CN118016106A - Emotional health analysis and support system for the elderly - Google Patents

Abstract

The present invention relates to the field of speech analysis technology, and further to an emotional health analysis and support system for the elderly. The system comprises: a speech acquisition device, a speech processing device and a speech emotion analysis device; the speech acquisition device is used to obtain the speech signal of the target elderly within a set time period; the speech processing device is used to apply a pre-emphasis filter to the collected speech signal to balance the spectrum, obtain a pre-processed signal, extract MFCC features, fundamental frequency features and energy features from the pre-processed signal as elements in a feature vector to form a feature vector; the speech emotion analysis device is used to perform emotion analysis on the feature vector of the speech segment using a pre-trained speech emotion analysis model, judge the emotional features of the speech segment, and issue a warning signal requiring emotional intervention. The present invention can monitor the emotional state of the elderly in real time, accurately identify emotional features, issue warning signals in a timely manner, and perform emotional intervention.

Description

Emotional health analysis and support system for the elderly

Technical Field

The present invention belongs to the technical field of speech analysis, and in particular relates to an emotional health analysis and support system for the elderly.

Background Art

With the continuous development of society and the aging of the population, the emotional health of the elderly has attracted increasing attention. The emotional health of the elderly has an important impact on their quality of life and social participation. Therefore, it is of great significance to develop a system that can timely monitor and support the emotional health of the elderly. Traditional emotional health monitoring methods mainly rely on face-to-face assessments by medical institutions or professionals. This method has problems such as high resource consumption, high cost, and poor real-time performance, and it is difficult to meet the needs of large-scale emotional health monitoring. Therefore, it is of great practical significance to develop a system that can automatically monitor the emotional health of the elderly and provide support and intervention when necessary.

In the past few years, with the development of speech processing and sentiment analysis, more and more research has focused on using speech signals for emotion recognition and health monitoring. Traditional speech sentiment analysis methods mainly rely on machine learning-based models, such as support vector machines (SVMs) and deep neural networks (DNNs). These methods extract features from speech signals and train them with labeled emotion categories to achieve emotion recognition and classification. However, these methods often require a large amount of labeled data and a complex model training process, and in practical applications, they have problems such as low accuracy and poor generalization ability.

In addition to machine learning-based methods, some methods based on signal processing and feature extraction have been proposed for speech emotion analysis. For example, emotion recognition is performed using features such as the fundamental frequency, energy, and Mel-frequency cepstral coefficients (MFCC) of speech signals. These methods usually have good real-time performance and applicability, but their accuracy and robustness still need to be improved.

In addition, some attempts have been made in academia and industry to develop systems for monitoring the emotional health of the elderly. These systems usually include components such as voice acquisition devices, emotion analysis algorithms, and support intervention mechanisms, and aim to monitor and support the emotional health of the elderly by analyzing their voice signals. However, existing systems often lack in-depth analysis of the elderly's voice characteristics and effective emotion recognition algorithms, resulting in poor results in practical applications, and problems such as high false alarm rates and low accuracy.

Therefore, in the field of emotional health analysis and support for the elderly, a system is needed that can combine speech processing technology and sentiment analysis algorithms to accurately monitor and support the emotional state of the elderly. Such a system should have the ability to efficiently extract and analyze the characteristics of the elderly's voice signals, and at the same time combine advanced sentiment analysis algorithms to accurately judge the emotional state of the elderly and intervene in a timely manner. At the same time, the system should also consider the personalized needs and privacy protection issues of the elderly to ensure its acceptability and reliability in practical applications.

Summary of the invention

The main purpose of the present invention is to provide an intelligent control system for primary frequency modulation of the power grid, which realizes comprehensive monitoring and support for the emotional health of the elderly through the collection, processing and emotional analysis of voice signals, and has important application prospects and social significance. The present invention can monitor the emotional state of the elderly in real time, accurately identify emotional characteristics, issue early warning signals in a timely manner, conduct emotional intervention and support, etc., which will provide effective protection and support for the emotional health of the elderly.

In order to solve the above problems, the technical solution of the present invention is achieved as follows:

An emotional health analysis and support system for the elderly, the system comprising: a speech collection device, a speech processing device and a speech emotion analysis device; the speech collection device is used to obtain the speech signal of the target elderly within a set time period, and perform preliminary signal analysis on the speech signal to determine whether speech emotion analysis is required, specifically including: statistical analysis of the average speech energy and the proportion of speech frequencies of the speech signal; the average speech energy is defined as the ratio of the total energy of the speech signal to the time period within the set time period; the proportion of speech frequencies is defined as the ratio of the length of the speech signal to the time period within the set time period; if the average speech energy or the proportion of speech frequencies are both within the corresponding threshold ranges, it is determined that speech emotion analysis is not required, otherwise, it is determined that speech emotion analysis is required; the speech processing device is used to collect the speech signals in the speech collection device When it is determined that speech emotion analysis is needed, a pre-emphasis filter is applied to the collected speech signal to balance the spectrum to obtain a preprocessed signal, MFCC features, fundamental frequency features and energy features are extracted from the preprocessed signal as elements in a feature vector to form a feature vector, and based on the feature vector, the preprocessed signal is divided into speech segments and non-speech segments using a zero crossing rate method; the speech emotion analysis device is used to perform emotion analysis on the feature vector of the speech segment using a pre-trained speech emotion analysis model to determine the emotional features of the speech segment, the emotional features include: positive emotional features, neutral emotional features and negative emotional features, if the ratio of the total number of frames of the speech segment with negative emotional features within a set time period to the length of the time period exceeds a set threshold, it is determined that the target elderly person is in a negative emotion, and a warning signal requiring emotional intervention is issued.

Furthermore, the voice collection device includes: a collection unit, an enhancement unit, a preliminary analysis unit and a noise separation unit; the collection unit is used to determine whether the voice signal emitted is emitted by the target elderly person through voice recognition within a set time period, and if so, to collect the voice signal; the enhancement unit is used to enhance the voice signal to obtain an enhanced voice signal; the preliminary analysis unit performs preliminary analysis of the voice signal to determine whether voice emotion analysis is required; the noise separation unit is used to separate background noise from the voice signal when it is determined that voice emotion analysis is required.

Furthermore, the enhancement unit performs signal enhancement on the speech signal to obtain an enhanced speech signal, which includes: using the following formula to enhance the speech signal: Perform short-time Fourier transform based on the autoregressive model to obtain the time-frequency representation :

;

in, represents the time segment index of the short-time Fourier transform; represents the frequency index; is the window length for each time segment; is the window function; are the coefficients of the autoregressive model, is the order of the autoregressive model; is the imaginary number symbol; Time domain index; the time-frequency representation is enhanced using a Wiener filter with nonlinear dynamic range compression characteristics through the following formula:

;

in, and Represent the power spectrum estimation of noise and speech signal respectively; In order to enhance the frequency domain representation of speech signals, the enhanced frequency domain signals Perform inverse short-time Fourier transform to obtain the enhanced speech signal.

Furthermore, the noise separation unit, when determining that speech emotion analysis is required, separates background noise from speech signals by: representing the speech signals as waveform functions in the time domain ; Convert it to the frequency domain through short-time Fourier transform to get the frequency domain representation ; Let the length of the time period be , using length The window function is segmented, and the window length is , the overlap length between windows is ; Select the Hamming window as the window function and define the window function as:

;

in, represents the sampling index of the window; it is extended to a length of 1 by applying the window function to each time segment of the speech signal and applying zero padding , to obtain the window signal in the time domain ; For each window signal Applying discrete Fourier transform, we get the frequency domain representation ; Assume that the background noise is steady-state and linearly superimposed with the speech signal; use an adaptive filter in the frequency domain to model and estimate the background noise; Assume represents the clean spectrum of the speech signal, Represents the spectrum of the background noise; the frequency domain response of the adaptive filter is defined as:

;

in, It's in time The frequency domain response of the adaptive filter at ; using the following formula, using the adaptive filter Reconstruct the speech signal in the frequency domain to obtain the reconstructed signal:

;

in, To reconstruct the signal; convert the reconstructed signal back to the time domain to obtain the speech signal after separating the background noise from the speech signal .

Furthermore, when the voice collection device determines that voice emotion analysis is required, the voice processing device applies a pre-emphasis filter to the collected voice signal to balance the spectrum, and then divides the voice signal into overlapping frames by the following formula:

;

;

in, is the original signal; It is the signal after pre-emphasis; is the pre-emphasis coefficient; Indicates frame; is the frame length; It is the frame shift, which describes the overlap between adjacent frames.

Furthermore, the speech processing device performs window function processing on each frame signal, applies discrete Fourier transform, and then processes it through Mel filter to extract MFCC features. The formula is as follows:

;

;

;

in, No. The first samples; represents the discrete Fourier transform processed Frequency domain coefficients Represents the number of points of the discrete Fourier transform, which is also the number of samples in each speech frame; equal Represents the number of independent frequency components in the discrete Fourier transform; It is The Mel filter is The Mel filter is a set of overlapping triangular bandpass filters, which is used to simulate the frequency perception of the human ear and perform nonlinear Mel scale conversion on the frequency. is the number of Mel filters, representing the number of frequency bands divided on the Mel scale; In order to pass the The result of applying a Mel filter to the coefficients of the discrete Fourier transform and taking the logarithm represents the The logarithmic energy of a frequency band; For the Mel-frequency cepstral coefficients are The result of applying discrete cosine transform is to convert the logarithmic energy spectrum of the Mel filter into cepstral coefficients in the time domain, reduce the correlation between features, and highlight the spectral shape characteristics; is the number of MFCC features finally extracted.

Furthermore, the fundamental frequency characteristics Calculated by the following formula:

;

in, is the autocorrelation function; yes The peak position of is the sampling frequency; energy characteristics The calculation is done by the following formula:

;

in, For the The energy of a speech frame; the obtained feature vector is:

.

Furthermore, the pre-trained speech sentiment analysis model is a three-part support vector model, and its category labels are:

,

Represents three emotional features, category labels Each element in corresponds to an emotion category; the speech emotion analysis model is expressed using the following formula:

;

;

;

;

;

in, Represents the normal vector of the decision hyperplane, which is used to define the classification boundary; represents the bias term, also known as the intercept, which is a parameter in the support vector machine model used to shift the classification boundary; A weight vector representing an additional decision function, used to define additional classification boundaries; The bias term of the additional decision function is used to shift the additional classification boundary; represents the slack variable, indicating the degree of deviation from the hyperplane allowed; represents the regularization parameter, The importance of the slack variable is controlled. The larger its value, the more severe the penalty for misclassification. represents an indicator variable; is an additional decision function; the feature vectors of historical speech segments and the corresponding category labels are collected as model training test data, and a part of the model training test data is used as training data to train the speech sentiment analysis model. The training goal is to find an optimal decision boundary to correctly classify the samples in the training set to the greatest extent and maintain the generalization ability of the model; the part of the model training test data other than the training data is used as test data, and the trained support vector machine model is evaluated using the test data. The evaluation indicator is accuracy. If the accuracy exceeds the set accuracy threshold, the training is stopped, otherwise, the parameters of the model are adjusted and the training continues until the accuracy exceeds the set accuracy threshold.

Furthermore, additional decision functions Use the following formula to express it:

.

The emotional health analysis and support system for the elderly of the present invention has the following beneficial effects:

First, the present invention realizes comprehensive monitoring of the emotional state of the elderly by combining the voice signal collection device, processing device and emotion analysis device. The voice collection device can obtain the voice signal of the target elderly within a set time period, and perform preliminary analysis to determine whether emotion analysis is needed. The collected voice signal is balanced by the pre-emphasis filter of the processing device, and then divided into overlapping frames, and features such as MFCC, fundamental frequency and energy features are extracted to prepare for subsequent emotion analysis. The emotion analysis device uses a pre-trained voice emotion analysis model to analyze the extracted features and determine the emotional state of the elderly, thereby realizing real-time monitoring of the emotional health of the elderly.

Secondly, the present invention adopts a sentiment analysis model based on machine learning, combined with methods such as support vector machines, which can accurately judge the emotional state of the elderly. The pre-trained speech sentiment analysis model uses the collected speech features to effectively identify the emotional features in the speech signal, including positive emotions, neutral emotions and negative emotions, so as to accurately judge the emotional state of the elderly. Through the training and testing of the model, the model parameters can be continuously optimized, the accuracy and robustness of sentiment analysis can be improved, and reliable support can be provided for the monitoring of the emotional health of the elderly.

Thirdly, the present invention also provides a real-time intervention mechanism, which can send out early warning signals in time when the emotional state of the elderly is abnormal, and provide emotional intervention and support. By monitoring the results of emotional analysis, when the system detects that the elderly are in a negative emotional state, it can send out early warning signals in time to prompt relevant staff or family members to intervene and support. This timely intervention mechanism can effectively prevent the occurrence of emotional health problems in the elderly and improve their quality of life and sense of happiness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the system structure of the emotional health analysis and support system for the elderly provided in an embodiment of the present invention.

DETAILED DESCRIPTION

In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

Embodiment 1: Referring to FIG1 , an emotional health analysis and support system for the elderly, the system comprises: a speech acquisition device, a speech processing device and a speech emotion analysis device; the speech acquisition device is used to acquire the speech signal of the target elderly within a set time period, and perform preliminary signal analysis on the speech signal to determine whether speech emotion analysis is required, specifically comprising: statistically analyzing the average speech energy and speech frequency ratio of the speech signal; the average speech energy is defined as the ratio of the total energy of the speech signal to the time period within the set time period; the speech frequency ratio is defined as the ratio of the length of the speech signal to the time period within the set time period; if the average speech energy or the speech frequency ratio are both within their respective corresponding threshold ranges, it is determined that speech emotion analysis is not required, otherwise, it is determined that speech emotion analysis is required; the speech processing device is used to perform preliminary signal analysis on the speech signal during the speech acquisition process; When the sound collection device determines that speech emotion analysis is needed, a pre-emphasis filter is applied to the collected speech signal to balance the spectrum to obtain a preprocessed signal, MFCC features, fundamental frequency features and energy features are extracted from the preprocessed signal as elements in a feature vector to form a feature vector, and based on the feature vector, the preprocessed signal is divided into speech segments and non-speech segments using a zero crossing rate method; the speech emotion analysis device is used to perform emotion analysis on the feature vector of the speech segment using a pre-trained speech emotion analysis model to determine the emotion characteristics of the speech segment, the emotion characteristics include: positive emotion characteristics, neutral emotion characteristics and negative emotion characteristics, if the ratio of the total number of frames of the speech segment with negative emotion characteristics within a set time period to the length of the time period exceeds a set threshold, it is determined that the target elderly person is in a negative emotion, and a warning signal requiring emotional intervention is issued.

Specifically, the voice acquisition device uses a built-in microphone or an external device to capture the voice signal of the target elderly within a set time period. When the target elderly interacts with the system, or a voice signal occurs within a specific time period or within a monitoring period set by the system, the acquisition device will start working. The voice signal is transmitted to the system in the form of an analog electrical signal for further processing and analysis. The voice processing device receives the voice signal from the voice acquisition device, preprocesses it, and extracts features. In the preprocessing stage, a pre-emphasis filter is usually applied to balance the spectrum and eliminate the high-frequency part of the voice signal. Then, sound features such as Mel-frequency cepstral coefficients (MFCC), fundamental frequency features, and energy features are extracted from the preprocessed signal. These feature extraction processes can be implemented using digital signal processing technology.

MFCC features are a set of coefficients obtained by Fourier transforming the speech signal and then performing discrete cosine transform on the spectrum on the Mel frequency scale. These coefficients reflect the characteristics of the speech signal in the Mel frequency domain, including the harmonic structure and formant of the sound. MFCC features can well characterize the spectral structure and acoustic characteristics of the speech signal, and provide a relatively comprehensive description of the speech content of the speech signal. In sentiment analysis, MFCC features can capture information such as the sound quality and pitch changes of the speech signal, thereby providing important features that help judge emotions. Fundamental frequency refers to the most important frequency component in the sound, which determines the pitch of the sound. Fundamental frequency features are usually used to describe the pitch or tone of the sound. Fundamental frequency features can capture the pitch changes of the speech signal, such as ups and downs, changes in intonation, etc. These changes in tone are often closely related to the emotional state, such as high pitch when the mood is high and low pitch when the mood is depressed. Therefore, fundamental frequency features are of great significance for sentiment analysis. Fundamental frequency is the lowest frequency periodic oscillation in the speech signal, which usually corresponds to the pitch of the sound. Fundamental frequency features are usually used to describe the pitch of the speech signal or the height of the sound. The fundamental frequency feature is chosen because pitch is one of the important indicators of expressing emotions in speech signals. People tend to change the pitch of their voices in different emotional states, for example, the voice will be higher when they are happy and lower when they are frustrated. Therefore, the fundamental frequency feature can provide important clues about the emotional information in the speech signal. The energy feature is the energy distribution of the speech signal, which is usually used to describe the intensity or volume of the speech signal. The energy feature is chosen because emotional expression is usually accompanied by changes in the intensity and volume of the voice. For example, the voice will be louder when angry and lower when sad. Therefore, the energy feature can provide important information about the intensity of emotions in the speech signal.

Energy, fundamental frequency and MFCC represent different aspects of speech signals. Energy features reflect the intensity and amplitude changes of speech signals, fundamental frequency features reflect the pitch changes of speech signals, and MFCC features extract the spectral features of speech signals. The combination of these three features can more comprehensively capture various emotion-related information in speech signals, including the intensity of emotions, changes in pitch, and spectral characteristics of speech. Energy, fundamental frequency and MFCC each represent different aspects of speech signals, and their combination can improve the accuracy of the speaker's emotional state. For example, when the energy of the voice increases, the fundamental frequency becomes higher, and the spectral characteristics of the voice change, it means that the speaker is in an angry or excited emotional state; and when the energy of the voice decreases, the fundamental frequency becomes lower, and the spectral characteristics of the voice change, it means that the speaker is in a sad or depressed emotional state. Since each feature has its own unique way of expressing information, there is complementarity and independence between them. Therefore, the combination of these three features can enhance the robustness of the sentiment analysis system, reduce the sensitivity to environmental noise and individual differences of speakers, and improve the generalization ability and applicability of the system. The three features of energy, fundamental frequency and MFCC can be regarded as abstract representations of different aspects of speech signals. Combining these features can provide a more comprehensive and multi-dimensional feature space, which helps to better describe the emotional information in speech signals and improve the performance of sentiment analysis systems.

Embodiment 2: The speech collection device includes: a collection unit, an enhancement unit, a preliminary analysis unit and a noise separation unit; the collection unit is used to determine whether the voice signal emitted is emitted by the target elderly person through voice recognition within a set time period, and if so, to collect the voice signal; the enhancement unit is used to enhance the voice signal to obtain an enhanced voice signal; the preliminary analysis unit performs preliminary analysis of the voice signal to determine whether voice emotion analysis is required; the noise separation unit is used to separate background noise from the voice signal when it is determined that voice emotion analysis is required.

Specifically, voice recognition is used to determine whether the voice signal is emitted by the target elderly person, so as to avoid subsequent processing and analysis of some invalid voice signals. In daily life, there will be various irrelevant sounds. If all voice signals are analyzed and processed indiscriminately, it will cause a waste of system resources. The noise separation unit separates the background noise in the environment from the target voice signal through noise suppression algorithms, such as adaptive filters and spectral subtraction, which improves the accuracy and reliability of subsequent sentiment analysis and enables the system to better focus on the voice sentiment analysis of the target elderly person.

Performing a preliminary signal analysis on the speech signal to determine whether speech emotion analysis is required, specifically including: statistically analyzing the average speech energy and speech frequency ratio of the speech signal; the average speech energy is defined as the ratio of the total energy of the speech signal to the time period within a set time period; the speech frequency ratio is defined as the ratio of the length of the speech signal to the time period within a set time period; if the average speech energy or the speech frequency ratio are both within their respective corresponding threshold ranges, it is determined that speech emotion analysis is not required, otherwise, it is determined that speech emotion analysis is required. The average speech energy reflects the sound intensity of the speech signal within a certain period of time. By analyzing the average speech energy, the volume of the speech signal can be preliminarily understood, so as to determine whether the speaker is emitting a strong speech with emotional color. A high-energy speech signal indicates that the speaker is in an emotional state such as excitement or anger, while a low-energy speech signal indicates that the speaker is relatively calm or negative. The speech frequency ratio reflects the proportion of the speech signal in the total time period. By analyzing the speech frequency ratio, the speaker's speech activity and frequency can be preliminarily understood, that is, the number and frequency of the speaker's speech within a certain period of time. A higher proportion of speech frequency indicates that the speaker has greater emotional fluctuations and speaks frequently, while a lower proportion of speech frequency indicates that the speaker is relatively stable or inactive. Through the basic feature analysis of the speech signal, we can preliminarily understand the current emotional state of the speaker, including the intensity and activity of the emotion. If the energy and frequency of the speech signal are within a relatively calm range, it means that the speaker is relatively calm or normal, and no further emotional analysis is required. On the contrary, if the energy and frequency of the speech signal exceed the preset threshold range, it indicates that the speaker is in a state of excitement, anger, frustration, etc. At this time, further speech emotion analysis is required to better understand the speaker's emotional state and provide corresponding support and intervention. The enhancement unit enhances the collected speech signal by applying signal processing techniques, such as filtering and noise reduction, to eliminate existing noise, distortion and other interference, and obtain a clearer and more reliable enhanced speech signal, providing better input for subsequent analysis.

Embodiment 3: The method for enhancing the speech signal by the enhancement unit to obtain the enhanced speech signal includes: using the following formula to enhance the speech signal Perform short-time Fourier transform based on the autoregressive model to obtain the time-frequency representation :

;

The original speech signal is divided into several short time periods, and a window function is used to truncate each short time period. The role of the window function is to limit the time domain length of the signal in each time segment and gradually decay it to zero to reduce the spectrum leakage effect during time domain analysis. In each time segment, the time domain signal is converted to the frequency domain using Fourier transform to obtain the spectrum representation of the speech signal in each time period. This process makes it possible to observe the energy distribution of the speech signal at different frequencies, so as to better understand the acoustic characteristics of the speech signal. In addition, the second term in the formula is the part of the autoregressive model, which is used to capture the autocorrelation of the speech signal. The autoregressive model assumes that there is a linear relationship between the speech signal at the current moment and the signal at several previous moments. By solving the autoregressive coefficient, the autocorrelation information of the speech signal can be obtained, which further enhances the representation ability of the speech signal.

in, represents the time segment index of the short-time Fourier transform; represents the frequency index; is the window length for each time segment; is the window function; are the coefficients of the autoregressive model, is the order of the autoregressive model; is the imaginary number symbol; Time domain index; the time-frequency representation is enhanced using a Wiener filter with nonlinear dynamic range compression characteristics through the following formula:

;

Molecular part It represents the power spectrum of the original speech signal in the frequency domain, that is, the energy distribution of the speech signal at different frequencies. By calculating the power spectrum of the original speech signal, we can understand the strength of the speech signal at different frequencies. It represents the ratio between the original speech signal power spectrum and the noise power spectrum. Represents the power spectrum ratio of noise to speech signal. The larger the ratio, the higher the proportion of noise in the signal, and vice versa. The function of this part is to compare the power spectrum of signal and noise, so as to suppress the noise in the enhancement process. The Wiener filter is a classic signal processing filter with nonlinear dynamic range compression characteristics, which can effectively suppress noise and improve the quality of the signal. Part of it is to use the principle of Wiener filter, by comparing the power spectrum of the original speech signal with the power spectrum of the noise, to enhance the signal. and Represent the power spectrum estimation of noise and speech signal respectively; In order to enhance the frequency domain representation of speech signals, the enhanced frequency domain signals Perform inverse short-time Fourier transform to obtain the enhanced speech signal.

Embodiment 4: A noise separation unit, when determining that speech emotion analysis is required, separates background noise from a speech signal, comprising: representing the speech signal as a waveform function in the time domain ; Convert it to the frequency domain through short-time Fourier transform to get the frequency domain representation ; Let the length of the time period be , using length The window function is segmented, and the window length is , the overlap length between windows is ; Select the Hamming window as the window function and define the window function as:

;

in, represents the sampling index of the window; it is extended to a length of 1 by applying the window function to each time segment of the speech signal and applying zero padding , to obtain the window signal in the time domain ; For each window signal Applying discrete Fourier transform, we get the frequency domain representation ; Assume that the background noise is steady-state and linearly superimposed with the speech signal; use an adaptive filter in the frequency domain to model and estimate the background noise; Assume represents the clean spectrum of the speech signal, Represents the spectrum of the background noise; the frequency domain response of the adaptive filter is defined as:

;

in, It's in time The frequency domain response of the adaptive filter at ; using the following formula, using the adaptive filter Reconstruct the speech signal in the frequency domain to obtain the reconstructed signal:

;

in, To reconstruct the signal; convert the reconstructed signal back to the time domain to obtain the speech signal after separating the background noise from the speech signal .

Specifically, first, the original speech signal is represented as a waveform function in the time domain Then, the speech signal in the time domain is converted to the frequency domain through short-time Fourier transform to obtain the frequency domain representation The purpose of this step is to convert the speech signal from the time domain to the frequency domain so that it can be processed in the frequency domain. The speech signal represented in the frequency domain is segmented, and the length of each segment is , and apply the Hamming window function to reduce the spectral leakage effect. The purpose of this step is to divide the speech signal into multiple window segments so that each window segment can be processed independently later, and the spectral leakage effect can be reduced by applying the window function. Zero padding is applied to each window signal to extend it to a length of , and then perform discrete Fourier transform on the expanded window signal to obtain the frequency domain representation The purpose of this step is to convert each window signal into the frequency domain for processing in the frequency domain. Assuming that the background noise is steady-state and linearly superimposed with the speech signal, an adaptive filter in the frequency domain is used to model and estimate the background noise. The purpose of this step is to estimate the spectral characteristics of the background noise so that it can be separated and processed later. According to the clean spectrum of the speech signal and the spectrum of background noise , calculate the frequency domain response of the adaptive filter The purpose of this step is to calculate the frequency domain response of the adaptive filter based on the spectral characteristics of the speech signal and background noise, so as to reconstruct the speech signal in the subsequent processing. The frequency domain response of the adaptive filter is used to reconstruct the speech signal in the frequency domain to obtain the reconstructed signal The purpose of this step is to denoise the speech signal based on the frequency domain response of the adaptive filter, separate the background noise component, and thus obtain a cleaner speech signal. Convert the reconstructed signal back to the time domain to obtain a clean speech signal after separating the background noise from the speech signal. The purpose of this step is to convert the reconstructed signal represented in the frequency domain back to the time domain for subsequent speech sentiment analysis and other processing.

Embodiment 5: When the speech acquisition device determines that speech emotion analysis is required, the speech processing device applies a pre-emphasis filter to the acquired speech signal to balance the spectrum, and then divides the speech signal into overlapping frames according to the following formula:

;

;

in, is the original signal; It is the signal after pre-emphasis; is the pre-emphasis coefficient; Indicates frame; is the frame length; It is the frame shift, which describes the overlap between adjacent frames.

Specifically, the formula represents the application of pre-emphasis filter. The pre-emphasis filter is used to balance the spectrum of the speech signal and enhance the energy of the high-frequency part, thereby improving the signal-to-noise ratio of the speech signal. In this formula, represents the clean speech signal after background noise separation. Represents the pre-emphasized speech signal, is the pre-emphasis coefficient. The function of the pre-emphasis filter is to reduce the attenuation of the high-frequency part of the voice signal during transmission, thereby improving the clarity and recognizability of the voice signal. It represents the process of dividing the pre-emphasized speech signal into overlapping frames. In this formula, Indicates frame, is the frame length, is the frame shift (or frame shift step), which indicates the overlap length between adjacent frames. By dividing the speech signal into frames, the speech signal can be processed in segments in time, so that the speech signal in each frame can be considered as a short-term steady state, making it easier to perform subsequent feature extraction and analysis.

Embodiment 6: The speech processing device performs window function processing on each frame signal, applies discrete Fourier transform, and then processes it through Mel filter to extract MFCC features. The formula is as follows:

;

;

;

in, No. The first samples; represents the discrete Fourier transform processed frequency domain coefficients; Represents the number of points of the discrete Fourier transform, which is also the number of samples in each speech frame; equal Represents the number of independent frequency components in the discrete Fourier transform; It is The Mel filter is The Mel filter is a set of overlapping triangular bandpass filters, which is used to simulate the frequency perception of the human ear and perform nonlinear Mel scale conversion on the frequency. is the number of Mel filters, representing the number of frequency bands divided on the Mel scale; In order to pass the The result of applying a Mel filter to the coefficients of the discrete Fourier transform and taking the logarithm represents the The logarithmic energy of a frequency band; For the Mel-frequency cepstral coefficients are The result of applying discrete cosine transform is to convert the logarithmic energy spectrum of the Mel filter into cepstral coefficients in the time domain, reduce the correlation between features, and highlight the spectral shape characteristics; is the number of MFCC features finally extracted.

Specifically, first, the discrete Fourier transform (DFT) part in the formula plays the role of converting the time domain speech signal into a frequency domain representation. By performing a discrete Fourier transform on each frame of the speech signal, the speech signal can be analyzed in the frequency domain to obtain the amplitude and phase information of different frequency components. This step is crucial for speech emotion analysis because different emotional states are often reflected in the spectral structure of the speech signal, such as anger causing an increase or decrease in the sound frequency, while sadness causes a change in the sound frequency. Next, the Mel filter processing part maps the frequency domain signal after the discrete Fourier transform to the Mel frequency scale. The Mel frequency scale is more in line with the perceptual characteristics of the human auditory system and can better capture the important frequency components in the speech signal. The Mel filter divides the spectrum into a series of triangular bandpass filters and weights the energy in different frequency ranges to more accurately represent the human auditory perception of different frequencies. The role of this step is to highlight the important frequency components of the speech signal while suppressing the unimportant frequency components, thereby improving the distinguishability and robustness of the feature. Finally, the MFCC feature extraction part converts the frequency domain energy spectrum processed by the Mel filter into a set of cepstrum coefficients. These cepstral coefficients reflect the spectral characteristics of the speech signal on the Mel frequency axis, including the sound color and resonance characteristics of the speech signal. By using discrete cosine transform (DCT), the correlation between features can be reduced and the frequency domain energy spectrum can be converted into a more compact and stable feature representation. MFCC features have good discriminability and robustness, and are suitable for speech sentiment analysis and other speech processing tasks.

The first part of the formula For each speech frame Perform discrete Fourier transform (DFT). Discrete Fourier transform is a technique for converting time domain signals to frequency domain. It converts speech signals from time domain representation to frequency domain representation and obtains frequency domain coefficients. ,in represents the frequency index, represents the number of samples per speech frame. Next, the second part of the formula The figure shows the processing of the Mel filter. The Mel filter is a set of filters used to mimic the frequency perception of the human ear. It performs a nonlinear Mel scale transformation on the frequency axis. Here, Indicates The Mel filter is The gain at each frequency point, is the number of independent frequency components in the discrete Fourier transform. Through the processing of the Mel filter, the frequency axis can be converted to the Mel scale, which is more in line with the auditory characteristics of the human ear, and the logarithmic energy spectrum is obtained. Finally, the third part of the formula It shows the MFCC feature extraction process. MFCC is a commonly used method for extracting speech signal features. It converts the logarithmic energy spectrum into Applying discrete cosine transform, a set of cepstral coefficients, or MFCCs, are obtained. These coefficients can reduce the correlation between features and highlight the spectral shape characteristics of speech signals, providing a more recognizable feature representation.

Example 7: Baseband Characteristics Calculated by the following formula:

;

in, is the autocorrelation function; yes The peak position of is the sampling frequency; energy characteristic The calculation is done by the following formula:

;

in, For the The energy of a speech frame; the obtained feature vector is:

.

Specifically, first, the fundamental frequency characteristics The calculation is done by the autocorrelation function The autocorrelation function is a function that describes the correlation between a signal and its delayed version. At different delays The autocorrelation value under , the peak position of the autocorrelation function can be found . Fundamental frequency characteristics The sampling frequency is defined as The peak position of the autocorrelation function The fundamental frequency characteristic In speech signal analysis, it is usually used to represent the basic frequency of speech, that is, the pitch of the sound. The calculation is done by calculating the energy of the speech frame The energy is obtained by taking the difference of It is a speech frame The sum of the squares of the amplitudes of all samples in represents the overall energy of the speech frame. Energy characteristics It is the difference between the energies of adjacent speech frames, which is used to indicate the energy changes between speech frames. Energy characteristics In speech signal analysis, it is often used to represent the energy changes of speech, such as the intensity changes of the sound or the recognition of pauses and active parts of speech. Finally, the feature vector obtained is Including fundamental frequency characteristics , Energy characteristics , and a series of cepstral coefficients extracted by MFCC These MFCC coefficients reflect the spectral characteristics of the speech signal on the Mel frequency axis, and have good discrimination and robustness. By combining the fundamental frequency characteristics, energy characteristics and MFCC coefficients, the spectral and energy characteristics of the speech signal can be more comprehensively described, providing richer feature information for subsequent speech emotion analysis.

Example 8: The pre-trained speech sentiment analysis model is a three-part support vector model, and its category label is:

,

Represents three emotional features, category labels Each element in corresponds to an emotion category; the speech emotion analysis model is expressed using the following formula:

;

;

;

;

;

in, Represents the normal vector of the decision hyperplane, which is used to define the classification boundary; represents the bias term, also known as the intercept, which is a parameter in the support vector machine model used to shift the classification boundary; A weight vector representing an additional decision function, used to define additional classification boundaries; The bias term of the additional decision function is used to shift the additional classification boundary; represents the slack variable, indicating the degree of deviation from the hyperplane allowed; represents the regularization parameter, The importance of the slack variable is controlled. The larger its value, the more severe the penalty for misclassification. represents an indicator variable; is an additional decision function; the feature vectors of historical speech segments and the corresponding category labels are collected as model training test data, and a part of the model training test data is used as training data to train the speech sentiment analysis model. The training goal is to find an optimal decision boundary to correctly classify the samples in the training set to the greatest extent and maintain the generalization ability of the model; the part of the model training test data other than the training data is used as test data, and the trained support vector machine model is evaluated using the test data. The evaluation indicator is accuracy. If the accuracy exceeds the set accuracy threshold, the training is stopped, otherwise, the parameters of the model are adjusted and the training continues until the accuracy exceeds the set accuracy threshold.

Specifically, Example 8 describes a support vector machine (SVM) model for speech sentiment analysis, the goal of which is to The emotion classification is carried out, and the emotion categories are divided into three categories: negative emotion ( ), neutral sentiment ( ) and positive emotions ( ). The model learns an optimal decision boundary to distinguish speech of different emotion categories by optimizing an objective function. First, the objective function consists of two parts. The first part is to minimize the weight vector The square norm of , the goal of this part is to find a suitable decision hyperplane to maximize the correct classification of samples in the training set. At the same time, in order to avoid overfitting the training data, the objective function also includes a regularization term ,in is a regularization parameter used to control the complexity of the model, and is a slack variable, indicating that a certain degree of misclassification is allowed. The second part is a description of the constraints. The constraints ensure that the classification of the training samples on the decision boundary is correct. For each training sample, its feature vector The corresponding category label The distance between the inner product and the boundary is at least Such constraints ensure that most samples are correctly classified, and at the same time, by introducing slack variables A certain degree of classification error is allowed. Additional constraints include an auxiliary decision function , which is used in parallel with the main decision function to improve the classification performance of the model. This auxiliary decision function is also subject to similar constraints to ensure that it is on the boundary of the correct classification. In addition, is an indicator variable, according to the category label The value of determines whether additional constraints are enabled. If the category label For neutral emotions, Set to 1, indicating that additional constraints are enabled, otherwise set to 0, indicating that additional constraints are not enabled. During the model training process, the feature vectors of the historical speech segments and the corresponding category labels are used as training and testing data. This part of the data is divided into training data and test data, where the training data is used to train the model and the test data is used to evaluate the performance of the model. The evaluation indicator is usually the accuracy, that is, the proportion of samples that are correctly classified. If the accuracy reaches the set threshold, the training is stopped, otherwise the model parameters are adjusted and the training continues until the set accuracy threshold is reached.

Example 9: Additional decision functions Use the following formula to express it:

.

Specifically, additional decision functions The introduction of is to increase the classification of a category. In the support vector machine (SVM) model, the original decision function can usually only handle classification problems of two categories, but in practical applications, sometimes more emotional categories need to be processed. The role of the additional decision function is to enable the support vector machine model to handle classification problems of more than two categories by introducing additional classification boundaries, thereby improving the applicability and performance of the model. Specifically, the additional decision function defines an additional weight vector and the bias term , providing an additional classification boundary for each additional emotion category. These additional classification boundaries together with the original classification boundaries constitute a multi-category classification model, which enables the model to distinguish and identify multiple different emotion categories at the same time. By adjusting the parameters of the additional classification boundaries, the classification accuracy of each emotion category can be adjusted, thereby achieving more accurate emotion classification.

As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (9)

1. An emotional health analysis and support system for the elderly, characterized in that the system comprises: a speech acquisition device, a speech processing device and a speech emotion analysis device; the speech acquisition device is used to acquire the speech signal of the target elderly within a set time period, and perform preliminary signal analysis on the speech signal to determine whether speech emotion analysis is needed, specifically including: statistically analyzing the average speech energy and speech frequency ratio of the speech signal; the average speech energy is defined as the ratio of the total energy of the speech signal to the time period within the set time period; the speech frequency ratio is defined as the ratio of the length of the speech signal to the time period within the set time period; if the average speech energy or the speech frequency ratio are both within the corresponding threshold range, it is determined that speech emotion analysis is not needed, otherwise, it is determined that speech emotion analysis is needed; the speech processing device is used to perform a statistical analysis on the speech signal in the speech When the acquisition device determines that speech emotion analysis is needed, a pre-emphasis filter is applied to the collected speech signal to balance the spectrum to obtain a preprocessed signal, MFCC features, fundamental frequency features and energy features are extracted from the preprocessed signal as elements in a feature vector to form a feature vector, and based on the feature vector, the preprocessed signal is divided into speech segments and non-speech segments using a zero crossing rate method; the speech emotion analysis device is used to perform emotion analysis on the feature vector of the speech segment using a pre-trained speech emotion analysis model to determine the emotion characteristics of the speech segment, the emotion characteristics include: positive emotion characteristics, neutral emotion characteristics and negative emotion characteristics, if the ratio of the total number of frames of the speech segment with negative emotion characteristics within a set time period to the length of the time period exceeds a set threshold, it is determined that the target elderly person is in a negative emotion, and a warning signal requiring emotional intervention is issued. 2. The emotional health analysis and support system for the elderly as described in claim 1 is characterized in that the voice collection device includes: a collection unit, an enhancement unit, a preliminary analysis unit and a noise separation unit; the collection unit is used to determine whether the voice signal emitted is emitted by the target elderly person through voice recognition within a set time period, and if so, to collect the voice signal; the enhancement unit is used to enhance the voice signal to obtain an enhanced voice signal; the preliminary analysis unit performs preliminary analysis of the voice signal to determine whether voice emotion analysis is needed; the noise separation unit is used to separate background noise from the voice signal when it is determined that voice emotion analysis is needed. 3. The emotional health analysis and support system for the elderly as claimed in claim 2, characterized in that the enhancement unit performs signal enhancement on the voice signal to obtain the enhanced voice signal, comprising: using the following formula to enhance the voice signal Perform short-time Fourier transform based on the autoregressive model to obtain the time-frequency representation : ; in, represents the time segment index of the short-time Fourier transform; represents the frequency index; is the window length for each time segment; is the window function; are the coefficients of the autoregressive model, is the order of the autoregressive model; is the imaginary number symbol; Time domain index; the time-frequency representation is enhanced using a Wiener filter with nonlinear dynamic range compression characteristics through the following formula: ; in, and Represent the power spectrum estimation of noise and speech signal respectively; In order to enhance the frequency domain representation of speech signals, the enhanced frequency domain signals Perform inverse short-time Fourier transform to obtain the enhanced speech signal. 4. The emotional health analysis and support system for the elderly as claimed in claim 3, characterized in that the noise separation unit, when determining that speech emotion analysis is required, separates background noise from speech signals by: representing the speech signals as waveform functions in the time domain ; Convert it to the frequency domain through short-time Fourier transform to get the frequency domain representation ; Let the length of the time period be , using length The window function is segmented, and the window length is , the overlap length between windows is ; Select the Hamming window as the window function and define the window function as: ; in, represents the sampling index of the window; it is extended to a length of 1 by applying the window function to each time segment of the speech signal and applying zero padding , to obtain the window signal in the time domain ; For each window signal Applying discrete Fourier transform, we get the frequency domain representation ; Assume that the background noise is steady-state and linearly superimposed with the speech signal; use an adaptive filter in the frequency domain to model and estimate the background noise; Assume represents the clean spectrum of the speech signal, Represents the spectrum of the background noise; the frequency domain response of the adaptive filter is defined as: ; in, It's in time The frequency domain response of the adaptive filter at ; using the following formula, using the adaptive filter Reconstruct the speech signal in the frequency domain to obtain the reconstructed signal: ; in, To reconstruct the signal; convert the reconstructed signal back to the time domain to obtain the speech signal after separating the background noise from the speech signal . 5. The emotional health analysis and support system for the elderly as claimed in claim 4 is characterized in that the speech processing device, when the speech acquisition device determines that speech emotion analysis is required, applies a pre-emphasis filter to the collected speech signal to balance the spectrum by the following formula, and then divides the speech signal into overlapping frames: ; ; in, is the original signal; It is the signal after pre-emphasis; is the pre-emphasis coefficient; Indicates frame; is the frame length; It is the frame shift, which describes the overlap between adjacent frames. 6. The emotional health analysis and support system for the elderly as claimed in claim 5, characterized in that the speech processing device performs window function processing on each frame signal, applies discrete Fourier transform, and then processes it through Mel filter to extract MFCC features, and the formula is as follows: ; ; ; in, No. The first samples; represents the discrete Fourier transform processed frequency domain coefficients; Represents the number of points of the discrete Fourier transform, which is also the number of samples in each speech frame; equal Represents the number of independent frequency components in the discrete Fourier transform; It is The Mel filter is The Mel filter is a set of overlapping triangular bandpass filters, which is used to simulate the frequency perception of the human ear and perform nonlinear Mel scale conversion on the frequency. is the number of Mel filters, representing the number of frequency bands divided on the Mel scale; In order to pass the The result of applying a Mel filter to the coefficients of the discrete Fourier transform and taking the logarithm represents the The logarithmic energy of a frequency band; For the Mel-frequency cepstral coefficients are The result of applying discrete cosine transform is to convert the logarithmic energy spectrum of the Mel filter into cepstral coefficients in the time domain, reduce the correlation between features, and highlight the spectral shape characteristics; is the number of MFCC features finally extracted. 7. The emotional health analysis and support system for the elderly as claimed in claim 6, characterized in that the fundamental frequency feature Calculated by the following formula: ; in, is the autocorrelation function; yes The peak position of is the sampling frequency; energy characteristics The calculation is done by the following formula: ; in, For the The energy of a speech frame; the obtained feature vector is: . 8. The emotional health analysis and support system for the elderly as claimed in claim 7, characterized in that the pre-trained speech emotion analysis model is a three-point support vector model, and its category label is: , Represents three emotional features, category labels Each element in corresponds to an emotion category; the speech emotion analysis model is expressed using the following formula: ; ; ; ; ; in, Represents the normal vector of the decision hyperplane, which is used to define the classification boundary; represents the bias term, also known as the intercept, which is a parameter in the support vector machine model used to shift the classification boundary; A weight vector representing an additional decision function, used to define additional classification boundaries; The bias term of the additional decision function is used to shift the additional classification boundary; represents the slack variable, indicating the degree of deviation from the hyperplane allowed; represents the regularization parameter, The importance of the slack variable is controlled. The larger its value, the more severe the penalty for misclassification. represents an indicator variable; is an additional decision function; the feature vectors of historical speech segments and the corresponding category labels are collected as model training test data, and a part of the model training test data is used as training data to train the speech sentiment analysis model. The training goal is to find an optimal decision boundary to correctly classify the samples in the training set to the greatest extent and maintain the generalization ability of the model; the part of the model training test data other than the training data is used as test data, and the trained support vector machine model is evaluated using the test data. The evaluation indicator is accuracy. If the accuracy exceeds the set accuracy threshold, the training is stopped, otherwise, the parameters of the model are adjusted and the training continues until the accuracy exceeds the set accuracy threshold. 9. The emotional health analysis and support system for the elderly as claimed in claim 8, characterized in that the additional decision function Use the following formula to express it: .