Guqin playing audio feature extraction and tone quality evaluation method

Abstract

The application relates to the technical field of audio signal processing, and discloses a guzheng performance audio feature extraction and tone quality evaluation method, which comprises the following steps: receiving a discrete-time audio sequence signal and calculating a first-order differential sequence of logarithmic energy; calling an a priori pitch harmonic distribution matrix to compare and judge the differential sequence by using an adaptive threshold to determine a physical starting transient starting point; rewriting a framing step length according to the physical starting transient starting point, cutting an audio section into a transient pulse subframe and a decay steady-state subframe; extracting a high-dimensional differential mel-frequency cepstrum coefficient vector from the transient pulse subframe, and searching for state transition in a Markov tone quality evaluation state transition matrix to output a tone quality evaluation score vector; and the application uses event-driven adaptive framing to realize orthogonal isolation representation of the transient state and the steady state, eliminates time-frequency resolution constraints, and improves feature recognition under dense plucking working conditions.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, and more specifically, to a method for extracting audio features of guzheng performance and evaluating its sound quality. Background Technology

[0002] Currently, feature extraction of guzheng performance audio typically employs short-time Fourier transform with a fixed time window or extracts Mel-frequency cepstral coefficients at a fixed order. These methods assume that the input audio is approximately stationary within the fixed time window, and then calculate high-dimensional acoustic features for subsequent analysis and evaluation based on this assumption. However, in the process of feature extraction and sound quality evaluation of guzheng performance audio, the guzheng strings generate significant transient impact energy at the moment of plucking; during playing actions such as vibrato, pressing, glissando, and trills, the fundamental frequency changes continuously and exhibits complex nonlinear drift. Conventional equal-step framing methods cannot simultaneously accommodate both types of information. If the time window is shortened to more accurately capture the plucking start point, the frequency resolution during the long-tail decay stage and trill process will decrease; if the time window is extended to obtain clearer string harmonic spectral lines, the long time window will smooth the energy edges at the moment of plucking. Thus, a contradiction arises between time-domain resolution and frequency-domain resolution.

[0003] To address the above issues, common improvements include increasing the signal sampling rate or adding additional hardware sensors to track the physical movement of the strings. However, these methods increase system hardware costs and computational power consumption. In addition to the limitations of hardware improvements such as increasing the sampling rate or adding external sensors, existing technologies also have shortcomings from a purely software perspective, such as evaluation and control methods. For example, Chinese invention patent CN113782059B discloses a method and device for evaluating musical instrument audio and a non-transient storage medium. It uses an acoustic model to forcibly align the audio to be evaluated with a reference score and determines the score based on the average value and the maximum element value of the acoustic likelihood vector.

[0004] Therefore, when dense strumming causes strong harmonic aliasing and violent string pressing causes significant fundamental frequency drift, without prior constraints related to the pitch and harmonic distribution of the guzheng strings, relying solely on data-driven processing or hardware measurement is still insufficient to eliminate parasitic interference components caused by aliasing echoes. Furthermore, the increased data dimensionality may lead to overfitting of the eigenvectors, making it difficult for the state transition matrix used for subsequent sound quality evaluation to converge stably. How to overcome the limitations of fixed window length and stationarity assumptions, while maintaining a conventional single-channel sensing link and standard sampling rate, adaptively adjust the time step of audio sequence frame calculation, separate the plucking transient and attenuation steady state, and reduce feature distortion caused by multi-string resonance aliasing and fundamental frequency drift, becomes the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention provides a method for extracting audio features of guzheng performances and evaluating their sound quality, comprising the following steps:

[0006] Step S101: Receive the discrete-time audio sequence signal and calculate the first-order difference sequence of the logarithmic energy of the discrete-time audio sequence signal using the short-time logarithmic energy difference rule;

[0007] Step S102: Retrieve the solidified prior pitch harmonic distribution matrix and perform adaptive threshold comparison judgment on the logarithmic energy first-order difference sequence based on the prior pitch harmonic distribution matrix. When the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive start threshold spontaneously generated by the background noise base signal, determine the physical starting transient start point of string plucking.

[0008] Step S103: Rewrite the timing step of the audio frame extraction process according to the physical start transient starting point, and independently divide the audio segment within 15ms after the physical start transient starting point into transient pulse subframes, and divide the audio segment from the subsequent point to the energy decay boundary point into decay steady state subframes.

[0009] Step S104: Call the first-order cepstral transform operator on the transient pulse subframe to extract the high-dimensional differential Mel cepstral coefficient vector, and perform the guzheng sound quality evaluation state transition search in the preset Markov sound quality evaluation state transition matrix based on the high-dimensional differential Mel cepstral coefficient vector to output the sound quality evaluation score vector.

[0010] Preferably, step S104 further includes the following sub-steps: Step S1041, detecting the energy attenuation slope of the attenuation steady-state subframe, and triggering a short-time energy attenuation compensation rule to calculate the dimensionless gain compensation factor when the absolute value of the energy attenuation slope is lower than the steady-state empirical threshold; Step S1042, applying in-situ weighted penalty to the current frame cepstral matrix using the dimensionless gain compensation factor to truncate the parasitic interference components introduced by aliasing echoes and improve the energy resolution and signal-to-noise ratio of the high-dimensional differential Mel-Cepstral coefficient vector in the frequency domain.

[0011] Preferably, step S101 includes the following sub-steps: Step S1011, performing sliding windowing processing on the received discrete-time audio sequence signal to divide it into overlapping short-time audio frame sequences; Step S1012, calculating the sum of squares of transient energy of the short-time audio frames in the short-time audio frame sequence, and performing a logarithmic transformation on the sum of squares of transient energy to generate a short-time logarithmic energy time-series signal; Step S1013, performing first-order difference calculation on the short-time logarithmic energy time-series signal using the short-time logarithmic energy difference rule, thereby outputting a logarithmic energy first-order difference sequence.

[0012] Preferably, step S102 includes the following sub-steps: Step S1021, retrieve the a priori pitch harmonic distribution matrix, which contains the fundamental frequency distribution and characteristic harmonic component distribution of the 21 strings of the guzheng under the standard string order; Step S1022, perform joint spatial constraint mapping on the logarithmic energy first-order difference sequence and the a priori pitch harmonic distribution matrix to remove the background harmonic aliasing components generated by multi-string resonance; Step S1023, after removing the background harmonic aliasing components, when the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive initiation threshold, confirm the current energy mutation position as the physical initiation transient starting point.

[0013] Preferably, step S103 includes the following sub-steps: Step S1031, in response to the determined physical start transient origin, the constant time step of the conventional audio framing process is rewritten into an adaptive non-uniform step triggered by the physical start transient origin; Step S1032, an audio segment with a length of 15ms starting from the physical start transient origin is extracted, and a narrow window function with high time domain resolution is applied to the audio segment to truncate it, so as to independently construct transient pulse subframes; Step S1033, the energy decay termination point after the physical start transient origin is located, and the signal between the audio trail after 15ms and the energy decay termination point is independently divided into decay steady-state subframes.

[0014] Preferably, step S104 includes the following sub-steps: Step S1043, performing a discrete Fourier transform on the transient pulse subframe to convert the time-domain discrete audio signal to the frequency domain, and calculating the transient power spectrum matrix; Step S1044, performing energy-weighted summation on the transient power spectrum matrix through a set of triangular filter banks arranged according to Mel frequency standards to obtain the logarithmic energy spectrum vector of the Mel filter banks; Step S1045, applying a first-order cepstral transform operator to the logarithmic energy spectrum vector of the Mel filter banks to perform a discrete cosine transform, thereby extracting the high-dimensional difference Mel cepstral coefficient vector.

[0015] Preferably, step S104 includes the following sub-steps: Step S1046, locally mapping the extracted high-dimensional difference Mel-Cepstral coefficient vector to the acoustic sound quality evaluation hidden Markov model to activate the sound quality evaluation state space; Step S1047, calculating the hidden state sequence transition probability between adjacent frames using the Markov sound quality evaluation state transition matrix, and searching for the maximum probability convergence path through dynamic programming; Step S1048, performing multi-dimensional weighted scoring based on the likelihood probability value corresponding to the maximum probability convergence path, thereby generating and outputting the sound quality evaluation score vector.

[0016] Preferably, step S102 includes the following sub-steps: step S1024, when receiving discrete-time audio sequence signals, extracting the background noise floor signal of the preceding silent audio segment; step S1025, calculating the variance envelope time-series curve of the background noise floor signal, and using the maximum value of the variance envelope time-series curve as the reference gain boundary of the adaptive start-up threshold, thereby spontaneously generating the adaptive start-up threshold.

[0017] Preferably, step S101 includes the following sub-steps: step S1014, before receiving the discrete-time audio sequence signal, a bandpass filter is used to perform high-frequency anti-aliasing filtering and low-frequency DC component filtering on the one-dimensional original discrete audio waveform stream; step S1015, the filtered audio waveform stream is input to the analog-to-digital conversion unit to perform time-domain discretization sampling, thereby generating a discrete-time audio sequence signal that satisfies the time continuity constraint.

[0018] The present invention has at least the following beneficial effects:

[0019] 1. In the extraction of audio features from Guzheng performance, by calculating the logarithmic energy first-order difference sequence of discrete-time audio sequence signals and combining it with the prior pitch harmonic distribution matrix to determine the physical starting point of transient response, this invention can segment the audio according to the actual starting point of Guzheng plucking. After the transient pulse subframe formed at the moment of plucking and the attenuated steady-state subframe formed by the subsequent aftersound are processed separately, the smoothing effect of the fixed time window on the plucking impact is weakened, and the frequency details in the long-tail attenuation and glissando process are not easily mixed up in the same time window. As a result, the transient timbre features produced by right-hand plucking and the frequency change features produced by left-hand glissando and vibrato can be preserved separately, and the audio features reflect the details of Guzheng performance more clearly.

[0020] 2. The prior pitch harmonic distribution matrix contains the fundamental frequency distribution and harmonic components of the guzheng strings under the standard string sequence. When determining the physical starting point of the transient, this invention uses this matrix to constrain the first-order difference sequence of logarithmic energy, so that the background harmonics formed by multi-string resonance are not easily mistaken for the new plucking starting point. In this way, when there is dense strumming or strong background harmonics, the determination of the starting position no longer relies solely on increasing the sampling rate or adding external sensors, which can reduce the interference of multi-string resonance on the starting point positioning and improve the accuracy of the plucking transient segmentation.

[0021] 3. To address the issue of overlapping aftertones of the previous string and the starting pulse of the next string during dense string plucking, this invention determines whether the aftertone still affects the current frame by attenuating the energy attenuation slope of the steady-state subframe. When the conditions are met, a dimensionless gain compensation factor is used to weight the cepstral matrix of the current frame. This reduces the interference components caused by aliased aftertones, reduces the contamination of subsequent plucking features by preceding aftertones, and makes the effective timbre information in the high-dimensional differential Mel cepstral coefficient vector more prominent, thereby improving the stability and signal-to-noise ratio of feature extraction in dense plucking scenarios. Attached Figure Description

[0022] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings, in which several embodiments of the invention are illustrated by way of example and not limitation, wherein:

[0023] Figure 1 This is a flowchart of the adaptive frame segmentation and sound quality state transition search process for the performance audio of this invention;

[0024] Figure 2 This is a structural diagram of the acoustic signal acquisition and digital processing system of the present invention. Detailed Implementation

[0025] The principles and spirit of the present invention will now be described with reference to several exemplary embodiments in conjunction with the accompanying drawings. It should be understood that these embodiments are provided merely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0026] A method for extracting audio features of guzheng performance and evaluating its sound quality includes the following steps:

[0027] Step S101: Receive the discrete-time audio sequence signal and calculate the first-order difference sequence of the logarithmic energy of the discrete-time audio sequence signal using the short-time logarithmic energy difference rule;

[0028] Step S102: Retrieve the solidified prior pitch harmonic distribution matrix and perform adaptive threshold comparison judgment on the logarithmic energy first-order difference sequence based on the prior pitch harmonic distribution matrix. When the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive start threshold spontaneously generated by the background noise base signal, determine the physical starting transient start point of string plucking.

[0029] Step S103: Rewrite the timing step of the audio frame extraction process according to the physical start transient starting point, and independently divide the audio segment within 15ms after the physical start transient starting point into transient pulse subframes, and divide the audio segment from the subsequent point to the energy decay boundary point into decay steady state subframes.

[0030] Step S104: Call the first-order cepstral transform operator on the transient pulse subframe to extract the high-dimensional differential Mel cepstral coefficient vector, and perform the guzheng sound quality evaluation state transition search in the preset Markov sound quality evaluation state transition matrix based on the high-dimensional differential Mel cepstral coefficient vector to output the sound quality evaluation score vector.

[0031] Preferably, step S104 further includes the following sub-steps: Step S1041, detecting the energy attenuation slope of the attenuation steady-state subframe, and triggering a short-time energy attenuation compensation rule to calculate the dimensionless gain compensation factor when the absolute value of the energy attenuation slope is lower than the steady-state empirical threshold; Step S1042, applying in-situ weighted penalty to the current frame cepstral matrix using the dimensionless gain compensation factor to truncate the parasitic interference components introduced by aliasing echoes and improve the energy resolution and signal-to-noise ratio of the high-dimensional differential Mel-Cepstral coefficient vector in the frequency domain.

[0032] Preferably, step S101 includes the following sub-steps: Step S1011, performing sliding windowing processing on the received discrete-time audio sequence signal to divide it into overlapping short-time audio frame sequences; Step S1012, calculating the sum of squares of transient energy of the short-time audio frames in the short-time audio frame sequence, and performing a logarithmic transformation on the sum of squares of transient energy to generate a short-time logarithmic energy time-series signal; Step S1013, performing first-order difference calculation on the short-time logarithmic energy time-series signal using the short-time logarithmic energy difference rule, thereby outputting a logarithmic energy first-order difference sequence.

[0033] Preferably, step S102 includes the following sub-steps: Step S1021, retrieve the a priori pitch harmonic distribution matrix, which contains the fundamental frequency distribution and characteristic harmonic component distribution of the 21 strings of the guzheng under the standard string order; Step S1022, perform joint spatial constraint mapping on the logarithmic energy first-order difference sequence and the a priori pitch harmonic distribution matrix to remove the background harmonic aliasing components generated by multi-string resonance; Step S1023, after removing the background harmonic aliasing components, when the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive initiation threshold, confirm the current energy mutation position as the physical initiation transient starting point.

[0034] Preferably, step S103 includes the following sub-steps: Step S1031, in response to the determined physical start transient origin, the constant time step of the conventional audio framing process is rewritten into an adaptive non-uniform step triggered by the physical start transient origin; Step S1032, an audio segment with a length of 15ms starting from the physical start transient origin is extracted, and a narrow window function with high time domain resolution is applied to the audio segment to truncate it, so as to independently construct transient pulse subframes; Step S1033, the energy decay termination point after the physical start transient origin is located, and the signal between the audio trail after 15ms and the energy decay termination point is independently divided into decay steady-state subframes.

[0035] Preferably, step S104 includes the following sub-steps: Step S1043, performing a discrete Fourier transform on the transient pulse subframe to convert the time-domain discrete audio signal to the frequency domain, and calculating the transient power spectrum matrix; Step S1044, performing energy-weighted summation on the transient power spectrum matrix through a set of triangular filter banks arranged according to Mel frequency standards to obtain the logarithmic energy spectrum vector of the Mel filter banks; Step S1045, applying a first-order cepstral transform operator to the logarithmic energy spectrum vector of the Mel filter banks to perform a discrete cosine transform, thereby extracting the high-dimensional difference Mel cepstral coefficient vector.

[0036] Preferably, step S104 includes the following sub-steps: Step S1046, locally mapping the extracted high-dimensional difference Mel-Cepstral coefficient vector to the acoustic sound quality evaluation hidden Markov model to activate the sound quality evaluation state space; Step S1047, calculating the hidden state sequence transition probability between adjacent frames using the Markov sound quality evaluation state transition matrix, and searching for the maximum probability convergence path through dynamic programming; Step S1048, performing multi-dimensional weighted scoring based on the likelihood probability value corresponding to the maximum probability convergence path, thereby generating and outputting the sound quality evaluation score vector.

[0037] Preferably, step S102 includes the following sub-steps: step S1024, when receiving discrete-time audio sequence signals, extracting the background noise floor signal of the preceding silent audio segment; step S1025, calculating the variance envelope time-series curve of the background noise floor signal, and using the maximum value of the variance envelope time-series curve as the reference gain boundary of the adaptive start-up threshold, thereby spontaneously generating the adaptive start-up threshold.

[0038] Preferably, step S101 includes the following sub-steps: step S1014, before receiving the discrete-time audio sequence signal, a bandpass filter is used to perform high-frequency anti-aliasing filtering and low-frequency DC component filtering on the one-dimensional original discrete audio waveform stream; step S1015, the filtered audio waveform stream is input to the analog-to-digital conversion unit to perform time-domain discretization sampling, thereby generating a discrete-time audio sequence signal that satisfies the time continuity constraint.

[0039] Example 1: In this example, the digital signal processor receives discrete-time audio sequence signals. This signal corresponds to the transient impact pulse of the right hand plucking during the guzheng performance, as well as the continuous glissando and vibrato introduced by the left hand pressing the string. When multiple string resonances exist at the same time, the background harmonics are easy to overlap, and the fundamental frequency will also drift nonlinearly with the playing action. Conventional equal step frame division is difficult to retain the frequency details in the plucked transient edge and the decayed tail note at the same time.

[0040] Digital signal processors read discrete-time audio sequence signals using a fixed window sliding mechanism, following the short-time logarithmic energy difference rule. Within each window, the squares of the sampled amplitudes are summed and their logarithms are taken. Then, a first-order difference is made with the logarithmic energy of the previous adjacent window to obtain the first-order difference sequence of logarithmic energy, which characterizes the transient energy jump. : ,in, It is a first-order difference sequence of log-energy. It is a discrete-time audio sequence signal. For the time-domain variable of the sliding sampling point, The base number of sampling points is fixed at 128.

[0041] Subsequently, the digital signal processor retrieves the priori pitch harmonic distribution matrix stored in the external static memory. This matrix contains the fundamental frequency distribution and characteristic harmonic component distribution under the standard string order of the guzheng. The processor will then use the logarithmic energy first-order difference sequence. With the prior pitch harmonic distribution matrix Implement joint spatial constraint mapping: for the current decision time, As a temporal energy mutation quantity, and called The frequency domain distribution of the target string sequence is used to weight and calibrate the differential amplitude, so that the background harmonic aliasing component generated by multi-string resonance is weakened before the oscillation determination. After this mapping process, the energy mutation response that matches the fundamental frequency and characteristic harmonic components of the target string sequence is retained, and the coherent coupling energy terms that do not match the current plucked string sequence are filtered out.

[0042] When the a priori pitch harmonic distribution matrix The differential amplitude after constraint calibration exceeds the adaptive start-up threshold generated by the background noise floor signal. At this point, the digital signal processor determines the current energy mutation position as the physical start of the string plucking transient. In response to this start, the processor terminates the original equally spaced frame timing and rewrites the constant time step of the conventional audio frame segmentation process into an adaptive non-uniform step triggered by the physical start of the transient. Starting from this start, a 15ms audio segment is extracted, and a narrow window function with high time domain resolution is applied to the audio segment to independently construct transient pulse subframes. For the audio trail after 15ms, the processor continuously calculates the short-time energy of subsequent frames and takes the first frame position where the short-time energy drops to close to the reference energy corresponding to the background noise floor signal and remains stable as the energy attenuation boundary point. Then, the signal between 15ms and this energy attenuation boundary point is divided into attenuated steady-state subframes.

[0043] After dividing the signal into transient pulse subframes and decaying steady-state subframes, the digital signal processor (DSP) processes the two types of subframes separately. For transient pulse subframes, the processor performs a discrete Fourier transform to convert the time-domain discrete audio signal to the frequency domain and calculates the transient power spectrum matrix. Then, the transient power spectrum matrix is ​​passed through a set of triangular filters arranged according to the Mel frequency standard to perform energy-weighted summation, obtaining the Mel filter bank logarithmic energy spectrum vector. Subsequently, a first-order cepstral transform operator is applied to this vector to perform a discrete cosine transform, extracting the high-dimensional difference Mel cepstral coefficient vector. For attenuation steady-state subframes, the processor calls the fine spectral line feature entropy operator to calculate the spectral line information entropy vector that reflects the attenuation law of string harmonics. When it is necessary to align and record transient and steady-state features, the processor processes the high-dimensional difference Mel-frequency cepstral coefficient vector. Perform first-order mean normalization to converge the numerical range to between 0 and 1; simultaneously, adjust the spectral line information entropy vector. A dimensionless linear transformation is performed using the maximum extremum mapping function, which also converges the numerical range to between 0 and 1. The two are then concatenated in situ according to time sequence to form a combined acoustic feature time-series tensor. It is used to store the time correspondence between transient impact characteristics and steady-state decay characteristics.

[0044] During the sound quality evaluation stage, the digital signal processor extracts the high-dimensional differential Mel-frequency cepstral coefficient vector. The local mapping is applied to the Hidden Markov Model for acoustic quality evaluation, activating the quality evaluation state space. This state space includes states with good timbre, slight distortion, severe aliasing, and lossy timbre. The initial probability distribution is set so that the good timbre state accounts for 100%. The pre-set Markov quality evaluation state transition matrix is ​​trained under supervision using 1000 pre-collected standard performance audio clips. The transition frequency between the four states in adjacent frames is statistically analyzed, solidifying into a 4x4 probability transition matrix. During the state transition search process, the high-dimensional difference Mel-frequency cepstral coefficient vector... The observation vector is used in the calculation of the transmission probability, which is given by a multivariate Gaussian mixture model. The processor uses the Markov sound quality evaluation state transition matrix to calculate the hidden state sequence transition probability between adjacent frames, and searches for the path with the maximum probability convergence through the Viterbi algorithm. The logarithm of the likelihood probability corresponding to this path is extracted as the basis for multidimensional weighted scoring. The score is mapped to the interval from 0 to 100 points to generate a sound quality evaluation score vector. .

[0045] Under conditions of intensive plucking, the physical aftersound of the preceding string may coherently overlap with the starting pulse of the following string. In this case, the digital signal processor detects the energy attenuation slope of the preceding attenuated steady-state subframe. The slope is determined by the short-time energy difference between adjacent attenuation steady-state subframes and the corresponding time interval. When the energy attenuation slope... When the absolute value is lower than the fixed steady-state empirical threshold, the processor triggers the short-term energy decay compensation rule and calls the reverse energy dissipation drag operator to calculate the dimensionless gain compensation factor. : ,in, It is a dimensionless gain compensation factor. It is a dimensionless calibration constant and is fixed at 0.15. It is a dimensionless calibration constant and is fixed at 1.24. The processor utilizes the energy decay slope. Apply in-situ weighted penalty to the cepstral matrix of the current frame, specifically by... As each diagonal element of the diagonal matrix, it is multiplied left by the high-dimensional difference Mel cepstral matrix of the current frame. Through this attenuation weight, the high-order interference components in the cepstral matrix are suppressed row by row, thereby truncating the parasitic interference components introduced by aliasing echoes and improving the energy resolution and signal-to-noise ratio of the high-dimensional difference Mel cepstral coefficient vector in the frequency domain.

[0046] When the playing string produces a large-span glissando motion, causing a nonlinear instantaneous drift in the fundamental frequency, the digital signal processor constructs a one-dimensional intermediate representation vector of the fundamental frequency change rate. Before evaluating the fundamental frequency monotonic variation over three consecutive frames, the processor executes a time-domain normalization interface program. This program divides the fundamental frequency variation over the three consecutive frames by the total actual time span of those three frames to obtain the instantaneous fundamental frequency drift rate within a standard time unit. This drift rate is then multiplied by the 20ms time constant corresponding to a standard constant frame length, mapping it to an equivalent nominal frame variation. When the absolute value of the fundamental frequency monotonic variation over three consecutive frames exceeds 45Hz, the processor determines that the current state has reached a safe boundary condition. It then reduces the temporal step size of the Markov sound quality evaluation state transition matrix from a standard single-frame step to a 0.5 discrete step, and discards transition path branches that do not satisfy the pitch continuity physical constraint. Subsequently, it continues to calculate the likelihood probability value corresponding to the path with the highest probability convergence, and outputs the sound quality evaluation score vector. It is then stored in an external memory for subsequent data presentation on the digital evaluation terminal.

[0047] Through the above processing, the digital signal processor, under the conditions of a single-channel sensor link and standard sampling rate, first utilizes the logarithmic energy first-order difference sequence. and the prior pitch harmonic distribution matrix Determine the physical starting point of the transient state, and then rearrange the timeline steps of the audio frame extraction process around this starting point to process transient pulse subframes and decaying steady-state subframes separately; high-dimensional differential Mel-frequency cepstral coefficient vector. Enter the Markov sound quality evaluation state transition matrix to complete the guzheng sound quality evaluation state transition search, and decay the energy attenuation slope of the steady-state subframe. This is used to trigger the short-time energy attenuation compensation rule. Therefore, the reverberation aliasing and fundamental frequency drift caused by glissando under dense string plucking are incorporated into the same discrete calculation process, resulting in an output sound quality evaluation score vector. It can reflect the transient impact characteristics and changes in sound quality of guzheng performance audio.

[0048] Example 2: In this example, the audio signal analysis method relies on an acoustic testing system. This system uses a single-channel acoustic acquisition unit to acquire audio samples of guzheng performances. The sampling frequency is 48kHz and the sampling accuracy is 24 bits. The acquisition objects include guzheng audio at different playing speeds and different plucking forces. In order to form test conditions that include strong coherent resonance and harmonic overlap of multiple strings, the digital signal processor superimposes Gaussian white noise with a signal-to-noise ratio of 20dB and power frequency interference harmonics with a frequency of 50Hz into the input discrete-time audio sequence signal to obtain the original noisy source to be tested.

[0049] In the short-time logarithmic energy difference rule, the number of basic sampling points Used to balance the ability to capture transient boundaries in plucked strings with low-frequency string harmonic resolution, smaller It helps to shorten the detection window for energy mutations, but it weakens the clustering of low-spectral lines; a larger It can improve frequency domain energy stability, but it easily smooths out the oscillation edges caused by right-hand flicking. When the spectral bandwidth of the flicking transient energy impact waveform is greater than 4kHz, the digital signal processor will... The value is set to 128, and this value is used as the uniform sampling window length for each group of test samples.

[0050] The experiment involved setting up 10 gradient samples. The control group used the traditional fixed-window-length framing method, while the experimental group used a method for extracting audio features and evaluating the sound quality of guzheng performance, as described in this embodiment. For each sample, the digital signal processor first received the discrete-time audio sequence signal, calculated the first-order difference sequence of logarithmic energy, and then retrieved the prior pitch harmonic distribution matrix. Adaptive threshold comparison was performed on the difference sequence to determine the physical starting transient point. Subsequently, the time axis step size of the audio framing extraction process was rewritten based on this starting point. The audio segment within 15ms after the starting point was divided into transient pulse subframes, and the audio segment between the subsequent point and the energy decay boundary point was divided into decaying steady-state subframes. For transient pulse subframes, the processor extracted the high-dimensional differential Mel-frequency cepstral coefficient vector and input the vector into the Markov sound quality evaluation state transition matrix to complete the guzheng sound quality evaluation state transition search and output the sound quality evaluation score vector. For decaying steady-state subframes, the processor synchronously recorded the variance of the spectral information entropy vector to observe the fluctuation of the long-tailed harmonic energy distribution.

[0051] In two samples with an input signal-to-noise ratio of 20dB, the physical start transient capture deviations of the fixed window length framing method were 12.4ms and 14.2ms, respectively, while the capture deviations of the method in this embodiment were 1.15ms and 1.32ms, respectively. After the transient pulse subframes were extracted, the spectral distances of the high-dimensional differential Mel-frequency cepstral coefficient vectors in the control group were 4.25 and 4.62, respectively, while those in the experimental group were 1.12 and 1.25, respectively. The variances of the spectral information entropy vectors of the attenuated steady-state subframes were 0.85 and 0.92 in the control group, and 0.14 and 0.16 in the experimental group, respectively. After the above data were entered into the Markov sound quality evaluation state transition matrix, the score accuracy rates of the control group were 72.5% and 70.1%, respectively, while those of the experimental group were 94.2% and 93.8%, respectively.

[0052] When the input signal-to-noise ratio decreased to 15dB, the starting point deviations of the control group for the third and fourth samples increased to 18.5ms and 19.1ms, respectively, while those of the experimental group were 1.68ms and 1.75ms, respectively. Under the same sample, the cepstral distances of the control group were 5.18 and 5.35, respectively, while those of the experimental group were 1.41 and 1.48, respectively. The entropy variances of the control group were 1.14 and 1.21, respectively, while those of the experimental group were 0.22 and 0.25, respectively. Due to the reduction in the positioning deviation of the physical starting transient point, the flicking impact edge in the transient pulse subframe was preserved separately, and the inter-frame difference of the high-dimensional difference Mel-frequency cepstral coefficient vector in the state transition search was more stable. The corresponding score accuracies were 65.4% and 63.8% for the control group, and 91.5% and 90.9% for the experimental group, respectively.

[0053] When the input signal-to-noise ratio was further reduced to 10dB, in the fifth and sixth samples, the fixed window length framing method was affected by the superposition of background noise and multi-string resonance, with starting deviations reaching 24.6ms and 26.3ms, respectively. After constraining and calibrating the logarithmic energy first-order difference sequence using the prior pitch harmonic distribution matrix, the starting deviations in this embodiment were 2.21ms and 2.45ms, respectively. At this time, the cepstral distances of the control group were 6.42 and 6.78, and the cepstral distances of the experimental group were 1.85 and 1.92, respectively. The entropy variances of the control group were 1.58 and 1.67, and the entropy variances of the experimental group were 0.34 and 0.38, respectively. After these differences entered the sound quality evaluation stage, the scoring accuracy rates of the control group were 56.2% and 54.7%, respectively, and the scoring accuracy rates of the experimental group were 87.4% and 86.1%, respectively.

[0054] In the seventh and eighth samples with an input signal-to-noise ratio of 5dB, the starting biases of the control group were 32.8ms and 35.1ms, respectively, while those of the experimental group were 3.82ms and 4.12ms. The corresponding cepstral distances were 7.95 and 8.24 for the control group and 2.54 and 2.68 for the experimental group. The corresponding entropy variances were 2.15 and 2.31 for the control group and 0.52 and 0.58 for the experimental group. Because the experimental group first independently divided the 15ms audio segment after the physical transient starting point into transient pulse subframes and then separated the subsequent decaying tail sounds, the aliasing effect of the preceding echo on the subsequent starting pulse was limited to a smaller range. In terms of the accuracy corresponding to the sound quality evaluation score vector, the control group achieved 45.3% and 42.1%, respectively, while the experimental group achieved 81.3% and 79.6%.

[0055] When the input signal-to-noise ratio drops to 0dB, the ninth and tenth samples exhibit the strongest noise interference. The starting deviations for the control group are 45.6ms and 48.2ms, respectively, while those for the experimental group are 8.55ms and 9.18ms, respectively, both remaining within 10ms. At this point, the cepstral distances for the control group are 10.42 and 11.15, respectively, while those for the experimental group are 4.35 and 4.62, respectively. The entropy variances for the control group are 3.42 and 3.68, respectively, while those for the experimental group are 1.12 and 1.21, respectively. After inputting these features into the Markov sound quality evaluation state transition matrix, the score accuracy for the control group decreases to 31.4% and 28.9%, respectively, while the score accuracy for the experimental group is 68.4% and 65.7%, respectively.

[0056] The data changes from the above 10 samples show that as the input signal-to-noise ratio gradually decreases from 20dB to 0dB, the physical start-up transient capture deviation of the fixed window length framing method gradually increases from 12.4ms to 48.2ms, while the capture deviation of the method in this embodiment changes from 1.15ms to 9.18ms under the same noise conditions. In terms of cepstral distance, the control group increases from 4.25 to 11.15, and the experimental group increases from 1.12 to 4.62. In terms of entropy variance of attenuated steady-state subframes, the control group increases from 0.85 to 3.68, and the experimental group increases from 0.14 to 1.21. In terms of scoring accuracy, the control group decreases from 72.5% to 28.9%, and the experimental group decreases from 94.2% to 65.7%. These data correspond to the same noisy input and the same evaluation process, reflecting the impact of adaptive non-uniform step size on oscillation positioning deviation, cepstral feature fluctuation, and state transition search results under low signal-to-noise ratio conditions.

[0057] Further examining the fundamental frequency nonlinear drift caused by the large-span glissando of the left hand, the digital signal processor records the monotonic variation of the fundamental frequency in consecutive frames based on the one-dimensional fundamental frequency change rate intermediate representation vector. When the fundamental frequency variation in consecutive frames remains within 45Hz per frame, the pitch continuity physical constraint can limit the candidate paths in the Markov sound quality evaluation state transition matrix, making the state transition probability converge stably along adjacent frames. When the fundamental frequency variation in consecutive frames exceeds 45Hz and continues to rise above 60Hz, the instantaneous fundamental frequency drift caused by the change in string tension causes the likelihood probability of the transition path to saturate, and the accuracy corresponding to the sound quality evaluation score vector decreases accordingly. This phenomenon corresponds to the tension release, energy dissipation, and frequency drift that occur in the guzheng strings during large-span string pressing and glissando, indicating that the 45Hz boundary condition can serve as the triggering basis for the state transition search in this embodiment.

[0058] Therefore, in this embodiment, under the single-channel acquisition conditions of 48kHz and 24bit, a noisy test input is constructed by superimposing 20dB Gaussian white noise and 50Hz power frequency interference harmonics. The test data shows that after the discrete-time audio sequence signal is processed by the short-time logarithmic energy difference rule, the prior pitch harmonic distribution matrix constraint, the adaptive threshold comparison judgment, and the adaptive non-uniform step size framing, a high-dimensional differential Mel-frequency cepstral coefficient vector for sound quality evaluation score vector output can be formed. After this vector enters the Markov sound quality evaluation state transition matrix, it can still complete the guzheng sound quality evaluation state transition search under the conditions of strong background noise, multi-string resonance, and large dynamic glissando.

[0059] Example 3: In this example, the digital signal processor receives a discrete-time audio sequence signal obtained with user authorization. This signal corresponds to the guzheng performance audio under different performance environments. It includes strong transient impact pulses generated by the right hand plucking, as well as continuous glissando and vibrato introduced by the left hand pressing the strings. Since the amplitude and frequency distribution of the background noise are not fixed in different acquisition environments, using only a fixed threshold can easily cause the physical starting point of the transient to shift. At the same time, unplucked strings may also be excited by the surrounding sound field and generate coherent air coupling vibration. If this part of the unplucked parasitic harmonics is not processed by prior constraints, it is easy to be confused with the abrupt front in the logarithmic energy first difference sequence.

[0060] To determine the adaptive trigger threshold, the digital signal processor (DSP) first continuously collects 512 discrete data points in a silent background environment as the background noise basis signal before receiving the audio stream. It then calculates the variance envelope timing curve of this background noise basis signal and takes the maximum value of the variance envelope timing curve as the reference gain boundary. Combined with a fixed scaling factor of 3, it generates the adaptive trigger threshold. This scaling factor is used to ensure that the trigger threshold crosses the main fluctuation range of the background noise, thereby reducing the probability of false triggering caused by pure background noise. For the generation of the adaptive trigger threshold, the processor can also simultaneously calculate the short-time average energy of the background noise basis signal and verify this short-time average energy against the aforementioned reference gain boundary, so that the adaptive trigger threshold is simultaneously constrained by the background noise energy level and the maximum value of the variance envelope.

[0061] In this embodiment, the prior pitch harmonic distribution matrix solidified in the external static memory adopts a two-dimensional scalar architecture. Its row components correspond one-to-one with the standard string order of the 21 strings of the guzheng, and its column components correspond to the frequency domain sampling points of the spectral lines. The values ​​in this matrix are derived from the fundamental frequency distribution and characteristic harmonic component distribution of the guzheng strings under standard tension conditions. After the digital signal processor calculates the first-order difference sequence of logarithmic energy, it calls the row component of the prior pitch harmonic distribution matrix corresponding to the current target string order and performs joint spatial constraint mapping with the first-order difference sequence of logarithmic energy. During the mapping process, the processor performs weighted calibration on the difference amplitude according to the fundamental frequency and characteristic harmonic components of the target string order, filters out coherent coupling energy terms that do not match the current target string order, and strips the background harmonic aliasing components generated by multi-string resonance.

[0062] After calibration constrained by the a priori pitch harmonic distribution matrix, if the amplitude of the logarithmic energy first-order difference sequence is greater than the adaptive start threshold at three consecutive sampling points, the digital signal processor confirms that the current energy mutation position is the physical start transient origin and triggers the rewriting of the time axis timing step of the audio frame extraction process. Subsequently, starting from this physical start transient origin, the processor extracts an audio segment of 15ms in length, truncates it through a narrow window function, and divides it into transient pulse subframes. For the audio trail after 15ms, the processor continues to track the short-term energy decrease process and divides the signal from the subsequent point to the energy decay boundary into decay steady-state subframes.

[0063] The aforementioned 15ms timeline timing step corresponds to the physical contact excitation period of the right-hand plucking action. If this step is too short, the number of sampling points in the transient pulse subframe will be insufficient, and the low-frequency concentration of the power spectrum after the discrete Fourier transform will decrease. If this step is too long, the non-stationary signal during the long-tail decay period will enter the transient pulse subframe, weakening the independence of the transient impact characteristics. The 0.15 and 1.24 used in the short-time energy attenuation compensation rule are dimensionless calibration constants, which are used to adjust the compensation intensity and the denominator bias, respectively. When dense strumming produces strong coherent aftereffects, this set of calibration constants limits the in-situ weighted penalty amplitude of the current frame cepstral matrix to avoid excessive suppression of the normal oscillation pulse.

[0064] For the attenuated steady-state subframe, the digital signal processor performs a discrete Fourier transform to obtain a discrete frequency domain spectral line amplitude sequence of length 512. Then, it calculates the normalized probability distribution based on the total energy of the amplitude sequence, and performs negative logarithmic information entropy multiplication and summation on the normalized probability distribution to obtain the spectral line information entropy vector. The processor aligns this spectral line information entropy vector with the high-dimensional differential Mel-frequency cepstral coefficient vector in chronological order to form a combined acoustic feature time-series tensor, which is used to record the correspondence between the plucked transient and steady-state attenuation.

[0065] Constrained by the adaptive initiation threshold and the prior pitch harmonic distribution matrix, random background noise and coherent harmonics generated by non-plucked strings are weakened at the front end of the feature extraction link. The digital signal processor then inputs the extracted high-dimensional differential Mel-Cepstral coefficient vector into the Markov sound quality evaluation state transition matrix, and calculates the state transition probability trajectory by combining the steady-state attenuation information recorded in the combined acoustic feature time tensor. The sound quality evaluation score vector is then output and stored in the external memory. In this way, the plucked initiation, attenuation tail tone and environmental noise in the discrete-time audio sequence signal enter the corresponding processing links, and the determination of the physical initiation transient starting point no longer depends solely on the fixed static threshold.

[0066] Example 4: This example combines Figures 1 to 2 This paper describes a method for extracting audio features from guzheng performances and evaluating their sound quality. Figure 1 As shown, step S101 receives a discrete-time audio sequence signal and calculates the first-order logarithmic energy difference sequence of the discrete-time audio sequence signal using the short-time logarithmic energy difference rule; step S102 retrieves the fixed prior pitch harmonic distribution matrix, which performs adaptive threshold comparison judgment on the first-order logarithmic energy difference sequence, and determines the physical start transient starting point when the amplitude exceeds the adaptive start threshold generated by the background noise base signal; step S103 rewrites the time axis timing step of the frame extraction process according to the physical start transient starting point, independently divides the audio segment within 15ms after the starting point into transient pulse subframes, and divides the audio segment between the subsequent point and the energy attenuation boundary point into attenuation steady-state subframes; step S104 calls the first-order cepstral transform operator to extract the high-dimensional differential Mel-frequency cepstral coefficient vector for the transient pulse subframes, and completes the state transition search in the preset Markov sound quality evaluation state transition matrix to output the sound quality evaluation score vector.

[0067] like Figure 2 As shown, the acoustic acquisition unit acquires audio signals and inputs them to the digital signal processor. After receiving the discrete-time audio sequence signal and calculating the logarithmic energy first-order difference sequence, the digital signal processor determines the physical start transient origin. This determination process retrieves the prior pitch harmonic distribution matrix and automatically generates an adaptive start threshold. Then, it divides the transient pulse subframes and attenuated steady-state subframes. With the assistance of triggering the short-time energy attenuation compensation rule, it extracts the high-dimensional differential Mel-frequency cepstral coefficient vector, thereby completing the guzheng sound quality evaluation state transition search. Finally, the output sound quality evaluation score vector is transmitted to the digital evaluation terminal.

[0068] Example 5: In this example, before deploying a method for extracting audio features and evaluating the sound quality of a guzheng performance in a broadcast space, the digital signal processor first performs audio sound field environment calibration. The processor controls the speaker unit to output a standard white noise signal, and the single-channel acoustic acquisition unit simultaneously acquires a one-dimensional time-domain physical impedance response stream containing spatial reflection features. The walls, floor, and objects in the broadcast space will dissipate sound energy to different degrees at different frequency bands. If the subsequent feature extraction is performed directly, the amplitude-frequency distortion introduced by spatial reflection will be superimposed on the transient power spectrum matrix. Therefore, before formally receiving the discrete-time audio sequence signal, the processor first converts the one-dimensional time-domain physical impedance response stream into an environmental channel transfer function and generates a spatial acoustic equalization compensation vector accordingly. .

[0069] During the conversion process, the digital signal processor performs a 512-point discrete Fourier transform on the one-dimensional time-domain physical impedance response flow to obtain a frequency-domain complex response sequence. Subsequently, it calculates the square of the modulus of this complex response sequence, constructs a spatial power spectrum, and performs sliding window mean smoothing on the spatial power spectrum to reduce the impact of discrete noise on individual frequency points. After smoothing, the processor takes the reciprocal of the energy at each frequency point as the spatial acoustic equalization compensation vector. The values ​​of each dimension component, and Stored in an internal register for subsequent frame power spectrum correction.

[0070] After the discrete-time audio sequence signal is input, the digital signal processor calculates the first-order difference sequence of logarithmic energy according to the predetermined process, and retrieves the prior pitch harmonic distribution matrix to perform adaptive threshold comparison judgment on the difference sequence. When the difference amplitude after being constrained by the prior pitch harmonic distribution matrix exceeds the adaptive start threshold, the current energy change position is confirmed as the physical start transient starting point. The processor then rewrites the constant time step of the conventional audio framing process into an adaptive non-uniform step size, divides the audio segment within 15ms after the physical start transient starting point into transient pulse subframes, and divides the subsequent audio segment to the energy decay boundary point into decay steady-state subframes.

[0071] For transient pulse subframes, the digital signal processor performs a discrete Fourier transform to calculate the transient power spectrum matrix. When this matrix enters the spatial correction stage, the processor calls the spatial acoustic equalization compensation vector in frequency order. The columns of the transient power spectrum matrix are compared with... The components at corresponding frequency points are multiplied point by point to reduce the amplified frequency bands and compensate for the attenuated frequency bands in the broadcast space. After this correction, the transient power spectrum matrix is ​​then subjected to energy-weighted summation through a triangular filter bank arranged with Mel frequency standards, and processed by a first-order cepstral transform operator to extract the high-dimensional difference Mel cepstral coefficient vector.

[0072] Within the studio space, the background noise power may change due to personnel movement, air conditioning airflow, or equipment power supply noise. The digital signal processor continuously monitors the background noise floor signal corresponding to the preceding silent audio segment. When the change in background noise power relative to the previous stable period is greater than 6dB, the processor calls the average energy value within the last 3 time windows and performs a sliding update on the adaptive start threshold. The updated adaptive start threshold continues to serve as the threshold for determining the physical start transient, ensuring that the start point positioning is consistent with the current background noise level of the studio space, with the positioning deviation controlled within 2ms.

[0073] After spatial acoustic equalization compensation is completed, the high-dimensional differential Mel-Cepstral coefficient vector is fed into the Markov sound quality evaluation state transition matrix. The digital signal processor searches for the path with the maximum probability convergence based on the hidden state sequence transition probabilities between adjacent frames, and performs multi-dimensional weighted scoring based on the likelihood probability value corresponding to the path to generate a sound quality evaluation score vector. The sound quality evaluation score vector is stored in an external memory for retrieval by a digital evaluation terminal; simultaneously, the spatial acoustic equalization compensation vector... The data is retained in the internal register until the next studio environment calibration to regenerate the compensation data.

[0074] Example 6: In this example, after the guzheng pickup channel has been working for a long time, the pickup sensitivity and the gain of the preamplifier circuit will slowly drift. Before receiving the discrete-time audio sequence signal each time, the digital signal processor first reads the preceding silent segment audio, extracts the background noise base signal, and performs a moving average filter on the signal to obtain the current silent base amplitude. The processor compares the current silent base amplitude with the initial silent energy baseline. When the difference between the two exceeds the preset deviation threshold, the channel baseline calibration process is started; if it does not exceed the deviation threshold, the original adaptive start threshold is used to continue the subsequent start point determination.

[0075] In the channel baseline calibration process, the digital signal processor generates a variance envelope timing curve based on the background noise floor signal, and uses the maximum value of this curve as the reference gain boundary for the adaptive start threshold. For baseline changes caused by channel gain drift, the processor calculates an adaptive correction factor based on the ratio between the current sliding window mean and the initial silent energy baseline. , and then Adaptive start threshold Make a proportional correction to obtain the correction threshold. This processing only changes the value of the threshold as it is updated with the channel baseline drift, without changing the discrete-time audio sequence signal itself or the subsequent calculation method of the logarithmic energy first-order difference sequence.

[0076] After threshold correction, the digital signal processor continues to calculate the first-order logarithmic energy difference sequence of the discrete-time audio sequence signal according to the short-time logarithmic energy difference rule, and retrieves the prior pitch harmonic distribution matrix. It then performs an adaptive threshold comparison and judgment on the difference sequence after constraint. If the difference amplitude after processing by the prior pitch harmonic distribution matrix exceeds the correction threshold... At that time, the processor will determine the current position as the physical start of the transient; if no more than [time] has elapsed... At this time, the non-starting point determination of the current frame is maintained, and the frame time step rewriting is not triggered.

[0077] After determining the physical transient starting point, the digital signal processor (DSP) rewrites the constant time step in the conventional audio framing process into an adaptive non-uniform step size. The processor extracts the audio segment within 15ms after the starting point as the transient pulse subframe, and divides the audio segment from this 15ms point to the energy decay boundary point into decaying steady-state subframes. After the transient pulse subframes are processed by discrete Fourier transform, Mel frequency standard triangular filter bank, and first-order cepstral transform operator, a high-dimensional differential Mel cepstral coefficient vector is extracted. This vector is then fed into the Markov sound quality evaluation state transition matrix. The processor searches for the maximum probability convergence path of the hidden state sequence of adjacent frames and generates a sound quality evaluation score vector based on the likelihood probability value corresponding to the path. .

[0078] Under test conditions where channel sensitivity decreases by 15%, if the sensitivity is still based on the value before drift... As a starting point determination threshold, the effective abrupt change amplitude in the log-energy first-order difference sequence is synchronously suppressed by channel attenuation, making it easy for the physical starting transient point to shift backward. After adopting the channel baseline calibration procedure of this embodiment... The starting point is adjusted synchronously with the current silent base amplitude. The threshold constraint generated by the background noise base signal continues to be used to determine the sound quality evaluation score vector. The numerical offset in the evaluation results decreases accordingly, and the evaluation results are consistent with the actual state of the pickup channel.

[0079] The above description is only a few preferred embodiments of the present invention and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, technical solutions formed by replacing the above-mentioned features with the technical features with similar functions disclosed in the embodiments of the present invention.

Claims

1. A method for extracting audio features of guzheng performance and evaluating its sound quality, characterized in that, Includes the following steps: Step S101: Receive the discrete-time audio sequence signal and calculate the first-order difference sequence of the logarithmic energy of the discrete-time audio sequence signal using the short-time logarithmic energy difference rule; Step S102: Retrieve the solidified prior pitch harmonic distribution matrix and perform adaptive threshold comparison judgment on the logarithmic energy first-order difference sequence based on the prior pitch harmonic distribution matrix. When the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive start threshold spontaneously generated by the background noise base signal, determine the physical starting transient start point of string plucking. Step S103: Rewrite the timing step of the audio frame extraction process according to the physical start transient starting point, and independently divide the audio segment within 15ms after the physical start transient starting point into transient pulse subframes, and divide the audio segment from the subsequent point to the energy decay boundary point into decay steady state subframes. Step S104: Call the first-order cepstral transform operator on the transient pulse subframe to extract the high-dimensional differential Mel cepstral coefficient vector, and perform the guzheng sound quality evaluation state transition search in the preset Markov sound quality evaluation state transition matrix based on the high-dimensional differential Mel cepstral coefficient vector to output the sound quality evaluation score vector.

2. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S104 further includes the following sub-steps: Step S1041, detect the energy attenuation slope of the attenuation steady-state subframe, and when the absolute value of the energy attenuation slope is lower than the steady-state empirical threshold, trigger the short-time energy attenuation compensation rule to calculate the dimensionless gain compensation factor; Step S1042, use the dimensionless gain compensation factor to implement in-situ weighted penalty on the current frame cepstral matrix to truncate the parasitic interference components introduced by aliasing echoes, and improve the energy resolution and signal-to-noise ratio of the high-dimensional differential Mel cepstral coefficient vector in the frequency domain.

3. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S101 includes the following sub-steps: Step S1011, performing sliding windowing processing on the received discrete-time audio sequence signal to divide it into overlapping short-time audio frame sequences; Step S1012, calculating the sum of squares of transient energy of the short-time audio frames in the short-time audio frame sequence, and performing a logarithmic transformation on the sum of squares of transient energy to generate a short-time logarithmic energy time-series signal; Step S1013, performing first-order difference calculation on the short-time logarithmic energy time-series signal using the short-time logarithmic energy difference rule, thereby outputting a first-order difference sequence of logarithmic energy.

4. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1021, retrieve the prior pitch harmonic distribution matrix, which contains the fundamental frequency distribution and characteristic harmonic component distribution of the 21 strings of the guzheng under the standard string order; Step S1022, perform joint spatial constraint mapping on the logarithmic energy first-order difference sequence and the prior pitch harmonic distribution matrix to remove the background harmonic aliasing components generated by multi-string resonance; Step S1023, after removing the background harmonic aliasing components, when the amplitude of the logarithmic energy first-order difference sequence exceeds the adaptive initiation threshold, confirm the current energy mutation position as the physical initiation transient starting point.

5. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S103 includes the following sub-steps: Step S1031, in response to the determined physical start transient origin, the constant time step of the conventional audio framing process is rewritten to an adaptive non-uniform step triggered by the physical start transient origin; Step S1032, an audio segment with a length of 15ms starting from the physical start transient origin is extracted, and a narrow window function with high temporal resolution is applied to the audio segment to truncate it, so as to independently construct transient pulse subframes; Step S1033: Locate the energy decay termination point after the physical start transient, and independently divide the signal between the audio trail 15ms later and the energy decay termination point into decay steady-state subframes.

6. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S104 includes the following sub-steps: Step S1043, performing a discrete Fourier transform on the transient pulse subframe to convert the time-domain discrete audio signal to the frequency domain and calculating the transient power spectrum matrix; Step S1044, performing energy-weighted summation on the transient power spectrum matrix through a set of triangular filter banks arranged according to Mel frequency standards to obtain the logarithmic energy spectrum vector of the Mel filter banks; Step S1045, applying a first-order cepstral transform operator to the logarithmic energy spectrum vector of the Mel filter banks to perform a discrete cosine transform, thereby extracting the high-dimensional difference Mel cepstral coefficient vector.

7. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S104 includes the following sub-steps: Step S1046, locally mapping the extracted high-dimensional difference Mel-Cepstral coefficient vector to the acoustic sound quality evaluation hidden Markov model to activate the sound quality evaluation state space; Step S1047, using the Markov sound quality evaluation state transition matrix to calculate the hidden state sequence transition probability between adjacent frames, and searching for the maximum probability convergence path through dynamic programming; Step S1048, performing multi-dimensional weighted scoring based on the likelihood probability value corresponding to the maximum probability convergence path, thereby generating and outputting the sound quality evaluation score vector.

8. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S102 includes the following sub-steps: Step S1024, when receiving discrete-time audio sequence signals, extracting the background noise base signal of the preceding silent audio segment; Step S1025, calculating the variance envelope time-series curve of the background noise base signal, and using the maximum value of the variance envelope time-series curve as the reference gain boundary of the adaptive start-up threshold, thereby spontaneously generating the adaptive start-up threshold.

9. The method for extracting audio features of guzheng performance and evaluating its sound quality according to claim 1, characterized in that, Step S101 includes the following sub-steps: Step S1014, before receiving the discrete-time audio sequence signal, a bandpass filter is used to perform high-frequency anti-aliasing filtering and low-frequency DC component filtering on the one-dimensional original discrete audio waveform stream; Step S1015, the filtered audio waveform stream is input to the analog-to-digital conversion unit to perform time-domain discretization sampling, thereby generating a discrete-time audio sequence signal that satisfies the time continuity constraint.

Images