A voice memory enhancement device and interaction method for cognitive impairment assistance

Abstract

The application relates to the technical field of speech recognition, in particular to a speech memory enhancement device for cognitive impairment assistance and an interaction method.The device comprises a speech preprocessing module, an acoustic segmentation module, a complexity evaluation module, a matching retrieval module and a memory feedback module.In the application, waveform cutting and frequency band energy calculation are performed on an interactive speech stream to construct an acoustic feature matrix, a boundary coordinate is generated by using feature difference sequence expansion direction reasoning to extract a coherent segmentation sequence, an arrangement entropy numerical value is calculated based on a frequency domain component barycenter position, a distance regularization is performed to find a minimalization path to construct a target playback address linked list by combining a mapping index, a digital code is read to be converted into a continuous analog parameter to perform driving amplification to generate a prompt audio, speech implicit features are extracted to replace a traditional subjective rough division mode, multi-dimensional distance regularization is used to realize automatic retrieval of historical memory fragments, and a compensation signal meeting the perception demand is outputted through dynamic conditioning to eliminate the mechanical feedback stiffness.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and in particular to a speech memory enhancement device and interaction method for assisting those with cognitive impairment. Background Technology

[0002] The field of speech recognition technology mainly involves methods and devices for converting human speech signals into processable text or instructions. It covers speech signal acquisition by acquiring sound waves through a microphone and converting them into electrical signals, framing and windowing continuous speech to form short speech segments, extracting acoustic feature sequences containing formant frequency energy distribution and cepstral features, comparing and probabilistically calculating the feature sequences frame by frame based on the matching relationship between acoustic models and language models, and outputting the corresponding text sequence in combination with dictionary rules. It also includes the process of speech synthesis to convert text into speech feedback and the organization of speech input and output in human-computer interaction.

[0003] Among them, traditional speech memory enhancement devices and interaction methods for assisting cognitive impairment refer to continuously recording the user's speech input through wearable or fixed speech acquisition devices, storing the speech signal in segments according to time sequence as audio files, and manually annotating it according to preset time tags or keyword lists. When the user needs to recall, the corresponding audio segment can be called and played by pressing a button or touch. At the same time, the preset reminder content is synthesized into speech output at a fixed speed by combining simple voice prompts, and the user can assist in memory retrieval and information prompts by repeating playback and manually selecting history records.

[0004] Existing technologies continuously record user voice input and store the signal in segments as audio files in chronological order. They rely on preset time tags for manual annotation and recall audio segments by pressing a button. This recording and coarse-grained segmentation operation mode makes it difficult to capture the deep acoustic features of speech. Relying solely on manual annotation brings heavy workload and subjective bias. At the same time, the feedback process can only output reminders at a fixed speech rate. The mechanical repetition cannot be dynamically adjusted to adapt to the perceptual state of specific groups, resulting in cumbersome memory recall and low accuracy of information compensation. Summary of the Invention

[0005] To address the challenges of existing technologies that continuously record user voice input and store the signal in chronological segments as audio files, relying on preset time tags for manual annotation and recalling audio segments via button presses, this recording and coarse-grained segmentation operation mode struggles to capture the deep acoustic features of speech. Excessive reliance on manual annotation leads to heavy workloads and subjective biases. Furthermore, the feedback mechanism can only output reminders at a fixed speaking speed, and the mechanical repetition cannot dynamically adjust to the perceptual state of specific groups, resulting in cumbersome memory recall and low accuracy in information compensation. Therefore, this invention provides a voice memory enhancement device and interaction method for assisting those with cognitive impairments.

[0006] On the one hand, a speech memory enhancement device for assisting cognitive impairment is provided, the device comprising: The speech preprocessing module acquires the interactive speech data stream received by the microphone array, divides it into a fixed time window to generate a discrete waveform frame sequence, performs Fourier transform and energy spectral density calculation, and extracts the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. The acoustic segmentation module performs differential operation on adjacent frames based on the acoustic feature matrix to establish a differential feature sequence, uses a long short-term memory network to perform evolution direction inference calculation to construct hidden layer state probability values, compares them with a preset segmentation threshold to extract coordinate intervals, and generates an acoustically stable segmentation sequence. The complexity assessment module extracts the local frequency domain component set based on the acoustically stable segmented sequence, extracts the frequency centroid position, and applies the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value. The matching and retrieval module extracts the historical feature cepstral matrix from the structured mapping index table based on the syllable arrangement entropy value, calculates the distortion distance between the local frequency domain component set and the historical feature cepstral matrix using the dynamic time warping algorithm, and sorts them in ascending order to construct a memory target replay index. The memory feedback module reads digital voice data based on the memory target playback index, inputs it into the digital-to-analog converter chip to generate an analog voltage signal, and then amplifies and converts the analog voltage signal through the speaker hardware driver circuit to generate a memory compensation prompt audio.

[0007] As a further aspect of the present invention, the acoustic feature matrix includes a frame index, a frequency band number, and cepstral coefficient values; the acoustically stable segmented sequence includes start coordinates, end coordinates, and segment duration; the local frequency domain component set includes frequency band energy distribution, harmonic components, and formant positions; the syllable arrangement entropy value is specifically a scalar entropy value that quantifies the complexity of syllables; the memory target playback index includes a matching priority order, cumulative distortion cost, and digital speech data address; and the memory compensation prompt audio includes audio content, sound pressure level, and playback sequence.

[0008] As a further aspect of the present invention, the speech preprocessing module includes: The waveform segmentation module acquires the interactive voice data stream received by the microphone array, and performs periodic segmentation processing on the voice data stream according to the preset time window length. It filters out continuous usable data segments whose amplitude exceeds the preset silent state amplitude reference value and generates a discrete waveform frame sequence. The frequency domain calculation submodule calls the discrete waveform frame sequence to perform discrete Fourier band transform operation, extracts the frequency band amplitude parameter of a single frame waveform, calculates the ratio of the square value of the frequency band amplitude parameter of a single frame waveform to the number of discrete sampling points of a single frame waveform, and generates the waveform energy spectral density value. The feature construction submodule maps the waveform energy spectral density value and the discrete waveform frame sequence to the Mel scale filter bank, extracts the Mel frequency cepstral coefficients and retains the values ​​that are greater than the preset environmental noise floor measurement threshold, and arranges them according to the sequential timing rules to generate an acoustic feature matrix.

[0009] As a further aspect of the present invention, the preset silent state amplitude reference value is obtained by extracting historical environmental audio sample data under non-interactive environmental conditions, calculating the arithmetic mean of the absolute values ​​of the amplitude of the historical environmental audio sample data, adding the arithmetic mean of the absolute values ​​of the amplitude to a set fixed amplitude offset parameter, and outputting the preset silent state amplitude reference value. The preset environmental noise floor measurement threshold is calculated and output by extracting the frequency band power distribution statistics of the background environmental noise data in the corresponding frequency band during non-interactive periods, multiplying the frequency band power distribution statistics with the set noise tolerance gain coefficient, and then outputting the preset environmental noise floor measurement threshold.

[0010] As a further aspect of the present invention, the acoustic segmentation module includes: The differential feature calculation submodule reads feature vector data within adjacent speech time frames based on the acoustic feature matrix, performs multidimensional matrix subtraction difference operation, extracts the calculated numerical result items, arranges and combines them according to the corresponding original acquisition time, and generates a differential feature sequence. The hidden state inference submodule calls the differential feature sequence and substitutes it into the set gated loop calculation processing logic to perform hidden state evolution inference operation. It calculates the numerical output of the dot product of the current forgetting weight and the previous memory state parameter, and adds the numerical output of the dot product of the current input gate activation value and the current candidate state parameter to generate the current memory cell state. It calculates the hidden state parameter through the output gate and the current memory cell state and obtains the hidden state probability value through the activation function mapping. The boundary determination segmentation module compares the probability value of the hidden layer state with the preset segmentation threshold, extracts the timestamp index node corresponding to the segment that crosses the preset segmentation threshold, performs coherence merging processing, extracts the time truncation coordinates of the beginning and end of the segment, and generates an acoustically stable segmented sequence.

[0011] As a further aspect of the present invention, the preset segmentation threshold is determined by extracting a set number of historical speech training samples as input to the global hidden layer probability mean output by the set gated loop calculation processing logic, and then superimposing the product of the probability distribution standard deviation corresponding to the global hidden layer probability mean and the set weight coefficient.

[0012] As a further aspect of the present invention, the complexity evaluation module includes: The frequency band component extraction submodule obtains the discrete spectral parameters within the extraction time window of the acoustically stable segmented sequence, performs a bandwidth limiting truncation and screening operation to collect the energy amplitude parameters corresponding to each frequency node in the specified frequency band interval, and establishes a local frequency domain component set. The frequency centroid extraction submodule calls the energy amplitude parameters of each item in the local frequency domain component set, uses the energy amplitude parameters as weighting coefficients to perform weighted and scale normalization processing on the corresponding frequency node index values, extracts the signal frequency domain energy distribution features, and obtains the frequency centroid coordinate values. The permutation entropy calculation submodule performs phase space delay reconstruction processing on the frequency centroid coordinate values, extracts the distribution frequency index of various mode states, calculates the corresponding proportion probability value based on the distribution frequency index, performs information quantization derivation on the proportion probability value to extract sequence complexity features, and generates syllable permutation entropy value.

[0013] As a further aspect of the present invention, the matching retrieval module includes: The mapping retrieval submodule queries the structured mapping index table based on the syllable arrangement entropy value, calculates the absolute value of the relative deviation between the syllable arrangement entropy value and the value of a single node record in the structured mapping index table, extracts the frequency order data parameter in the node item corresponding to the minimum value, and establishes a historical feature cepstral matrix. The distance normalization submodule calls the local frequency domain component set and the historical feature cepstral matrix, calculates the Euclidean distance difference between the component data in the local frequency domain component set and the data parameters in the historical feature cepstral matrix, constructs a two-dimensional aligned grid, finds the path to minimize the cumulative distance of nodes and counts the sum of values ​​along the path, and calculates the distortion distance. The index generation submodule performs numerical comparison operations on the distortion distance to generate an ascending-order data linked list. Based on the values ​​of each item in the ascending-order data linked list, it associates the original node identifiers, assigns retrieval order numbers, and performs parameter recombination and integration processing to construct a memory target replay index.

[0014] As a further aspect of the present invention, the memory feedback module includes: The voice reading submodule extracts the storage address addressing instruction based on the memory target replay index, calls the digital voice data in the preset sample dataset, performs a matching and filtering operation based on the digital voice data and the addressing instruction, extracts the digital coding parameters corresponding to each frequency node, and generates a discrete voice sampling sequence. The digital-to-analog decoding submodule performs amplitude continuous reconstruction on the digital-to-analog converter chip for the input of the discrete speech sampling sequence. It performs division operation on each digital coding parameter in the discrete speech sampling sequence with a preset amplitude reference, calculates the pulse modulation duty cycle coefficient and performs integration operation to generate analog continuous voltage amplitude. The power amplifier submodule calls the analog continuous voltage amplitude input speaker hardware driver circuit to perform gain adjustment operation, calculates the output power parameter by dividing the square of the analog continuous voltage amplitude by the hardware equivalent load impedance value, performs bias conditioning configuration based on the output power parameter, and generates memory compensation prompt audio.

[0015] On the other hand, a speech memory enhancement interaction method for assisting cognitive impairment, wherein the speech memory enhancement interaction method for assisting cognitive impairment is executed based on the aforementioned speech memory enhancement device for assisting cognitive impairment, includes the following steps: S1: Acquire the interactive voice data stream received by the microphone array, divide it into a fixed time window to generate a discrete waveform frame sequence, perform Fourier transform and energy spectral density calculation, and extract the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. S2: Based on the acoustic feature matrix, perform differential operation on adjacent frames to establish a differential feature sequence, use a long short-term memory network to perform evolution direction reasoning calculation to construct hidden layer state probability values, compare with a preset segmentation threshold to extract coordinate intervals, and generate an acoustically stable segmented sequence. S3: Based on the acoustically stable segmented sequence, extract the local frequency domain component set, extract the frequency centroid position, and apply the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value; S4: Based on the numerical query of the syllable arrangement entropy, extract the historical feature cepstral matrix from the structured mapping index table, use the dynamic time warping algorithm to calculate the distortion distance between the local frequency domain component set and the historical feature cepstral matrix and sort them in ascending order to construct a memory target replay index; S5: Based on the memory target playback index, read digital voice data, input the digital-to-analog converter chip to generate an analog voltage signal, and use the speaker hardware driver circuit to amplify and convert the analog voltage signal to generate a memory compensation prompt audio.

[0016] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: For interactive speech streams, waveform segmentation and frequency band energy calculation are performed to construct an acoustic feature matrix. The feature difference sequence is used to expand directional reasoning to extract boundary coordinates and generate a coherent segmented sequence. The permutation entropy value is calculated based on the centroid position of the frequency domain components. The distance regularization is performed in combination with the mapping index to find the minimum path and construct the target playback address linked list. The digital encoding is read and converted into continuous analog parameters to drive amplification and generate prompt audio. By extracting latent speech features to replace the traditional subjective coarse segmentation mode, multi-dimensional distance regularization is used to realize the automatic retrieval of historical memory fragments. The mechanical feedback is eliminated by dynamically adjusting the output compensation signal that meets the perception needs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the device of the present invention; Figure 2 This is a schematic diagram of the device frame of the present invention; Figure 3 This is a flowchart of the speech preprocessing module in this invention; Figure 4 This is a flowchart of the acoustic segmentation module in this invention; Figure 5 This is a flowchart of the complexity evaluation module in this invention; Figure 6 This is a flowchart of the matching and retrieval module in this invention; Figure 7 This is a flowchart of the memory feedback module in this invention; Figure 8 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] This invention provides a speech memory enhancement device for assisting those with cognitive impairments, such as... Figure 1-2 The diagram shown illustrates a speech memory enhancement device for assisting those with cognitive impairments. The device includes: The speech preprocessing module acquires the interactive speech data stream received by the microphone array, divides it into a fixed time window to generate a discrete waveform frame sequence, performs Fourier transform and energy spectral density calculation, and extracts the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. The acoustic segmentation module establishes a differential feature sequence by performing differential operations between adjacent frames based on the acoustic feature matrix, constructs the hidden layer state probability value by using a long short-term memory network for evolution direction inference calculation, compares it with a preset segmentation threshold to extract the coordinate interval, and generates an acoustically stable segmentation sequence. The complexity assessment module extracts the local frequency domain component set based on the acoustically stable segmented sequence, extracts the frequency centroid position, and applies the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value; The matching and retrieval module extracts the historical feature cepstral matrix from the structured mapping index table based on the numerical query of syllable arrangement entropy. It uses the dynamic time warping algorithm to calculate the distortion distance between the local frequency domain component set and the historical feature cepstral matrix and sorts them in ascending order to construct a memory target replay index. The memory feedback module reads digital voice data based on the memory target playback index, inputs it into the digital-to-analog converter chip to generate an analog voltage signal, and then amplifies and converts the analog voltage signal through the speaker hardware driver circuit to generate a memory compensation prompt audio.

[0022] The acoustic feature matrix includes frame index, frequency band number, and cepstral coefficient value; the acoustically stable segmented sequence includes start coordinates, end coordinates, and segment duration; the local frequency domain component set includes frequency band energy distribution, harmonic components, and formant positions; the syllable arrangement entropy value is specifically a scalar entropy value that quantifies the complexity of the syllable; the memory target playback index includes matching priority order, cumulative distortion cost, and digital speech data address; and the memory compensation cue audio includes audio content, sound pressure level, and playback timing.

[0023] Specifically, such as Figure 2 , 3 As shown, the speech preprocessing module includes: The waveform segmentation module acquires the interactive voice data stream received by the microphone array, and performs periodic segmentation processing on the voice data stream according to the preset time window length. It filters out continuous usable data segments whose amplitude exceeds the preset silent state amplitude reference value and generates a discrete waveform frame sequence. The waveform segmentation module extracts interactive voice data streams captured in real time by an 8-channel silicon microelectromechanical system (MEMS) microphone array located at the front end of the device, using a sampling frequency of 16000 Hz and a quantization depth of 16 bits, via a serial communication bus. Internally, the module uses a data buffer and an internal clock signal to truncate the continuously flowing voice data stream into independent digital audio slices every 256 milliseconds, with a preset time window length of 256 milliseconds. Subsequently, the module reads 500 historical environmental audio sample data segments (each 10 seconds long) from the memory, representing non-interactive periods. It then reads the quantized voltage values ​​corresponding to each sampling point in each historical environmental audio sample data segment, extracts the absolute values ​​of all quantized voltage values, accumulates all absolute values, and divides the sum by the total number of sampling points to calculate the arithmetic mean of the absolute amplitude of the historical environmental audio sample data. For example, when the cumulative absolute value of the extracted historical environmental audio sample data is 45000 mV and the total number of sampling points is 1000, the arithmetic mean of the absolute amplitude of the historical environmental audio sample data is calculated as 45 mV through division. Next, the module retrieves a pre-programmed fixed amplitude offset parameter from the flash memory chip. This parameter ranges from 10 mV to 30 mV, with a current setting of 15 mV. The arithmetic mean of the calculated 45 mV amplitude absolute value is added to the 15 mV fixed amplitude offset parameter, outputting a preset mute state amplitude baseline of 60 mV. For each 256-millisecond digital audio slice, the module extracts the amplitude values ​​of each sampling point within the slice, calculates the arithmetic mean of these amplitude values, and compares this arithmetic mean with the preset mute state amplitude baseline of 60 mV. If the arithmetic mean of a digital audio slice reaches or exceeds the preset mute state amplitude baseline of 60 mV, it is identified and selected as a continuous usable data segment. All selected continuous usable data segments are sequentially transcribed into static random access memory according to their time addresses, spliced ​​together to form continuous memory data blocks, and a discrete waveform frame sequence is generated.

[0024] The frequency domain calculation submodule calls the discrete waveform frame sequence to perform discrete Fourier band transform operation, extracts the frequency band amplitude parameter of a single frame waveform, calculates the ratio of the square value of the frequency band amplitude parameter of a single frame waveform to the number of discrete sampling points of a single frame waveform, and generates the waveform energy spectral density value. The frequency domain calculation submodule reads the previously generated discrete waveform frame sequence from the static random access memory via the internal high-speed data bus, and inputs each discrete waveform frame sequence into the core of the digital signal processing chip configured with 1024 sampling operation nodes. The chip's radix-2 time decimation algorithm logic is invoked to perform a discrete Fourier band transform operation on the discrete waveform frame sequence, converting the original time-series voltage amplitude into a complex distribution in the frequency dimension. During this operation, the real and imaginary parts of the complex data output at each frequency sampling point are extracted. The real and imaginary parts are squared, and the results of the two square operations are added together to extract the frequency band amplitude parameter of the single-frame waveform. Subsequently, the module reads the number of discrete sampling points for the single-frame waveform pre-registered in the configuration file. The specific value of this number depends on the cache capacity limit of the front-end hardware; currently, the number of discrete sampling points for the single-frame waveform is 512. This module obtains the calculated single-frame waveform frequency band amplitude parameter and squares it to obtain the square value. Then, it uses this squared value as the dividend and the number of discrete sampling points (512) of the single-frame waveform as the divisor to perform a division operation. For example, when the calculated result of the single-frame waveform frequency band amplitude parameter is 16, the module squares it to obtain 256, and then divides 256 by the number of discrete sampling points (512) of the single-frame waveform to calculate a ratio of 0.5. This module arranges the ratios of 0.5 calculated at each frequency node for each frame in ascending order of frequency and writes them into a specific register array to generate waveform energy spectral density values ​​characterizing the energy density of each frequency band.

[0025] The feature construction submodule maps the waveform energy spectral density value and discrete waveform frame sequence to the Mel scale filter bank, extracts the Mel frequency cepstral coefficients and retains the values ​​that are greater than the preset environmental noise floor measurement threshold, and arranges them according to the sequential time rule to generate an acoustic feature matrix. The feature construction submodule extracts the aforementioned waveform energy spectral density value and discrete waveform frame sequence, and synchronously transmits them to a Mel-scale filter bank structure containing 40 triangular bandpass interception logics. This module controls the Mel-scale filter bank to perform filtering and interception operations in a logarithmic distribution within the frequency range of 0 Hz to 8000 Hz. By multiplying the waveform energy spectral density value with the center response weight of each triangular bandpass interception logic and performing an accumulation operation, 40 frequency band energy output values ​​are obtained. Subsequently, a natural logarithmic operation is performed on these 40 frequency band energy output values, and the logarithmic operation result is substituted into the discrete cosine transform calculation matrix for matrix multiplication. The values ​​of the first 13 dimensions of the output matrix are extracted as Mel frequency cepstral coefficients. Simultaneously, background environmental noise data collected during 100 seconds of operation in a non-interactive period is read, and the average energy of the noise data in the corresponding 40 frequency bands is extracted as the frequency band power distribution statistics. The module reads the set noise tolerance gain coefficient from the system's internally defined environmental adaptability configuration document. This set noise tolerance gain coefficient has been calibrated through multiple outdoor environmental tests, and its value is fixed at 1.2. The aforementioned frequency band power distribution statistics are multiplied by the set noise tolerance gain coefficient of 1.2. For example, when the frequency band power distribution statistics for a certain frequency band are 10, 10 is multiplied by the set noise tolerance gain coefficient of 1.2, and the preset environmental noise floor measurement threshold is calculated to be 12. This module iterates through each of the previously extracted Mel-frequency cepstral coefficients, comparing the coefficient values ​​with the preset environmental noise floor measurement threshold of 12. All values ​​below 12 are discarded, retaining only Mel-frequency cepstral coefficient values ​​greater than the preset environmental noise floor measurement threshold. For the retained values, the accompanying original acquisition timestamp is read, and the data is arranged and filled column-wise in a two-dimensional data structure according to the chronological order, ultimately generating an acoustic feature matrix.

[0026] Specifically, such as Figure 2 , 4 As shown, the acoustic segmentation module includes: The differential feature calculation submodule, based on the extracted acoustic feature matrix, reads the feature vector data within adjacent speech time frames, performs multidimensional matrix subtraction difference operation, extracts the calculated numerical result items, arranges and combines them according to the corresponding original acquisition time, and generates a differential feature sequence. The differential feature calculation submodule reads the previously generated acoustic feature matrix from the front-end cache storage area. This acoustic feature matrix contains multiple 13-dimensional column vectors arranged in time frames. With a time step parameter set to 1, a loop traversal reading operation is performed on the data within the acoustic feature matrix, simultaneously reading the feature vector data corresponding to the current time frame and the feature vector data from the adjacent previous time frame. For the feature vector data from the adjacent speech time frames, the internal vector operation logic unit is invoked. The feature vector data of the current time frame is used as the minuend matrix, and the feature vector data of the previous time frame is used as the subtrahend matrix. Multidimensional matrix subtraction is performed dimension-by-dimensional to extract the first-order numerical result. Next, the feature vector data of the current time frame and the feature vector data of the previous time frame are similarly subjected to multidimensional matrix subtraction. The difference obtained from the two subtraction operations is further subtracted to extract the second-order numerical result. For example, when the feature value of a certain dimension in the current speech time frame is 8, and the corresponding feature value in the adjacent previous speech time frame is 5, a subtraction difference operation of 8 minus 5 is performed, and the first-order calculated value result is extracted as 3. Subsequently, the module obtains the first-order and second-order calculated value results for each time frame, extracts the corresponding original acquisition timestamp attribute bound within the result item, and arranges them in ascending order of timestamp values. In the horizontal dimension, the module combines and concatenates the original 13-dimensional feature vector, 13-dimensional first-order calculated value result, and 13-dimensional second-order calculated value result, expanding the feature dimension of a single time frame to 39 dimensions, and finally vertically concatenates to generate a differential feature sequence with continuous temporal correlation characteristics.

[0027] The hidden state inference submodule calls the differential feature sequence to substitute into the set gated loop calculation processing logic, performs the hidden state evolution inference operation, calculates the current forgetting weight to participate in the dot product of the previous memory state parameter, and superimposes the current input gate activation value and the dot product of the current candidate state parameter to generate the current memory cell state. The hidden state parameter is calculated through the output gate and the current memory cell state and mapped through the activation function to obtain the hidden state probability value. The hidden state inference submodule constructs and invokes the internally deployed gated recurrent computation processing logic. This logic network includes an input layer that receives 39-dimensional input, a recurrent hidden layer with 128 operational neurons, and an output layer. The input layer uses a fully connected topology to map and transmit the numerical values ​​within the previously generated difference feature sequence to the recurrent hidden layer. The recurrent hidden layer is configured with forget gate, input gate, and output gate logic. All neurons use a logistic nonlinear activation mechanism to control the output range of the signal, defining it between 0 and 1. Simultaneously, a hyperbolic tangent nonlinear activation mechanism is used to control the amplitude range of the cell candidate state, defining it between -1 and +1. The hidden state value from the previous time step is concatenated with the difference feature sequence input data from the current time step. Then, the concatenated matrix is ​​multiplied by the preset forget gate weight matrix in the system. The forget gate bias term is added to the multiplication result, and the resulting sum is substituted into the logistic nonlinear activation mechanism to calculate the current forget weight. Next, the current forgetting weight is multiplied by the previous memory state parameter (i.e., the cell state matrix of the previous time step) to produce a numerical output. For example, if the current forgetting weight is 0.8 after activation calculation, and the value of a node in the previous memory state parameter is 0.5, the operation of multiplying 0.8 by 0.5 is performed, resulting in a numerical output of 0.4. Simultaneously, this module performs a multiplication-addition operation on the input gate weight matrix and the current input difference feature sequence, and substitutes this into the activation mechanism to obtain the current input gate activation value parameter. Assume this current input gate activation value parameter is calculated to be 0.3. And through a similar operation, the current candidate state parameter is calculated to be 0.8. Then, the numerical output of the previous memory state parameter multiplied by the calculated current forgetting weight of 0.4 is added to the numerical output of the previous memory state parameter multiplied by the current input gate activation value parameter of 0.3 and the current candidate state parameter of 0.8, resulting in a numerical superposition operation. By adding 0.4 and 0.24, the value 0.64 is obtained, generating the current memory cell state. This current memory cell state of 0.64 is substituted into the hyperbolic tangent nonlinear activation mechanism and multiplied with the output gate state to calculate the hidden layer state parameters. After being mapped by the activation function, the hidden layer state probability values ​​of each time node are finally obtained in the output layer.

[0028] The boundary determination and segmentation module compares the probability values ​​of the hidden layer states with the preset segmentation thresholds, extracts the timestamp index nodes that cross the preset segmentation thresholds, performs coherence merging processing, extracts the time truncation coordinates of the beginning and end of the segments, and generates an acoustically stable segmented sequence. The boundary determination and segmentation module extracts a predetermined number of historical speech training samples from the system's built-in database. This predetermined number, fixed at 500 after long-term operational verification, is then fed into the aforementioned gated loop calculation logic to perform a full-process operation. All hidden layer probability values ​​generated by the output layer are collected, summed, and divided by the total number of time nodes in the samples to extract the global hidden layer probability mean. The dispersion of the probability values ​​relative to the mean is calculated by summing the squared differences. The variance is then calculated by dividing the sum by the corresponding number of nodes, and the square root is taken to extract the standard deviation of the probability distribution corresponding to the global hidden layer probability mean. A pre-recorded weighting coefficient, set to 1.5 according to the normal distribution confidence interval rule, is retrieved from the storage area. The extracted global hidden layer probability mean, probability distribution standard deviation, and weighting coefficient 1.5 are then loaded into the central processing unit's arithmetic unit. For example, when the calculated global hidden layer probability mean is 0.65 and the standard deviation of the probability distribution corresponding to the global hidden layer probability mean is 0.12, the module first multiplies 0.12 with a set weight coefficient of 1.5 to obtain a product of 0.18. Then, the module adds the global hidden layer probability mean of 0.65 to the aforementioned product 0.18, and uses this addition to determine the preset segmentation threshold as 0.83. For each hidden layer state probability value calculated in real-time, a numerical magnitude comparison is performed with the preset segmentation threshold of 0.83. When multiple consecutive hidden layer state probability values ​​are greater than 0.83 within a certain time period, the timestamp index nodes corresponding to the probability values ​​greater than the preset segmentation threshold are extracted. For groups of adjacent timestamp index nodes with a time interval of less than 50 milliseconds, the module performs a continuity merging process, grouping them into the same continuous segment. For each merged segment, the specific time values ​​of its start and end times are extracted as the time cutoff coordinates of the segment's beginning and end. Finally, based on the coordinates, the speech data of the corresponding time period is extracted to form independent audio blocks, generating an acoustically stable segmented sequence.

[0029] Specifically, such as Figure 2 , 5 As shown, the complexity evaluation module includes: The frequency band component extraction submodule obtains discrete spectral parameters within the time window of the acoustically stable segmented sequence extraction, performs a bandwidth limiting and truncation screening operation to collect the energy amplitude parameters corresponding to each frequency node in the specified frequency band interval, and establishes a local frequency domain component set. The frequency band component extraction submodule activates its internal time-series data stream analyzer to obtain the previously generated acoustically stable segmented sequence. For each independent audio block within the acoustically stable segmented sequence, a time window is defined based on the extracted segment start and end time truncation coordinates. The discrete spectral parameters cached during the original frequency domain calculation within the time window are extracted. Since the energy reflecting the core semantic information in the speech signal is mostly concentrated in the mid-to-low frequency region, this module reads the lower and upper frequency thresholds stored in the configuration register. The lower frequency threshold is set to 300 Hz, and the upper frequency threshold is set to 3400 Hz. This module uses these two frequency thresholds as truncation instructions and substitutes them into the bandwidth limiting truncation filtering operation logic to perform range comparison operations on all frequency sampling points within the discrete spectral parameters. For nodes with frequencies less than 300 Hz or greater than 3400 Hz, this module forcibly resets their corresponding energy amplitude parameters to 0. For each frequency node within the specified frequency band of 300 Hz to 3400 Hz, this module collects its corresponding energy amplitude parameters in their entirety. The module then performs array-based arrangement of the truncated and filtered frequency node indices and their corresponding energy amplitude parameters in ascending frequency order, and stores them in dedicated two-dimensional storage space in memory, thereby establishing a local frequency domain component set. For example, when a frequency node within the specified frequency band is 1000 Hz and its energy amplitude parameter is 25, the values ​​1000 and 25 are combined into a single record and stored in memory.

[0030] The frequency centroid extraction submodule calls the energy amplitude parameters of each component in the local frequency domain, uses the energy amplitude parameters as weighting coefficients to perform weighted and scale normalization processing on the corresponding frequency node index values, extracts the frequency domain energy distribution characteristics of the signal, and obtains the frequency centroid coordinate values. The frequency centroid extraction submodule reads the local frequency domain component set from the internal memory. It extracts the index values ​​of each frequency node and their corresponding energy amplitude parameters stored within the local frequency domain component set. Using each energy amplitude parameter as a weighting coefficient, it performs a multiplication operation on the corresponding frequency node index value to calculate the product of each pair of frequency nodes and energy amplitudes. Next, it performs a summation operation on all calculated product values ​​within the specified frequency band to extract the numerator of the signal frequency domain energy distribution feature. Simultaneously, it performs a separate summation operation on all energy amplitude parameters contained in the local frequency domain component set to extract the total energy amplitude as the denominator of the signal frequency domain energy distribution feature. The aforementioned numerator of the signal frequency domain energy distribution feature is divided by the denominator of the signal frequency domain energy distribution feature to perform weighted and scale normalization processing. For example, when the local frequency domain component set contains only three frequency nodes, 500 Hz, 1000 Hz, and 1500 Hz, and the corresponding energy amplitude parameters (i.e., weighting coefficients) are 10, 20, and 10, respectively. First, the product of frequency and energy amplitude is calculated, resulting in values ​​of 5000, 20000, and 15000. This module sums these three products to obtain a numerator of 40000. Then, the energy amplitude parameters 10, 20, and 10 are summed to obtain a denominator of 40. Finally, 40000 is divided by 40, yielding a result of 1000 Hz. This module outputs the core frequency point extracted based on weighted and scale-normalized processing, obtaining the frequency centroid coordinates.

[0031] The permutation entropy calculation submodule performs phase space delay reconstruction processing on the frequency centroid coordinate values, extracts the distribution frequency index of various mode states, calculates the corresponding proportion probability value based on the distribution frequency index, performs information quantization derivation on the proportion probability value to extract sequence complexity features, and generates syllable permutation entropy value. The permutation entropy calculation submodule receives a one-dimensional time series sequence composed of the frequency centroid coordinates output above. This module reads the system's preset embedding dimension parameter and time delay parameter. After multiple rounds of testing and calibration, the embedding dimension parameter is fixed at 3, and the time delay parameter is set to 1. Phase space delay reconstruction processing is performed on the one-dimensional sequence composed of frequency centroid coordinates. According to the extraction rules of embedding dimension parameter 3 and time delay parameter 1, the one-dimensional sequence is divided into multiple reconstruction vectors containing three consecutive coordinate elements. The three elements within each reconstruction vector are arranged in ascending order according to their numerical values, and the arrangement order of the original position indices is recorded. Since the embedding dimension is 3, there are a total of 6 permutations, i.e., mode states. This module traverses all reconstruction vectors, accumulates and extracts the distribution frequency index of each mode state. Subsequently, the distribution frequency index of a single mode state is divided by the total number of reconstruction vectors, and the corresponding percentage probability value is calculated based on the distribution frequency index. For example, when the total number of reconstructed vectors is 100, and the frequency index of a certain permutation pattern is 25, the corresponding proportion probability value of this pattern is calculated as 0.25 through division. Then, the internal logarithmic operation core is invoked to calculate the logarithm of each calculated proportion probability value of 0.25 to the base of the natural constant. The 0.25 is multiplied by its logarithm, and the negative value is extracted. Then, all the negative values ​​of this type of product calculated for the six pattern states are accumulated. Through information quantification derivation of the proportion probability values, this module uses the final accumulated result as a sequence complexity feature characterizing the drastic change in sound within that time period, generating a syllable permutation entropy value.

[0032] Specifically, such as Figure 2 , 6 As shown, the matching and retrieval module includes: The mapping retrieval submodule queries the structured mapping index table based on the syllable arrangement entropy value, calculates the absolute value of the relative deviation between the syllable arrangement entropy value and the value of a single node record in the structured mapping index table, extracts the frequency order data parameter in the node item corresponding to the minimum value, and establishes a historical feature cepstral matrix. The mapping retrieval submodule queries a structured mapping index table in the solid-state storage medium based on the previously generated syllable arrangement entropy values. This structured mapping index table consists of multiple preset record nodes, each containing a baseline entropy reference value and the corresponding cepstral feature distribution configuration parameters.

[0033] Table 1: Structured Mapping Index Table

[0034] As shown in Table 1, the structured mapping index table records in detail the reference values ​​of the baseline permutation entropy corresponding to different node identifier sequence numbers, as well as the corresponding frequency order data parameter allocation baseline and energy matching correction factor. All parameters together constitute a static mapping standard system for different speech state characteristics of patients with cognitive impairment.

[0035] This module retrieves the syllable arrangement entropy value calculated at the current moment and subtracts it sequentially from the single-node record value in the structured mapping index table, i.e., the reference value of the arrangement entropy. The absolute value of the subtraction result is extracted to calculate the absolute value of the relative deviation between the syllable arrangement entropy value and the single-node record value in the structured mapping index table. For example, if the currently received syllable arrangement entropy value is 1.95, it is subtracted from the reference value of 1.80 for node 2 in Table 1, and the absolute value is calculated to yield a relative deviation of 0.15. Simultaneously, it is subtracted from the reference value of 2.45 for node 3, yielding an absolute value of 0.50. A numerical comparison operation is performed on all calculated absolute relative deviation values, and the minimum value is extracted. In this example, 0.15 is the minimum value. This module locates the node item corresponding to the minimum value, i.e., node 2, and reads the frequency order data parameters bound within this node item, i.e., the feature matrix construction template under the full-band equalization mode. Based on the extracted frequency order data parameters, the historical data cached by the feature construction submodule is called and organized into a two-dimensional matrix. Finally, the historical feature cepstral matrix is ​​established and output.

[0036] The distance normalization submodule calls the local frequency domain component set and the historical feature cepstral matrix, calculates the Euclidean distance difference between the data parameters of each component in the local frequency domain component set and the data parameters in the historical feature cepstral matrix, constructs a two-dimensional aligned grid, finds the path to minimize the cumulative distance of nodes and counts the sum of values ​​along the path, and calculates the twist distance. The distance normalization submodule retrieves the local frequency domain component set established by the aforementioned frequency band component extraction submodule from the shared memory area, and simultaneously calls the historical feature cepstral matrix output by the aforementioned mapping retrieval submodule. The local frequency domain component set contains M time-dimensional feature columns, and the historical feature cepstral matrix contains N time-dimensional feature columns. An M-row, N-column two-dimensional grid matrix is ​​generated in the internal arithmetic unit. Each coordinate node of the two-dimensional grid matrix is ​​traversed, and numerical subtraction is performed between the component data at the corresponding time point in the local frequency domain component set and the data parameter at the corresponding time point in the historical feature cepstral matrix. The difference is squared and then square-rooted to calculate the Euclidean distance difference between each component data point and the data parameter. All calculated Euclidean distance differences are filled into the corresponding grid coordinate points to construct a two-dimensional aligned grid. Subsequently, the dynamic programming accumulation operation rule is invoked, starting from the starting coordinates of the two-dimensional aligned grid, and for each grid node, the accumulated distance values ​​of its left, upper, and upper-left nodes are compared. The node with the smallest cumulative distance is selected as the next optimal node. This minimum cumulative distance is then added to the current grid node's Euclidean distance difference, and the cumulative distance of the current grid node is updated. Through progressive comparison and addition, the module finds the path that minimizes the cumulative distance of nodes and calculates the sum of the Euclidean distances of all nodes along that path. For example, on a path of length 5, the Euclidean distance differences between nodes are 2, 3, 1, 4, and 2. The sum of these values ​​is calculated to be 12. This sum is then normalized by dividing the total path length of 5, yielding a value of 2.4. This module uses 2.4 as the final dynamic time bending distance metric to calculate and output the warping distance.

[0037] The index generation submodule performs numerical comparison operations on the distortion distance to generate an ascending data linked list. Based on the values ​​of each item in the ascending data linked list, it associates the original node identifiers to assign retrieval order numbers and performs parameter recombination and integration processing to build a memory target replay index. The index generation submodule establishes an empty linked list structure. It inputs all the distortion distances calculated for each speech block into the comparator of the central processing unit for numerical comparison. Using a quicksort algorithm, a distortion distance is randomly selected as a benchmark value. All distortions less than the benchmark are placed on the left, and those greater are placed on the right. Through internal recursive calls, the left and right subsequences are continuously divided and compared, ultimately arranging all distortions in ascending order to generate an ascending-order linked list. Each distortion distance in the ascending-order linked list is read, and the associated original node identifier is called according to the positional order of the values. The first associated original node identifier is assigned a sequence number 1, the second a sequence number 2, and so on, incrementing the search order. For example, if an associated original node identifier originally has an address code of 1024 in memory, but its corresponding distortion distance is the 3rd in the ascending-order linked list, the number 3 is used as the search order number to establish a mapping with its address code 1024. This module reads the mapping data containing the retrieval sequence number and memory address, performs parameter reorganization and integration processing, encapsulates it into a query data packet structure of a specific format, stores it in an internal register, and builds a memory target replay index.

[0038] Specifically, such as Figure 2 , 7 As shown, the memory feedback module includes: The speech reading submodule extracts the storage address addressing instruction based on the memory target replay index, calls the digital speech data in the preset sample dataset, performs a matching and filtering operation based on the digital speech data and the addressing instruction, extracts the digital coding parameters corresponding to each frequency node, and generates a discrete speech sampling sequence. The speech reading submodule extracts the previously constructed memory target playback index, reading the associated original node identifier with retrieval sequence number 1 and the corresponding memory address mapping record encapsulated in the index data packet. Based on the data format within the memory target playback index, it converts it into an extraction storage address addressing command that the controller can recognize via the bus interface. This addressing command is used to access an externally connected high-capacity flash memory chip to retrieve high-precision digital speech data recorded in a pre-set sample dataset. All digital speech data is stored in a 16-bit pulse code modulation format. This module temporarily stores the read digital speech data in a high-speed buffer. Subsequently, it reads the frequency node list fed back by the front-end frequency band component extraction submodule and performs a matching and filtering operation based on the digital speech data in the temporary storage area and the frequency band boundaries extracted by the addressing command. This module compares each frequency channel identifier of the digital speech data one by one, extracting only data segments whose digital coding parameters match the aforementioned specified frequency nodes. For example, when an addressing instruction specifies the extraction of digital voice data within the address range of 2048 to 4096, and the frequency nodes are required to cover 1000 Hz, this module scans the address range and extracts the 16-bit binary numerical representation stored at the corresponding location of the 1000 Hz frequency node, i.e., the corresponding digital encoding parameter. The extracted corresponding digital encoding parameters are then concatenated and combined according to their chronological order, stored continuously in memory, and finally generate a discrete voice sampling sequence.

[0039] The digital-to-analog decoding submodule performs amplitude continuous reconstruction on the digital-to-analog converter chip for the input of discrete speech sampling sequences. It performs division operations on each digital coding parameter in the discrete speech sampling sequence with a preset amplitude reference, calculates the pulse modulation duty cycle coefficient and performs integration operations to generate analog continuous voltage amplitude. The digital-to-analog decoding submodule incorporates a high-precision digital-to-analog converter (DAC) chip core circuit. This module inputs the previously generated discrete speech sampling sequence into the DAC chip via a serial peripheral interface protocol. The DAC chip's internal digital filter upsamples and filters the discrete digital signal, performing amplitude continuous reconstruction on the speech discrete sampling sequence input to the DAC chip. The reference voltage parameter value provided by the system power supply circuit is read as a preset amplitude reference, which is calibrated to 3.3 volts. The various digital encoding parameters contained in the speech discrete sampling sequence are extracted; these digital encoding parameters correspond to integers between 0 and 65535. Each digital encoding parameter is divided by the full-scale value of 65535 to obtain a normalized ratio. This ratio is then multiplied by the preset amplitude reference of 3.3 volts, or the various digital encoding parameters are directly divided by the corresponding division value of the preset amplitude reference. The specific operation performed here involves using the decimal values ​​of the digital encoding parameters within the discrete speech sampling sequence as the dividend and the full-scale value of 65535 as the divisor to perform a division operation, calculating the decimal representing the signal duty cycle. For example, when a digital encoding parameter's decimal value is 32767, this module divides 32767 by 65535, calculating a pulse modulation duty cycle coefficient of approximately 0.5. Then, the pulse width modulation generator inside the digital-to-analog converter chip generates a square wave signal based on this duty cycle coefficient of 0.5. This signal is fed into a low-pass filter network composed of resistors and capacitors. The low-pass filter network performs an integration operation on the square wave signal based on the charging and discharging characteristics of the capacitors. As the integration time accumulates, the high-frequency abrupt square wave is smoothed, generating an analog continuous voltage amplitude fluctuating around 1.65 volts at the circuit output.

[0040] The power amplifier submodule calls the analog continuous voltage amplitude input speaker hardware driver circuit to perform gain adjustment operation, calculates the output power parameter by dividing the square of the analog continuous voltage amplitude by the hardware equivalent load impedance value, performs bias conditioning configuration based on the output power parameter, and generates memory compensation prompt audio. The power amplifier submodule consists of an audio driver amplifier chip and its peripheral bias circuitry, with its core employing a high-efficiency Class-D amplification topology. The analog continuous voltage amplitude generated earlier is retrieved via copper traces on the printed circuit board and input to the preamplifier pin of the speaker hardware driver circuit for gain adjustment. This module reads the hardware amplification scaling factor, determined by the ratio of the external feedback resistor value, from the driver circuit; this scaling factor is set to 10. Under the control of its internal arithmetic logic unit, the module multiplies the analog continuous voltage amplitude by the hardware amplification scaling factor and then squares the result by the hardware equivalent load impedance to calculate the output power parameter. For example, if the currently received analog continuous voltage amplitude is 1.5 volts, the module multiplies 1.5 by 10 to calculate an amplified voltage of 15 volts. Subsequently, using the nominal impedance of the load speaker (8 ohms) as the hardware equivalent load impedance, the internal monitoring logic calculates and evaluates the current output power parameter. In this example, the square of the amplified 15 volts (225) is applied to the 8-ohm load impedance and divided to calculate the output power parameter, which is 28.125 watts. This calculated output power parameter of 28.125 watts is then compared to the system's preset safety threshold of 30 watts. Since the safety threshold is not exceeded, the module performs bias conditioning based on the output power parameter, controlling the gate drive voltage of the amplification stage MOSFET to reach the rated turn-on level, ensuring that a large current flows smoothly into the speaker voice coil. Under the action of alternating current, the voice coil cuts the magnetic field lines, driving the diaphragm to vibrate, ultimately converting the amplified electrical signal into airborne sound waves, generating memory compensation prompt audio for playback to patients with cognitive impairment.

[0041] Please see Figure 8 The speech memory enhancement interaction method for assisting cognitive impairment is executed based on the aforementioned speech memory enhancement device for assisting cognitive impairment, and includes the following steps: S1: Acquire the interactive voice data stream received by the microphone array, divide it into a fixed time window to generate a discrete waveform frame sequence, perform Fourier transform and energy spectral density calculation, and extract the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. S2: Based on the acoustic feature matrix, perform differential operation on adjacent frames to establish a differential feature sequence, use a long short-term memory network to perform evolution direction reasoning calculation to construct hidden layer state probability values, compare with a preset segmentation threshold to extract coordinate intervals, and generate an acoustically stable segmented sequence. S3: Based on the acoustically stable segmented sequence, extract the local frequency domain component set, extract the frequency centroid position, and apply the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value; S4: Based on the numerical query of the syllable arrangement entropy, extract the historical feature cepstral matrix from the structured mapping index table, use the dynamic time warping algorithm to calculate the distortion distance between the local frequency domain component set and the historical feature cepstral matrix and sort them in ascending order to construct a memory target replay index; S5: Based on the memory target playback index, read digital voice data, input the digital-to-analog converter chip to generate an analog voltage signal, and use the speaker hardware driver circuit to amplify and convert the analog voltage signal to generate a memory compensation prompt audio.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A speech memory enhancement device for assisting cognitive impairment, characterized in that, The device includes: The speech preprocessing module acquires the interactive speech data stream received by the microphone array, divides it into a fixed time window to generate a discrete waveform frame sequence, performs Fourier transform and energy spectral density calculation, and extracts the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. The acoustic segmentation module performs differential operation on adjacent frames based on the acoustic feature matrix to establish a differential feature sequence, uses a long short-term memory network to perform evolution direction inference calculation to construct hidden layer state probability values, compares them with a preset segmentation threshold to extract coordinate intervals, and generates an acoustically stable segmentation sequence. The complexity assessment module extracts the local frequency domain component set based on the acoustically stable segmented sequence, extracts the frequency centroid position, and applies the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value. The matching and retrieval module extracts the historical feature cepstral matrix from the structured mapping index table based on the syllable arrangement entropy value, calculates the distortion distance between the local frequency domain component set and the historical feature cepstral matrix using the dynamic time warping algorithm, and sorts them in ascending order to construct a memory target replay index. The memory feedback module reads digital voice data based on the memory target playback index, inputs it into the digital-to-analog converter chip to generate an analog voltage signal, and then amplifies and converts the analog voltage signal through the speaker hardware driver circuit to generate a memory compensation prompt audio.

2. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The acoustic feature matrix includes a frame index, a frequency band number, and cepstral coefficient values. The acoustically stable segmented sequence includes start coordinates, end coordinates, and segment duration. The local frequency domain component set includes frequency band energy distribution, harmonic components, and formant positions. The syllable arrangement entropy value is specifically a scalar entropy value that quantifies the complexity of syllables. The memory target playback index includes a matching priority order, cumulative distortion cost, and digital speech data address. The memory compensation prompt audio includes audio content, sound pressure level, and playback sequence.

3. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The speech preprocessing module includes: The waveform segmentation module acquires the interactive voice data stream received by the microphone array, and performs periodic segmentation processing on the voice data stream according to the preset time window length. It filters out continuous usable data segments whose amplitude exceeds the preset silent state amplitude reference value and generates a discrete waveform frame sequence. The frequency domain calculation submodule calls the discrete waveform frame sequence to perform discrete Fourier band transform operation, extracts the frequency band amplitude parameter of a single frame waveform, calculates the ratio of the square value of the frequency band amplitude parameter of a single frame waveform to the number of discrete sampling points of a single frame waveform, and generates the waveform energy spectral density value. The feature construction submodule maps the waveform energy spectral density value and the discrete waveform frame sequence to the Mel scale filter bank, extracts the Mel frequency cepstral coefficients and retains the values ​​that are greater than the preset environmental noise floor measurement threshold, and arranges them according to the sequential timing rules to generate an acoustic feature matrix.

4. The speech memory enhancement device for assisting cognitive impairment according to claim 3, characterized in that, The preset silent state amplitude reference value is obtained by extracting historical environmental audio sample data under non-interactive environment conditions, calculating the arithmetic mean of the absolute values ​​of the amplitude of the historical environmental audio sample data, adding the arithmetic mean of the absolute values ​​of the amplitude to a set fixed amplitude offset parameter, and outputting the preset silent state amplitude reference value. The preset environmental noise floor measurement threshold is calculated and output by extracting the frequency band power distribution statistics of the background environmental noise data in the corresponding frequency band during non-interactive periods, multiplying the frequency band power distribution statistics with the set noise tolerance gain coefficient, and then outputting the preset environmental noise floor measurement threshold.

5. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The acoustic segmentation module includes: The differential feature calculation submodule reads feature vector data within adjacent speech time frames based on the acoustic feature matrix, performs multidimensional matrix subtraction difference operation, extracts the calculated numerical result items, arranges and combines them according to the corresponding original acquisition time, and generates a differential feature sequence. The hidden state inference submodule calls the differential feature sequence and substitutes it into the set gated loop calculation processing logic to perform hidden state evolution inference operation. It calculates the numerical output of the dot product of the current forgetting weight and the previous memory state parameter, and adds the numerical output of the dot product of the current input gate activation value and the current candidate state parameter to generate the current memory cell state. It calculates the hidden state parameter through the output gate and the current memory cell state and obtains the hidden state probability value through the activation function mapping. The boundary determination segmentation module compares the probability value of the hidden layer state with the preset segmentation threshold, extracts the timestamp index node corresponding to the segment that crosses the preset segmentation threshold, performs coherence merging processing, extracts the time truncation coordinates of the beginning and end of the segment, and generates an acoustically stable segmented sequence.

6. The speech memory enhancement device for assisting cognitive impairment according to claim 5, characterized in that, The preset segmentation threshold is determined by extracting a set number of historical speech training samples as input to the set gated loop calculation processing logic output global hidden layer probability mean, and superimposing the product of the probability distribution standard deviation corresponding to the global hidden layer probability mean and the set weight coefficient.

7. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The complexity evaluation module includes: The frequency band component extraction submodule obtains the discrete spectral parameters within the extraction time window of the acoustically stable segmented sequence, performs a bandwidth limiting truncation and screening operation to collect the energy amplitude parameters corresponding to each frequency node in the specified frequency band interval, and establishes a local frequency domain component set. The frequency centroid extraction submodule calls the energy amplitude parameters of each item in the local frequency domain component set, uses the energy amplitude parameters as weighting coefficients to perform weighted and scale normalization processing on the corresponding frequency node index values, extracts the signal frequency domain energy distribution features, and obtains the frequency centroid coordinate values. The permutation entropy calculation submodule performs phase space delay reconstruction processing on the frequency centroid coordinate values, extracts the distribution frequency index of various mode states, calculates the corresponding proportion probability value based on the distribution frequency index, performs information quantization derivation on the proportion probability value to extract sequence complexity features, and generates syllable permutation entropy value.

8. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The matching retrieval module includes: The mapping retrieval submodule queries the structured mapping index table based on the syllable arrangement entropy value, calculates the absolute value of the relative deviation between the syllable arrangement entropy value and the value of a single node record in the structured mapping index table, extracts the frequency order data parameter in the node item corresponding to the minimum value, and establishes a historical feature cepstral matrix. The distance normalization submodule calls the local frequency domain component set and the historical feature cepstral matrix, calculates the Euclidean distance difference between the component data in the local frequency domain component set and the data parameters in the historical feature cepstral matrix, constructs a two-dimensional aligned grid, finds the path to minimize the cumulative distance of nodes and counts the sum of values ​​along the path, and calculates the distortion distance. The index generation submodule performs numerical comparison operations on the distortion distance to generate an ascending-order data linked list. Based on the values ​​of each item in the ascending-order data linked list, it associates the original node identifiers, assigns retrieval order numbers, and performs parameter recombination and integration processing to construct a memory target replay index.

9. The speech memory enhancement device for assisting cognitive impairment according to claim 1, characterized in that, The memory feedback module includes: The voice reading submodule extracts the storage address addressing instruction based on the memory target replay index, calls the digital voice data in the preset sample dataset, performs a matching and filtering operation based on the digital voice data and the addressing instruction, extracts the digital coding parameters corresponding to each frequency node, and generates a discrete voice sampling sequence. The digital-to-analog decoding submodule performs amplitude continuous reconstruction on the digital-to-analog converter chip for the input of the discrete speech sampling sequence. It performs division operation on each digital coding parameter in the discrete speech sampling sequence with a preset amplitude reference, calculates the pulse modulation duty cycle coefficient and performs integration operation to generate analog continuous voltage amplitude. The power amplifier submodule calls the analog continuous voltage amplitude input speaker hardware driver circuit to perform gain adjustment operation, calculates the output power parameter by dividing the square of the analog continuous voltage amplitude by the hardware equivalent load impedance value, performs bias conditioning configuration based on the output power parameter, and generates memory compensation prompt audio.

10. A method for enhancing speech memory in assistive devices for cognitive impairment, characterized in that, The speech memory enhancement device for assisting cognitive impairment according to any one of claims 1-9 includes the following steps: S1: Acquire the interactive voice data stream received by the microphone array, divide it into a fixed time window to generate a discrete waveform frame sequence, perform Fourier transform and energy spectral density calculation, and extract the Mel frequency cepstral coefficients of each frame to construct an acoustic feature matrix. S2: Based on the acoustic feature matrix, perform differential operation on adjacent frames to establish a differential feature sequence, use a long short-term memory network to perform evolution direction reasoning calculation to construct hidden layer state probability values, compare with a preset segmentation threshold to extract coordinate intervals, and generate an acoustically stable segmented sequence. S3: Based on the acoustically stable segmented sequence, extract the local frequency domain component set, extract the frequency centroid position, and apply the permutation entropy calculation model to perform distribution frequency statistics and logarithmic weighted summation to extract the syllable permutation entropy value; S4: Based on the numerical query of the syllable arrangement entropy, extract the historical feature cepstral matrix from the structured mapping index table, use the dynamic time warping algorithm to calculate the distortion distance between the local frequency domain component set and the historical feature cepstral matrix and sort them in ascending order to construct a memory target replay index; S5: Based on the memory target playback index, read digital voice data, input the digital-to-analog converter chip to generate an analog voltage signal, and use the speaker hardware driver circuit to amplify and convert the analog voltage signal to generate a memory compensation prompt audio.

Images