A mobile terminal cough sound high trajectory and vocal cord disturbance combined extraction method based on frame correlation

Abstract

The application discloses a kind of based on frame self-correlation's mobile terminal cough sound high trajectory and vocal cord disturbance combined extraction method, belong to audio signal processing technical field.The method includes: receiving cough audio signal and carrying out frame with fixed frame length and frame shift, obtain frame sequence;The self-correlation calculation of each frame is executed in pre-set lag range, take the lag amount corresponding to the maximum peak of self-correlation function and convert into the fundamental frequency value of this frame, form frame-level fundamental frequency sequence;Based on the frame-level fundamental frequency sequence, calculate pitch statistical features, the pitch statistical features include fundamental frequency mean and fundamental frequency standard deviation;Based on the fundamental frequency difference absolute value of adjacent frame and the fundamental frequency mean, calculate frame-level frequency microdisturbance;Based on the root mean square amplitude difference absolute value of adjacent frame, calculate frame-level amplitude microdisturbance.The application directly reuses frame-level self-correlation pitch detection result in vocal cord disturbance feature extraction, without pitch period mark, applicable to cough class non-steady-state signal, realize the joint extraction of pitch trajectory and vocal cord disturbance feature on mobile terminal, give consideration to computational efficiency and extraction stability.

Description

Technical Field

[0001] This invention belongs to the field of audio signal processing technology, specifically relating to a method for jointly extracting cough pitch trajectory and vocal cord perturbation on a mobile device based on frame autocorrelation. Background Technology

[0002] Cough is a common symptom of respiratory diseases, and its acoustic characteristics have important applications in telemedicine, chronic respiratory disease monitoring, and infectious disease screening. The pitch trajectory of a cough reflects the change in the fundamental frequency of vocal cord vibration over time, while vocal cord perturbation characteristics, including frequency perturbations and amplitude perturbations, characterize the short-term instability of vocal cord vibration frequency and amplitude, respectively. Both are key acoustic indicators for assessing vocal cord function and respiratory pathological changes. In existing technologies, mainstream schemes for extracting pitch trajectories and vocal cord perturbation features rely on pitch period marking techniques. Specifically, schemes represented by Praat speech analysis software first accurately locate the start and end boundaries of each pitch period in the time domain, i.e., marking the signal periodically. Then, they extract the fundamental frequency value and amplitude peak value within each marked period, and finally calculate the frequency perturbation and amplitude perturbation from the periodically sequenced data. This scheme has high marking accuracy for steady-state vowel signals, but its core assumption in the pitch period marking algorithm is that the signal has a stable periodic structure within a local time window. Coughing sounds are non-steady-state, short-duration, and broadband burst signals, with rapid changes in glottal excitation within a short time, lacking the clear periodic structure of steady-state vowels. In actual tests, Praat's pitch period marking failure rate for cough segments exceeds 60%, with a large number of frames being incorrectly marked or missed, causing subsequent fundamental frequency extraction and perturbation feature calculations to be unstable. Another existing approach is to perform a global fundamental frequency estimation on the entire cough audio and output a single fundamental frequency value as a representative of the entire signal. While this approach avoids the difficulty of periodic labeling, the fundamental frequency of the cough sound exhibits a clear time-varying trajectory during the attack. The single fundamental frequency value completely loses the dynamic information of pitch changes over time, and it is impossible to deduce vocal cord disturbance characteristics from a single fundamental frequency value. Furthermore, both of the above schemes require a separate perturbation feature calculation process to be initiated after the fundamental frequency marking or global fundamental frequency estimation is completed. The data streams of the two stages are separated, resulting in high computational redundancy. Under the limited computing power of mobile terminals, traversing the audio data twice will further increase processing latency, which is not conducive to real-time or near-real-time cough acoustic analysis applications. Summary of the Invention

[0003] The technical problem this invention aims to solve is that in the analysis of cough sound signals on mobile devices, existing technologies rely on pitch period marking and cannot stably process non-steady cough signals. Furthermore, pitch feature extraction and vocal cord perturbation feature extraction are two independent processes, resulting in low computational efficiency. To address the aforementioned technical problems, this invention provides a method for jointly extracting cough pitch trajectory and vocal cord perturbation on a mobile device based on frame-based autocorrelation. The method includes the following steps: dividing the cough audio signal into frames with a fixed frame length and a fixed frame shift to obtain a frame sequence; performing autocorrelation calculation on each frame within a preset lag range, taking the lag value corresponding to the maximum peak value of the autocorrelation function, and converting the lag value into the fundamental frequency value of that frame; the fundamental frequency values ​​of all frames constitute a frame-level fundamental frequency sequence; calculating pitch statistical features based on the frame-level fundamental frequency sequence, including the fundamental frequency mean and fundamental frequency standard deviation; calculating frame-level frequency perturbation based on the ratio of the absolute value of the fundamental frequency difference between adjacent frames to the fundamental frequency mean; and calculating frame-level amplitude perturbation based on the ratio of the absolute value of the root mean square amplitude difference between adjacent frames to the root mean square amplitude mean of adjacent frames. In this invention, the framing operation divides the non-steady-state cough signal into a frame sequence with local short-time stationarity, enabling the autocorrelation algorithm to reliably detect periodic components within each frame. The preset lag range limits the reasonable frequency range for fundamental frequency search, avoiding false detections at excessively high or low frequencies. The frame-level fundamental frequency sequence fully preserves the trajectory information of pitch changes over time during a cough attack, while the fundamental frequency mean and standard deviation provide a statistical description of pitch levels for subsequent analysis. Frame-level frequency perturbations use the fundamental frequency difference between adjacent frames instead of the periodic frequency perturbations that require periodic marking in traditional schemes, and frame-level amplitude perturbations use the frame-level root mean square amplitude instead of the periodic amplitude peak value. This allows the calculation of frequency and amplitude perturbations to directly reuse the results of frame-level autocorrelation pitch detection without the need for additional periodic marking procedures. The entire method is executed by the mobile terminal processor, adapting to the computational resource constraints of mobile devices. The beneficial effects of this invention are as follows: First, robustness to non-stationary cough signals. The frame-based autocorrelation method does not rely on the assumption of global periodicity of the signal, but rather detects periodic components within a local short time window of each frame. Although the cough signal is generally non-stationary, it still possesses local quasi-periodicity within millisecond-level short frames, and the autocorrelation algorithm can reliably extract the fundamental frequency within this local range. Compared with the traditional Praat pitch period marking scheme, the method of this invention significantly improves the success rate of fundamental frequency extraction for cough segments, and avoids large-scale missed detections and false detections due to the non-stationary characteristics of the signal. Second, computational efficiency. The output of frame-level autocorrelation pitch detection—the frame-level fundamental frequency sequence and the frame-level root mean square amplitude sequence—is directly reused for subsequent frequency perturbation and amplitude perturbation calculations. Pitch extraction and perturbation extraction share the same data path, avoiding the two independent traversals of pitch marking followed by perturbation calculation in traditional schemes. On mobile terminal processors, this method reduces the total computational load of data processing and shortens the latency from audio input to feature output. Third, dual-feature joint extraction. This invention outputs both pitch statistical features and vocal cord perturbation features simultaneously after one frame segmentation and autocorrelation calculation, unlike traditional solutions which require deploying two separate algorithm modules. This joint extraction mechanism allows mobile terminals to simultaneously obtain the fundamental frequency distribution information of cough sounds and vocal cord instability indicators in a single processing flow, providing complete multi-dimensional acoustic features for downstream applications such as cough classification and disease assessment. Attached Figure Description Figure 1 This is a flowchart of the frame-by-frame autocorrelation pitch trajectory extraction in an embodiment of the present invention.

[0004] Figure 2 This is a schematic diagram illustrating the calculation of frame-level frequency perturbation and frame-level amplitude perturbation in an embodiment of the present invention. The meanings of the markings in the diagram are as follows: 101-Cough audio signal input, 102-Frame segmentation operation, 103-Frame sequence, 104-Frame-by-frame autocorrelation calculation, 105-Peak hysteresis extraction, 106-Hysteresis to fundamental frequency value conversion, 107-Frame-level fundamental frequency sequence, 108-Silent frame filtering, 109-Effective frame fundamental frequency sequence. 201 - Frame-level fundamental frequency sequence input; 202 - Absolute value of fundamental frequency difference between adjacent frames; 203 - Mean value of fundamental frequency; 204 - Frame-level frequency perturbation calculation; 205 - Frame-level frequency perturbation value; 206 - Frame sequence; 207 - Frame-by-frame root mean square amplitude calculation; 208 - Frame-level root mean square amplitude sequence; 209 - Absolute value of root mean square amplitude difference between adjacent frames; 210 - Mean value of root mean square amplitude; 211 - Frame-level amplitude perturbation calculation; 212 - Frame-level amplitude perturbation value. Figure 2 The meanings of the markings are as follows: 201-Frame-level fundamental frequency sequence input, 202-Absolute value of fundamental frequency difference between adjacent frames, 203-Mean fundamental frequency, 204-Frame-level frequency perturbation calculation, 205-Output frame-level frequency perturbation value, 206-Frame sequence input, 207-Frame-by-frame root mean square amplitude calculation, 208-Frame-level root mean square amplitude sequence, 209-Absolute value of root mean square amplitude difference between adjacent frames, 210-Mean mean root mean square amplitude, 211-Frame-level amplitude perturbation calculation, 212-Output frame-level amplitude perturbation value. Detailed Implementation The technical solution of the present invention will now be described in detail with reference to the accompanying drawings. like Figure 1As shown in labels 101 to 109, the frame-by-frame autocorrelation pitch trajectory extraction process of this embodiment includes: after inputting the cough audio signal (label 101), performing a frame-by-frame operation (label 102) to obtain a frame sequence (label 103), performing autocorrelation calculation on each frame of the frame sequence (label 104) to extract the peak hysteresis (label 105), converting the hysteresis into a fundamental frequency value (label 106) to obtain a frame-level fundamental frequency sequence (label 107), and outputting an effective frame fundamental frequency sequence (label 109) after filtering through silent frames (label 108). like Figure 2 As shown in labels 201 to 212, the calculation process for frame-level frequency perturbation and amplitude perturbation includes: calculating the absolute value of the fundamental frequency difference between adjacent frames (label 202) and the mean fundamental frequency (label 203) based on the frame-level fundamental frequency sequence input (label 201), and performing frame-level frequency perturbation calculation (label 204) to obtain the frame-level frequency perturbation value (label 205); simultaneously, performing frame-by-frame root mean square amplitude calculation (label 207) based on the frame sequence (label 206) to obtain the frame-level root mean square amplitude sequence (label 208), calculating the absolute value of the root mean square amplitude difference between adjacent frames (label 209) and the mean root mean square amplitude (label 210), and performing frame-level amplitude perturbation calculation (label 211) to obtain the frame-level amplitude perturbation value (label 212). The core concept of this invention is to transform the non-stationary cough signal into a local quasi-stationary frame on a millisecond time scale using a frame-segmentation operation. Within each frame, the fundamental frequency period is directly detected through autocorrelation calculation. The frame-level fundamental frequency difference is used to replace the traditional period-level difference to calculate the frequency perturbation, and the frame-level root mean square amplitude is used to replace the period-level amplitude peak to calculate the amplitude perturbation. Thus, the pitch trajectory of the cough sound and the vocal cord perturbation features can be jointly extracted without the need for a fundamental frequency period marker. In one specific implementation, the mobile terminal acquires a cough audio signal at a sampling rate of 16000 Hz using a built-in microphone, with a sampling bit depth of 16 bits and mono audio. The audio signal is first preprocessed, including DC bias removal and amplitude normalization, to make its mean zero and peak amplitude normalized to 1.0. The framing operation uses a fixed frame length of 2048 sampling points (corresponding to a 128-millisecond time window at a 16000 Hz sampling rate) and a frame shift of 512 sampling points (corresponding to a 32-millisecond inter-frame step), meaning that adjacent frames overlap by 1536 sampling points. After framing, a frame sequence is obtained, with the number of frames depending on the duration of the cough audio. For a typical cough episode lasting 1.5 seconds, approximately 85 frames are generated. The reason for choosing a 128-millisecond frame length is that the glottal period of a cough corresponds to a fundamental frequency range of approximately 100 Hz to 800 Hz in most populations. A 128-millisecond time window is sufficient to cover at least a dozen glottal periods, ensuring that the autocorrelation function contains sufficient periodic information. At the same time, the window length should not be too large to avoid time-varying effects on the intra-frame signal. Autocorrelation is calculated for each frame in the frame sequence. The autocorrelation function is defined as the inner product of the frame sample sequence and itself at different hysteresis. For a frame containing a signal x[n] with N sampling points, its autocorrelation function R[τ] at hysteresis τ is the sum of the products of x[n] and x[n+τ] in the range n from 0 to N-τ-1. The hysteresis range is set according to the target fundamental frequency range, with the lower limit being the sampling rate divided by 500 Hz (i.e., τ_min = 16000 / 500 = 32 sampling points, corresponding to the upper limit fundamental frequency of 500 Hz), and the upper limit being the sampling rate divided by 50 Hz (i.e., τ_max = 16000 / 50 = 320 sampling points, corresponding to the lower limit fundamental frequency of 50 Hz). Setting the lower limit of the hysteresis range to the sampling rate divided by 500 Hz rather than a lower value is to eliminate spurious autocorrelation peaks generated by high-frequency noise at extremely short hysteresis. The maximum peak position of the autocorrelation function is searched within the hysteresis range [32, 320]. If the ratio of the autocorrelation value R[τ_peak] at the maximum peak to the zero-hysteresis autocorrelation value R[0] is less than 0.3, the frame is considered a silent frame or a noise frame and is not included in the effective frame fundamental frequency sequence. For effective frames, the hysteresis τ_peak corresponding to the peak is converted into the fundamental frequency value F0 = 16000 / τ_peak, in Hertz. The fundamental frequency values ​​of all effective frames are arranged in chronological order to form a frame-level fundamental frequency sequence. In one test, an audio clip containing two cough bursts yielded 72 valid frames. The frame-level fundamental frequency sequence values ​​ranged from 132 Hz to 387 Hz, exhibiting a trajectory of first rising and then falling, consistent with the glottal dynamics of a cough burst. Pitch statistics were calculated based on frame-level fundamental frequency sequences. The fundamental frequency mean was the arithmetic mean of the fundamental frequency values ​​of all valid frames, and the fundamental frequency standard deviation was the standard deviation of the fundamental frequency values ​​of all valid frames relative to the fundamental frequency mean. In the test, the fundamental frequency mean of this cough audio segment was 231.4 Hz, and the fundamental frequency standard deviation was 42.7 Hz. The fundamental frequency median was 225.8 Hz, and the fundamental frequency interquartile range was the difference between the upper quartile (271.3 Hz) and the lower quartile (182.6 Hz) from the frame-level fundamental frequency sequence, which is 88.7 Hz. These statistical characteristics quantify the overall frequency level and frequency fluctuation of vocal cord vibration during a cough attack. The calculation process for frame-level frequency perturbation is as follows: Take the fundamental frequency values ​​F0[i] and F0[i+1] of adjacent valid frames from the frame-level fundamental frequency sequence, and calculate the absolute difference between them |F0[i+1] - F0[i]|. Calculate the arithmetic mean of the absolute differences for all adjacent valid frame pairs, and then divide by the mean fundamental frequency to obtain the frame-level frequency perturbation value. In the test, this value was 0.0247 (i.e., 2.47%), characterizing the frequency instability of the cough sound on the inter-frame time scale. The calculation of frame-level frequency perturbation directly reuses the frame-level fundamental frequency sequence generated in the autocorrelation stage, without the need for additional pitch period marking and data traversal. The calculation process for frame-level amplitude perturbation is as follows: For each frame in the frame sequence, the root mean square (RMS) amplitude is calculated. The RMS amplitude is the square root of the mean of the sum of squares of all sample points within the frame. The RMS amplitude values ​​of all frames are arranged in chronological order to form a frame-level RMS amplitude sequence. The arithmetic mean of the absolute values ​​of the differences in RMS amplitudes between adjacent frames is calculated, and then divided by the mean of the RMS amplitudes of all frames to obtain the frame-level amplitude perturbation value. In the test, this value was 0.0823 (i.e., 8.23%). Frame-level amplitude perturbation also directly reuses the results from the framing stage, eliminating the need for separate periodic amplitude peak extraction. It is worth noting that, unlike traditional methods that calculate frequency perturbations at the pitch period level, this invention calculates frequency and amplitude perturbations at the frame level. The rationale for the frame-level granularity lies in the fact that coughing sounds, as non-steady-state burst signals, exhibit glottal characteristic changes over a timescale of milliseconds to tens of milliseconds. A 32-millisecond frame shift corresponds to a frame rate of 31.25 Hz, effectively capturing such short-term changes. Overly fine periodic-level perturbation analysis is not only unhelpful for coughing signals but also introduces the inherent non-stationarity of the glottis into the perturbation measurement, leading to artificially high and unstable perturbation values. In practical mobile terminal deployments, the above method is integrated into mobile health monitoring applications as a software module. After a user triggers a cough recording, the audio data is processed locally on the terminal by the processor to perform the calculations of framing, autocorrelation, statistical features, and perturbation features, without relying on a cloud server. The entire processing time for a 1.5-second cough audio clip on mainstream mobile terminal processors is approximately 40 to 60 milliseconds, meeting the requirements for near real-time analysis. The output four feature values—fundamental frequency mean, fundamental frequency standard deviation, frame-level frequency perturbation, and frame-level amplitude perturbation—can be directly input into downstream disease classification models or anomaly detection algorithms. In another implementation, the mobile terminal acquires the cough audio signal at a sampling rate of 44100 Hz. Correspondingly, the fixed frame length for framing is 4096 sampling points (approximately 93 milliseconds), and the frame shift is 1024 sampling points. The lower limit of the hysteresis range for autocorrelation calculation is 44100 / 500 ≈ 88 sampling points, and the upper limit is 44100 / 50 ≈ 882 sampling points. The remaining steps are the same as in the aforementioned implementation. Different sampling rate configurations do not affect the core method flow of this invention; only the frame length, frame shift, and hysteresis range parameters need to be adjusted proportionally to the sampling rate. The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention. It should be noted that the specific parameter values, platform interface names, and system component names in the above embodiments are merely illustrative examples. Those skilled in the art will understand that adjustments can be made in actual deployments based on specific application scenarios and platform environments, but the core technical solutions of the present invention are not limited to the specific examples described above.

Claims

1. A method for jointly extracting cough pitch trajectory and vocal cord perturbation on a mobile device based on frame autocorrelation, characterized in that, Includes at least the following steps: Step 1: Divide the cough audio signal into frames with a fixed frame length and a fixed frame shift to obtain a frame sequence; wherein, the fixed frame length is the number of sampling points raised to the power of 2, and the fixed frame shift is one-quarter of the fixed frame length; Step 2: For each frame in the frame sequence, perform autocorrelation calculation within a preset hysteresis range, taking the ratio of the maximum peak value of the autocorrelation function to the zero-hysteresis autocorrelation value. Frames with a ratio lower than the validity threshold are marked as silent frames and discarded, while frames with a ratio not lower than the validity threshold are marked as valid frames. For each valid frame, take the hysteresis corresponding to the maximum peak value of the autocorrelation function and convert the hysteresis into the fundamental frequency value of that frame. The fundamental frequency values ​​of all valid frames constitute the valid frame fundamental frequency sequence. Wherein, the lower limit of the preset hysteresis range is the sampling rate of the cough audio signal divided by a first frequency threshold, and the upper limit of the preset hysteresis range is the sampling rate of the cough audio signal divided by a second frequency threshold. Step 3: Calculate pitch statistical features based on the effective frame fundamental frequency sequence. The pitch statistical features include the fundamental frequency mean, fundamental frequency standard deviation, fundamental frequency median, and fundamental frequency interquartile range. Step 4: Calculate the frame-level frequency perturbation based on the ratio of the average absolute value of the fundamental frequency difference between adjacent valid frames to the mean value of the fundamental frequency; Step 5: Calculate the frame-level amplitude perturbation based on the ratio of the average absolute value of the root mean square amplitude difference of adjacent valid frames to the average root mean square amplitude of all frames. Step 6: Calculate the root mean square amplitude sequence based on the root mean square amplitude value of each valid frame in the effective frame fundamental frequency sequence; Step 7: Combine the effective frame fundamental frequency sequence, fundamental frequency mean, fundamental frequency standard deviation, fundamental frequency median, fundamental frequency interquartile range, frame-level frequency perturbation, and frame-level amplitude perturbation into a seven-dimensional vocal tract functional feature vector and output it.

2. The method according to claim 1, characterized in that, The fixed frame length is 2048 sampling points, the fixed frame shift is 512 sampling points, the first frequency threshold is 500 Hz, the second frequency threshold is 50 Hz, and the validity threshold is 0.

3.

3. The method according to claim 1, characterized in that, The step of converting the hysteresis into the fundamental frequency value of the frame is specifically: dividing the sampling rate of the cough audio signal by the hysteresis to obtain the fundamental frequency value of the frame.

4. The method according to claim 1, characterized in that, The calculation of the frame-level frequency perturbation and the frame-level amplitude perturbation directly reuses the frame-level fundamental frequency sequence generated in step 2 and the root mean square amplitude sequence generated in step 6, without performing additional data traversal or pitch period marking operations separately.

5. The method according to claim 1, characterized in that, The seven-dimensional vocal cord functional feature vector includes: the complete time series of the fundamental frequency values ​​of the effective frames, the mean of the fundamental frequency, the standard deviation of the fundamental frequency, the median of the fundamental frequency, the interquartile range of the fundamental frequency, the percentage of frame-level frequency perturbation, and the percentage of frame-level amplitude perturbation.

6. The method according to claim 1, characterized in that, The seven-dimensional vocal cord functional feature vector is directly input into the downstream respiratory disease classification model or vocal cord abnormality detection algorithm as the input feature of the downstream model.

7. The method according to claim 1, characterized in that, The overlap rate between adjacent frames in the frame sequence is 75%.

8. A mobile terminal cough pitch trajectory and vocal cord perturbation joint extraction system based on frame autocorrelation, characterized in that, include: The framing module is used to segment the cough audio signal into frames with a fixed frame length and a fixed frame shift to obtain a frame sequence; wherein the fixed frame length is the number of sampling points raised to the power of 2, and the fixed frame shift is one-quarter of the fixed frame length; The silent frame filtering module is used to perform autocorrelation calculation on each frame in the frame sequence, take the ratio of the maximum peak value of the autocorrelation function to the zero-hysteresis autocorrelation value, mark frames with the ratio below the validity threshold as silent frames and discard them, and mark frames with the ratio not below the validity threshold as valid frames. The autocorrelation pitch detection module is used to perform autocorrelation calculation on the effective frames within a preset hysteresis range, take the hysteresis corresponding to the maximum peak value of the autocorrelation function, convert the hysteresis into the fundamental frequency value of the frame, and the fundamental frequency values ​​of all effective frames constitute the effective frame fundamental frequency sequence; wherein, the lower limit of the preset hysteresis range is the sampling rate of the cough audio signal divided by a first frequency threshold, and the upper limit of the preset hysteresis range is the sampling rate of the cough audio signal divided by a second frequency threshold; The pitch statistics module is used to calculate pitch statistics features based on the effective frame fundamental frequency sequence. The pitch statistics features include the fundamental frequency mean, fundamental frequency standard deviation, fundamental frequency median, and fundamental frequency interquartile range. The frequency perturbation calculation module is used to calculate frame-level frequency perturbations based on the ratio of the average absolute value of the fundamental frequency difference between adjacent valid frames to the mean value of the fundamental frequency. The amplitude perturbation calculation module is used to calculate frame-level amplitude perturbations based on the ratio of the average absolute value of the root mean square amplitude difference between adjacent valid frames to the average root mean square amplitude of all frames. The feature vector combination module is used to combine the effective frame fundamental frequency sequence, fundamental frequency mean, fundamental frequency standard deviation, fundamental frequency median, fundamental frequency interquartile range, frame-level frequency perturbation and frame-level amplitude perturbation into a seven-dimensional vocal cord functional feature vector and output it.

9. A mobile terminal device, characterized in that, The device includes a processor, a memory, and an audio acquisition module. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6. The audio acquisition module is used to acquire the cough audio signal.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images