction method for the basic sentiment parameters of the Mel spectrum is as follows: the linear spectrum of the frame data is filtered in the frequency domain through a set of triangular Mel filter banks. The Mel filter bank contains 40 triangular filters, with the center frequencies of adjacent filters overlapping by 50%. The center frequencies are evenly distributed on the Mel scale, and the logarithm of the output energy of each filter constitutes the Mel spectrum feature vector. This 40-dimensional feature vector is input into a pre-trained AI basic sentiment model, which is a three-layer fully connected neural network. The input layer has 40 neurons, the hidden layer has 64 neurons using rectified linear unit activation functions, and the output layer has 4 neurons using flexible maximum transfer functions, corresponding to four emotional states: happiness, sadness, anger, and neutrality. The model outputs a dimensionless raw output value of the AI ​​basic sentiment model. , The model selects the maximum probability value of flexibility among the four emotional states; simultaneously, the fully connected layers within the model output the weighted coefficients of the Mel-spectral features. , is the root mean square value of the weights corresponding to each dimension of the Mel spectrum eigenvector in the weight matrix from the hidden layer to the output layer. This coefficient is dimensionless.

[0042] Based on the above parameters, the processor executes the emotion feature fusion formula to calculate the basic emotion feature fusion value. The formula expression is:

[0043]

[0044] In the formula, The time-domain energy mean, with dimensions of ; The average energy of the frequency domain spectrum, with dimensions of ; These are the weighting coefficients for the characteristics of the Mel spectrum, and are dimensionless. The perturbation coefficient for human voice is dimensionless. The intermediate values ​​are fused to the basic sentiment features of the output, with the following dimensions: The formula consists of a product of two parts. The first part... Basic sentiment feature fusion item: It represents the global energy accumulation intensity of an audio signal in the time domain; the greater the energy, the higher the degree of emotion it usually corresponds to. Characterizing the energy distribution intensity of audio signals in the frequency domain, the degree of concentration of spectral energy in a specific frequency band is related to the emotion type; The coefficients are adjusted according to the emotional feature weighting rules of the Mel spectrum. In the fusion process, when the AI ​​basic emotion model determines that the current audio is highly correlated with a certain type of emotion... Increasing the value increases the contribution of frequency domain information to the fusion term, allowing the fusion result to take into account both time-domain global energy information and frequency-domain spectral energy distribution information. (The latter part...) For the perturbation correction term: hyperbolic tangent function The domain is all real numbers, and the range is Its functional expression is the ratio of hyperbolic sine to hyperbolic cosine. The function graph is symmetric about the origin, approximately linear near the origin, and asymptotically saturates at ±1 far from the origin. For any input perturbation coefficient... , Output a bounded value. hour , hour , hour , hour The higher the disturbance intensity, The closer it is to 1, The closer it is to 0, the stronger the attenuation effect on the fusion term; when the perturbation strength approaches 0, Approaching 0, the correction term approaches 1, and the fusion term is almost not attenuated. This operation allows the correction amplitude to adaptively adjust with the disturbance intensity, and the correction curve is smooth and without discontinuities. Three calculation examples are given below. Example 1: watt, watt, , ,but watt, , , Watt. Example 2: watt, watt, , ,but watt, , , Watt. Example 3: watt, watt, , ,but watt, , , Watt. Example 2 out of three. Minimum value, after correction The value that is retained the most is Watt; Example 3 Maximum value, fusion item The wattage is greater than the previous two examples, but after correction... The value is only The significant attenuation of watts indicates that the proportion of sentiment features retained under highly perturbed environments is extremely low, in order to reduce the interference of unreliable sentiment features on subsequent calculations.

[0045] After completing the basic emotional feature fusion, step S5 is executed, which calls the preset dynamic weighting factor and dimensional normalization coefficient to fuse the spatial reverberation attenuation feature, vocal perturbation feature, and basic emotional feature fusion value, and calculates the audio emotional assessment confidence score. The preset dynamic weighting factor includes the environmental interference dynamic weighting factor. and dynamic weighting factor of vocal perturbation Both are dimensionless parameters. The value of and Numerical association: The range of values ​​for is divided into five intervals according to a logarithmic scale, when When less than 0.05 , When it is between 0.05 and 0.15 , When it is between 0.15 and 0.4 , When it is between 0.4 and 0.8 , When greater than 0.8 , When it increases This is subsequently reduced to weaken the proportion of environmental interference terms in the calculation. The value of and Numerical association: The range of values ​​is linearly divided into four intervals, when When less than 1 , When it is between 1 and 2 , Between 2 and 3 , When greater than 3 ,when When it increases This is subsequently reduced to decrease the weight of the perturbation correction correlation term in the molecule under high perturbations. Dimensional normalization conversion coefficient. The dimensions are Its value is based on and The typical magnitude is determined, and the typical magnitude is: Watts magnitude, take Watt's reciprocal makes The product falls into the and , Comparable dimensionless numerical ranges.

[0046] Based on the above parameters and the feature data output from the preceding steps, the processor executes the comprehensive calculation formula for sentiment assessment confidence. The formula expression is:

[0047]

[0048] In the formula, The original output value of the AI ​​basic sentiment model is dimensionless. The environmental disturbance dynamic weighting factor is dimensionless. is a spatial reverberation attenuation characteristic value, dimensionless; The dynamic weighting factor for the sound perturbation is dimensionless. These are the dimensionless normalization conversion coefficients, with dimensions of ; The intermediate value for the fusion of basic sentiment features, with dimensions of ; The perturbation coefficient for human voice is dimensionless. The confidence level for the output audio emotion assessment is dimensionless. The numerator of this formula contains two terms. The first term... To correct for environmental interference in the sentiment data, the original output value of the AI-based sentiment model is multiplied by the dynamic weights of environmental interference and the spatial reverberation attenuation characteristics. The original probability of the model's judgment of the sentiment type of the current audio frame. and Jointly adjust the contribution of this item: when Larger spatiotemporal disturbances are weaker. Take a smaller value to decrease the proportion of this term in the numerator; when Smaller spatial interference means stronger spatial interference. A larger value is chosen to increase the proportion of this term in the molecule, allowing for greater reliance on environmental interference dimensional information for calibration under strong spatial interference. The second term... To correct for human voice perturbation in sentiment data, coefficients are transformed using dimensional normalization. Dimensional physical quantities Convert to a dimensionless value, then add it to the dynamic weight of the sound perturbation. Perform the product. according to The level is dynamically adjusted. The larger the disturbance, the stronger it is. The smaller the value, the lower the proportion of this term in the numerator, thus avoiding unreliability under high perturbations. The value has an excessive impact on the confidence score calculation. (Denominator) For the interference normalization constraint term: This represents the combined effect of environmental interference weight, vocal perturbation weight, and human vocal perturbation coefficient. , , When any value in the numerator increases, this term increases, the denominator increases as a whole, and the numerator is compressed as a whole; adding 1 ensures that this term is greater than 1 even under high interference, avoiding the denominator being equal to 0 or too small, which would lead to... Value overflow. The sum of the two terms in the numerator represents the strength of evidence positively supporting the current sentiment judgment, and the denominator represents the penalty coefficient for interference strength. Dividing the two yields the final confidence score. Three calculation examples are given below. Example 1: , , , , Watt's reciprocal watt, Then the first term of the numerator The second term in the numerator numerator total ; denominator ; Example 2: , , , , Watt's reciprocal watt, Then the first term of the numerator The second term in the numerator numerator total ; denominator ; Example 3: , , , , Watt's reciprocal watt, Then the first term of the numerator The second term in the numerator numerator total ; denominator ; This value exceeds 1, which is a constraint in practical applications. The upper limit is 1, that is, take Three examples The values ​​are 0.527, 0.162, and 1.0 respectively. Example 2 shows weak spatial interference but extremely strong vocal disturbance. Significantly reduced, Although the value is high, the confidence level is low. The value is only 0.162, reflecting the negative impact of voice disturbance on emotion judgment; in Example 3, spatial interference is strong but voice disturbance is weak. The retention rate is relatively high, the second term in the molecule contributes significantly, and the confidence level is high. Reaching the upper limit of 1.0 reflects a high degree of credibility in sentiment assessment under controllable perturbations.

[0049] All computational parameters in this method are uniquely identified, each with a fixed physical meaning and dimension, and all parameters follow dimensionless homogeneous computation rules. Taking the spatial reverberation attenuation characteristic calculation formula as an example, the parameters... and Both are units of sound pressure The ratio of the two values ​​is dimensionless; parameters Dimensions With parameters Dimensions The product has a dimension of 1, so exponentiation is legal; finally Dimensionless. Taking the formula for calculating the acoustic perturbation coefficient as an example, the parameters... and Both are units of frequency The ratio eliminates the dimension, and the square and root operations both apply to the dimensionless ratio, without introducing a dimension; and All are dimensionless, and ultimately Dimensionless. Taking the formula for fusing basic emotional features as an example, and The dimensions of all are The addition operator must meet the requirement of homogeneity of dimensions; Dimensionless, after product Maintain the original dimensions. Taking the comprehensive calculation formula for the confidence level of sentiment assessment as an example, Dimensions and Dimensions The product is dimensionless; both terms in the numerator are dimensionless; all terms in the denominator are dimensionless; the final product is dimensionless. Dimensionless. The dimensional transfer relationship of the four-layer operation formula is progressive, and the dimension of the output quantity of each layer matches the dimension of the input requirement of the next layer. The dimensional transformation path from the bottom acoustic physical quantity to the top dimensionless confidence is completely closed.

[0050] This method employs a fixed fourth-order operational timing logic to form a unidirectional serial data processing link. In the first order, the processor completes the spatial reverberation attenuation characteristic calculation and outputs... The value is latched into the first output register. In the second stage, the processor reads the value from the first output register. Value, combined with extracted human voice speech rate parameters Preset and table lookup Complete the calculation of human voice perturbation features and output. The value is latched into the second output register. In the third stage, the processor reads the value from the second output register. Value, combined with extracted energy parameters , and weighting coefficients Complete the basic sentiment feature fusion operation and output. The value is latched into the third output register. In the fourth stage, the processor reads the values ​​from the first, second, and third output registers. , , Value, call the preset , , and Complete the comprehensive calculation of the confidence level of the sentiment assessment and output the results. The output data of the preceding operation flows unidirectionally into the input port of the subsequent operation. Data synchronization between operations of different orders is ensured by register latching. The subsequent operation is only triggered after the previous operation is completed and the result is written to the register. There is no feedback path for subsequent data to be sent back to the preceding operation unit, nor is there a bypass channel for directly calling subsequent input data across a certain order.

[0051] After obtaining the audio emotion assessment confidence score, this method writes the standardized digital audio stream, all computational parameters, intermediate calculation results of the fourth-order operation, and the audio emotion assessment confidence score into an embedded storage medium. The embedded storage medium is a non-volatile memory connected to the processor's external bus, specifically a NAND flash memory chip with a capacity of at least 32 gigabytes, connected to the chip select pin of the serial peripheral interface bus. The storage format uses structured data blocks: the block header occupies 128 bytes, recording the timestamp, audio stream length, and parameter identifier code for this assessment. The timestamp is the Coordinated Universal Time (UTC) read from the real-time clock chip; the audio stream length is the total number of bytes in the standardized digital audio stream; and the parameter identifier code is a 16-byte universally unique identifier used to uniquely identify the parameter configuration version used in this assessment. The block body sequentially stores the binary data of the standardized digital audio stream, the extracted values ​​of five types of spatial acoustic parameters, and the parameters... Calculated values, parameters and and Values ​​and parameters Calculated values, parameters and and Values ​​and parameters Calculated values, parameters and and and Values ​​and parameters The final value is determined by separating each data segment with a fixed delimiter; a 32-bit cyclic redundancy check (CRC) is set at the end of the block, checking all bytes from the beginning of the block header to the CRC checksum. Simultaneously, the processor outputs audio emotion assessment confidence data via a communication interface, which is a universal asynchronous receiver / transmitter (UART) with a baud rate of 115200. The data frame format includes a start bit, 8 data bits, an even parity bit, and a stop bit. The output data frame includes an assessment identifier and... The numerical value and the evaluation identifier code are consistent with the parameter identifier code in the storage block header.

[0052] An AI voice evaluation system for strong interference scenarios is applied to the aforementioned AI audio emotion evaluation method. The system includes an audio signal acquisition module, a spatial acoustic feature analysis module, a human voice production feature analysis module, an AI emotion basic feature extraction module, a multi-feature operation module, and a data storage and result output module.

[0053] The audio signal acquisition module acquires the analog audio signal of the target human voice in the environment, performs the analog-to-digital conversion, outputs a standardized digital audio stream, and synchronously distributes it to other modules. The hardware of this module includes an electret condenser microphone, a signal conditioning circuit, and a distribution bus interface. The microphone converts sound pressure changes into an analog voltage signal, with a frequency response flatness of no more than ±2 dB within the 20 Hz to 20 kHz range. The signal conditioning circuit performs noise reduction sampling, signal amplitude calibration, and timing warping on the analog voltage signal. The core of the signal conditioning circuit is a field-programmable gate array (FPGA), and its internal logic units implement spectral subtraction noise reduction, amplitude calibration multipliers, and a dynamic time warping state machine, outputting a digital audio timing signal. The distribution bus interface integrates a direct memory access controller with a bus width of 32 bits and a clock frequency of 100 MHz. When the digital audio timing signal is generated, a bus request is triggered. After authorization by the bus arbiter, the data bus is occupied, and a frame of standardized digital audio stream data is simultaneously written to the designated address space of the dual-port random access memory of the spatial acoustic feature analysis module, the human voice feature analysis module, and the AI ​​emotion basic feature extraction module. The address space is pre-allocated as a contiguous memory region by the system initialization program.

[0054] The spatial acoustic feature analysis module is electrically connected to the audio signal acquisition module. It receives a standardized digital audio stream, extracts spatial acoustic parameters, and calculates the reverberation attenuation characteristic output value. The core hardware of this module is a digital signal processor (DSP), specifically a 32-bit fixed-point DSP with a 400 MHz operating frequency. Its internal firmware includes subroutines for extracting instantaneous audio sound pressure level amplitude, reading spatial reference sound pressure level amplitude, extracting spatial reverberation cutoff frequency, configuring a single-segment audio effective sampling duration register, and looking up a table for spatial acoustic damping coefficients. The processor reads a frame of standardized digital audio stream data from a dual-port random access memory (DRAM), executes the subroutines sequentially according to the instructions, and extracts… , , , , Substituting the five types of parameters into the reverberation attenuation characteristic calculation formula, we obtain... value. The value is temporarily stored in the output register of this module. The output register is in 32-bit floating-point format and is sent to the human voice feature analysis module and the multi-feature operation module through the inter-module serial peripheral interface bus. The serial peripheral interface is configured in master mode, with clock polarity and phase in mode 0, and the data frame length is 32 bits.

[0055] The human voice feature analysis module is electrically connected to the audio signal acquisition module and the spatial acoustic feature analysis module. It receives the standardized digital audio stream and reverberation attenuation feature output values, and calculates the human voice perturbation coefficient output value. The core hardware of this module is a digital signal processor (DSP), the same model as the spatial acoustic feature analysis module. Its internal firmware includes a real-time instantaneous speech rate extraction subroutine, a standard reference speech rate memory, and a lookup subroutine for muscle-based vocalization linkage correction factors. This processor reads the standardized digital audio stream from dual-port random access memory, performs syllable segmentation and speech rate calculation, and obtains the human voice perturbation coefficient output value. Read the standard reference speaking speed of the corresponding language from the internal read-only memory. The read-only memory stores a pre-programmed multilingual baseline speech rate constant table, indexed according to the language identifier code configured during system initialization. Value; the correction factor is read from memory based on the short-time energy judgment result of the current audio frame. Received via serial peripheral interface Value; Substitute into the formula for calculating the sound disturbance coefficient to obtain value. The value is temporarily stored in the output register of this module and then sent to the multi-feature operation module.

[0056] The AI-based emotion feature extraction module is electrically connected to the audio signal acquisition module. It receives a standardized digital audio stream and extracts the basic emotion parameters corresponding to the time-domain energy, frequency-domain spectrum, and Mel spectrum. The core hardware of this module is a digital signal processor (DSP), whose internal firmware includes subroutines for time-domain energy calculation, fast Fourier transform, Mel filter bank, and a pre-trained AI-based emotion model. The processor reads the standardized digital audio stream from a dual-port random access memory (DRAM) and calculates the mean time-domain energy. Perform a Fast Fourier Transform and calculate the mean energy of the frequency domain spectrum. Perform Mel filter bank filtering and logarithmic operation to obtain Mel spectrum feature vectors; input the feature vectors into the AI ​​basic sentiment model, and the model inference outputs... and . , , , The value is stored in the output register of this module and sent to the multi-feature operation module.

[0057] The multi-feature processing module is electrically connected to the spatial acoustic feature analysis module, the human voice feature analysis module, and the AI ​​emotion basic feature extraction module, respectively. It executes a fourth-order step-by-step serial substitution operation logic to complete multi-feature data processing and audio emotion assessment confidence calculation. The hardware core of this module is a central processing unit, specifically a 64-bit processor with an 800 MHz clock speed. The internal firmware implements a fourth-order operational timing state machine. The state machine transitions sequentially through four states: State S1 receives... Latched into an internal register; status S2 receives , , and The calculation yields And latch; Status S3 receives , , and The calculation yields And latch; status S4 received , , , , , , The calculation yields The state transition clock of the state machine is obtained by dividing the system clock by 16, and each state is allocated 500 clock cycles to complete the floating-point multiplication, addition, and exponentiation operations of the corresponding formula. , , The basis for the value and The quantization level is obtained from an internal lookup table, which is stored in the read-only memory on the CPU chip. The value is written to the output register of this module and sent to the data storage and result output module.

[0058] The data storage and output module is electrically connected to the multi-feature computation module, storing the algorithm program, model parameters, and evaluation data, and outputting audio emotion evaluation results. This module includes non-volatile memory and a communication interface. The non-volatile memory is NAND flash memory with a capacity of 128 gigabytes, divided into three storage areas: the first storage area occupies 16 gigabytes and stores the executable firmware programs of this module and the aforementioned modules' digital signal processors and central processing units, stored in binary image file format; the second storage area occupies 4 gigabytes and stores the weight parameters of the AI ​​basic emotion model, lookup table data, and preset constants, with the weight parameters arranged sequentially in single-precision floating-point format; the third storage area occupies 108 gigabytes and stores the standardized digital audio streams from each evaluation, all computational parameters, intermediate calculation results at each level, and the final confidence score in a circular queue. The data blocks are updated in a queue, with new data blocks overwriting the oldest ones when the queue is full. The communication interface is a universal asynchronous transmitter, configured with a baud rate of 115200, 8 data bits, 1 stop bit, and no parity. When a new data block is written to the third memory area... When the value is reached, the communication interface automatically reads it. The values ​​are then encapsulated into output data frames and sent out via physical links.

[0059] A computer-readable storage medium stores computer instructions that, when executed by a processor, implement the steps of the aforementioned AI audio emotion assessment method. This computer-readable storage medium is a storage partition within a solid-state drive (SSD), with the partition format being a File Allocation Table 32 file system, storing machine code sequences in binary file format. When a computer device equipped with this SSD powers on and loads the machine code sequence into memory, the processor sequentially executes: an instruction segment reading an analog audio signal from a microphone interface and converting it into a standardized digital audio stream; an instruction segment extracting five types of spatial acoustic parameters from the standardized digital audio stream and substituting them into a reverberation attenuation characteristic calculation formula; an instruction segment extracting human speech rate parameters from the standardized digital audio stream and substituting them into a vocal perturbation coefficient calculation formula; an instruction segment extracting time-domain energy, frequency-domain spectrum, and Mel-spectrum basic emotion parameters from the standardized digital audio stream and substituting them into an emotion basic feature fusion formula; an instruction segment substituting them into an emotion assessment confidence comprehensive calculation formula; and an instruction segment writing the entire process data to an embedded storage medium and sending the confidence value externally. The entire instruction segment covers the complete process from audio signal input to confidence result output. The order of formula operations and parameter dependencies are consistent with the steps described above. The instruction segment contains a sequence of assembly instructions for loading formula parameters from memory addresses, performing multiplication, addition, and exponentiation operations in the floating-point unit, and writing the results back to memory addresses. The calculation of each formula is completed in the processor's floating-point unit, and intermediate results are temporarily stored in the processor's general-purpose registers in single-precision floating-point format.

[0060] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-powered audio emotion testing method, characterized in that, Executed by the processor of a computer device, the following steps are included: S1. Collect the analog audio signal corresponding to the target human voice and convert the analog audio signal into a standardized digital audio stream; S2. Based on architectural acoustic modeling, the standardized digital audio stream is analyzed, spatial acoustic parameters are extracted, and spatial reverberation attenuation characteristics are calculated. S3. Based on the human voice rhythm modeling method, the standardized digital audio stream is analyzed, the human voice speech rate parameter is extracted, and the human voice perturbation feature is calculated by combining the spatial reverberation attenuation feature. S4. Extract the time-domain energy parameters, frequency-domain spectrum parameters, and Mel spectrum basic emotional parameters corresponding to the standardized digital audio stream, and calculate the basic emotional feature fusion value by combining the human voice perturbation features. S5. Call the preset dynamic weighting factor and dimensional normalization coefficient, and fuse the spatial reverberation attenuation feature, the human voice perturbation feature and the basic emotional feature fusion value to calculate the audio emotional assessment confidence level.

2. The AI ​​audio emotion testing method according to claim 1, characterized in that, The acquisition and conversion of the standardized digital audio stream described in S1 includes noise reduction sampling, signal amplitude calibration and timing regularization of the original analog audio signal, outputting a digital audio timing signal with continuous timing and uniform amplitude, and synchronously distributing the digital audio timing signal to the corresponding parameter parsing and calculation unit.

3. The AI ​​audio emotion testing method according to claim 1, characterized in that, The spatial acoustic parameters mentioned in S2 include the instantaneous sound pressure amplitude of the audio, the spatial reference sound pressure amplitude, the spatial reverberation cutoff frequency, the effective sampling duration of a single audio segment, and the spatial acoustic damping coefficient. The instantaneous sound pressure amplitude of the audio, the spatial reference sound pressure amplitude, the spatial reverberation cutoff frequency, the effective sampling duration of a single audio segment, and the spatial acoustic damping coefficient are all time-domain acoustic physical parameters extracted in real time from the standardized digital audio stream.

4. The AI ​​audio emotion testing method according to claim 1, characterized in that, The speech rate parameter in S3 includes the real-time instantaneous speech rate and the standard reference speech rate corresponding to the language. The calculation process of the speech perturbation features introduces spatial acoustic interference parameters and human voice muscle vocalization linkage correction parameters to complete the joint calculation of environmental interference data and human voice movement interference data.

5. The AI ​​audio emotion testing method according to claim 1, characterized in that, This method configures uniquely identified operational parameters, each with a fixed physical meaning and dimension. All operational parameters follow dimensionless homogeneous operational rules and are used for step-by-step calculation of multi-dimensional audio features and unified calculation of emotion assessment results.

6. The AI ​​audio emotion testing method according to claim 5, characterized in that, This method is configured with a fixed fourth-order operation timing logic, which sequentially completes the spatial reverberation attenuation feature operation, the human voice perturbation feature operation, the basic emotional feature fusion operation, and the comprehensive operation of emotional assessment confidence. The output data of the preceding operation is used as the input data of the subsequent operation, forming a unidirectional serial data operation link.

7. The AI ​​audio emotion testing method according to claim 5, characterized in that, In the basic emotional feature fusion operation of this method, the human voice perturbation feature data is numerically transformed using the hyperbolic tangent function. The transformed values ​​are then used in the fusion calculation of time-domain energy parameters and frequency-domain spectrum parameters.

8. The AI ​​audio emotion testing method according to claim 1, characterized in that, After obtaining the audio emotion assessment confidence score through S5 computation, the standardized digital audio stream, all computational parameters, intermediate calculation results of the fourth-order computation, and the audio emotion assessment confidence score are written into the embedded storage medium, while the audio emotion assessment confidence score data is output externally.

9. A voice assessment system, characterized in that, The AI ​​audio emotion assessment method according to any one of claims 1-8 includes an audio signal acquisition module, a spatial acoustic feature analysis module, a human voice production feature analysis module, an AI emotion basic feature extraction module, a multi-feature operation module, and a data storage and result output module. The audio signal acquisition module is used to acquire the target human voice audio analog signal in the environment, complete the conversion between analog and digital signals, output a standardized digital audio stream and distribute it synchronously to the other modules. The spatial acoustic feature analysis module is electrically connected to the audio signal acquisition module and is used to receive standardized digital audio streams, extract spatial acoustic parameters, and calculate reverberation attenuation feature output values. The human voice production feature analysis module is electrically connected to the audio signal acquisition module and the spatial acoustic feature analysis module. It is used to receive the standardized digital audio stream and the reverberation attenuation feature output value, and calculate the human voice production disturbance coefficient output value. The AI ​​emotion basic feature extraction module is electrically connected to the audio signal acquisition module and is used to receive standardized digital audio streams and extract basic emotion parameters corresponding to time-domain energy, frequency-domain spectrum and Mel spectrum. The multi-feature operation module is electrically connected to the spatial acoustic feature analysis module, the human voice production feature analysis module, and the AI ​​emotion basic feature extraction module, respectively, and is used to execute the fourth-order step-by-step serial substitution operation logic to complete the multi-feature data operation and audio emotion assessment confidence calculation. The data storage and result output module is electrically connected to the multi-feature operation module and is used to store algorithm programs, model parameters and evaluation data, and output audio emotion evaluation result data.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the steps of the AI ​​audio emotion testing method according to any one of claims 1-8.

Images