Intelligent old person voice assistant system and method based on voice recognition

Abstract

This invention discloses an intelligent voice assistant system and method for the elderly based on speech recognition, belonging to the field of intelligent speech recognition technology. It addresses the lack of dynamic adaptation mechanisms in current systems. Traditional speech recognition systems, which use a uniform model for training, often fail to consider the unique physiological characteristics of different elderly individuals, such as vocal cord aging and unclear pronunciation. This invention employs a k-means clustering method based on MFCC (Mel-frequency cepstral coefficients) to divide elderly speech data into multiple clusters with similar characteristics. A dynamic clustering analysis method, which comprehensively monitors data distribution patterns and noise ratios, is used to determine the triggering conditions for updating clusters. Each cluster is trained with a separate speech recognition model to determine the optimal model parameters for each cluster. This technology integrates data analysis methods such as feature extraction, cluster analysis, and scoring functions, enabling the speech recognition system to dynamically model physiological speech characteristics, thereby improving the accuracy of speech recognition for the elderly.

Description

Technical Field

[0001] This invention relates to the field of intelligent speech recognition technology, and more specifically, to an intelligent elderly voice assistant system and method based on speech recognition. Background Technology

[0002] Speech recognition has a wide range of applications, including voice calls, teleconferencing, smart homes, and service robots. With the nationwide population expansion and aging, the number of people with impaired speech communication is increasing year by year, making speech recognition of great significance to intelligent industries such as service robots.

[0003] The existing technology has the following shortcomings:

[0004] Current companion robots have low accuracy in recognizing the speech of the elderly. They typically simply combine all speech data into a speech recognition model, ignoring the impact of differences in speech features on the model. Improving the accuracy of companion robots in recognizing the speech of the elderly is of great significance to the elderly population.

[0005] To address the above problems, this invention proposes a solution. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a smart elderly voice assistant system and method based on speech recognition to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A voice assistant system for the elderly based on speech recognition includes:

[0009] The feature extraction module is used to extract MFCC features from the input speech signals of elderly users, including pre-emphasis, framing, windowing, fast Fourier transform, Mel filter bank energy calculation, logarithmic operation, discrete cosine transform and first-order and second-order difference processing, to generate a dataset containing MFCC coefficient statistics.

[0010] The dynamic clustering module uses the k-means clustering algorithm to divide the speech data into multiple clusters and initiates re-clustering based on preset trigger conditions. These trigger conditions comprehensively evaluate the statistical differences between the noise ratio and the data distribution pattern using a scoring function, which is defined as follows: ;

[0011] The model training module is used to generate training data by proportionally sampling each new cluster after triggering re-clustering, and to train multiple speech recognition models based on the sampled data, with each model corresponding to the optimal parameters of a different cluster.

[0012] The speech recognition module calls the corresponding speech recognition model to recognize the user's speech based on the cluster to which the speech belongs. The model is based on the ONNX format to accelerate inference.

[0013] In a preferred embodiment, the triggering condition of the dynamic clustering module satisfies at least one of the following:

[0014] 1. The amount of newly added data has reached a preset threshold;

[0015] 2. The calculation result of the scoring function exceeds the preset percentile of the historical scoring values.

[0016] In a preferred embodiment, the MFCC feature extraction includes:

[0017] Pre-emphasis processing uses a transfer function;

[0018] The frame length is 25ms, the frame shift is 10ms, and the sampling frequency is 16kHz.

[0019] The windowing process uses a Hamming window with a length of 400 sampling points;

[0020] The Mel filter bank contains 30 triangular filters.

[0021] In a preferred embodiment, the noise ratio is calculated as follows:

[0022] For the current batch of data, if the distance from a data point to the centroid of its cluster exceeds three times the standard deviation of the distance within the cluster, it is determined to be a noise point. The noise ratio η = number of noise points / N, where N is the total amount of data in the current batch.

[0023] In a preferred embodiment, the weights α and β are determined by the following steps:

[0024] Construct a data matrix containing the noise ratio normalized value Z and the KS statistic D;

[0025] Calculate the covariance matrix of the data matrix and extract the eigenvector corresponding to the largest eigenvalue;

[0026] After normalizing the eigenvectors, we obtain α and β.

[0027] In a preferred embodiment, the system further includes a hardware module:

[0028] The front-end human-computer interaction interface is used to receive user commands and display recognition results;

[0029] The backend embedded controller deploys the feature extraction module, dynamic clustering module, and speech recognition module.

[0030] Environmental sensors and IoT modules are used to collect environmental data and upload it to the cloud.

[0031] The intelligent speech recognition method for the elderly based on dynamic clustering includes the following steps:

[0032] MFCC features are extracted from the input speech signals of elderly users to generate a dataset containing MFCC coefficient statistics;

[0033] The k-means clustering algorithm is used to divide the data into multiple clusters, and the noise ratio and data distribution difference are monitored based on the scoring function. When the trigger condition is met, re-clustering is initiated.

[0034] For each new cluster after re-clustering, sample proportionally and use the sampled data to train multiple speech recognition models.

[0035] The corresponding speech recognition model is invoked based on the cluster to which the user's voice belongs, and the model is based on the ONNX format to accelerate inference.

[0036] In a preferred embodiment, the trigger condition for initiating re-clustering is:

[0037] When the amount of new data reaches the number of pre-reviewed entries, the current score is calculated and compared with the preset quantile of the historical score. If the score exceeds the preset quantile, re-clustering is triggered.

[0038] In a preferred embodiment, model training includes:

[0039] The sampled data was manually converted into text to generate the training set;

[0040] Train acoustic and language models independently for each cluster and optimize model parameters;

[0041] The recognition result is generated by combining the acoustic model score and the language model probability through a decoder.

[0042] In a preferred embodiment, the inference acceleration of the speech recognition model is achieved through the following steps:

[0043] Convert the trained HDF5 format model to ONNX format;

[0044] An embedded inference implementation is achieved by writing a C++ interface based on the ONNX Runtime framework.

[0045] The technical effects and advantages of the intelligent elderly voice assistant system and method based on speech recognition of this invention are as follows:

[0046] This invention uses dynamic clustering by monitoring both data distribution pattern and noise ratio. Each cluster is trained with a separate speech recognition model to determine the optimal model parameters for each cluster, thereby improving the accuracy of speech recognition. Attached Figure Description

[0047] Figure 1 This is a timing diagram of the intelligent elderly voice assistant system based on speech recognition of the present invention;

[0048] Figure 2 This is a flowchart of the intelligent elderly speech recognition method based on dynamic clustering of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention employs a k-means clustering method based on MFCC (Mel-frequency cepstral coefficients) to divide elderly people's speech data into multiple clusters with similar characteristics. It utilizes a dynamic clustering analysis method that comprehensively monitors data distribution patterns and noise ratios to determine the triggering conditions for updating clusters. Each cluster is trained with a separate speech recognition model to determine the optimal model parameters for each cluster. This technology integrates data analysis methods such as feature extraction, clustering analysis, and scoring functions, enabling the speech recognition system to dynamically model speech physiological characteristics, thereby improving the accuracy of speech recognition for the elderly.

[0051] Example 1, such as Figure 1 As shown, the present invention provides an intelligent voice assistant system and method for the elderly based on speech recognition, comprising the following steps:

[0052] The intelligent voice assistant system design mainly consists of a front-end and a back-end. The human-machine interface (HMI) screen is used to build the configuration interface as the device front-end for user interaction. The Jetson controller serves as the device back-end, on which various functional modules and a serial communication script library are deployed. The front-end communicates with the back-end system via a serial port. After the system is powered on, the HMI screen and controller start running. The user clicks the buttons on the interactive interface to activate the corresponding function, and sends the command information to the back-end via the serial port. After analyzing the command, it is executed locally.

[0053] Control chip selection. Jetson Orin Nano (controller 1) and Jetson Nano (controller 2) were selected as the device controller units. Target detection and virtual painting modules were deployed on controller 2, and voice interaction, speech generator and other functions were deployed on controller 1.

[0054] Other hardware selections. USB peripherals were chosen. Given the controller's abundant USB ports, a USB microphone and USB speaker were selected as audio input and output peripherals. (Intelligent)

[0055] The target detection function of the voice device only needs to detect the presence of a target, so a monocular camera is sufficient. A 1080P USB monocular camera was selected as the video input device for target detection and virtual painting. Furthermore, since the virtual painting function requires a video display device to display the canvas, an HDMI mini projector was chosen as the video output device.

[0056] Both the voice receiver and the cloud monitoring model of the intelligent voice device rely on environmental data. Therefore, an SHT-30 temperature and humidity sensor was used to detect the ambient temperature and humidity. The "FanYi FBox" IoT module was selected to enable cloud storage and visualization of data related to the voice interaction device. To apply the intelligent voice device in various indoor and outdoor scenarios, a suitable power supply module also needed to be designed. Therefore, a switching power supply and an adjustable step-down chip were selected to form the power supply module. The switching power supply has an AC220V input and a DC24V output, supporting a maximum power of 150W. The adjustable step-down module supports a maximum DC38V input and an adjustable DC1.25~36V output. The controller, FBox IoT module, and HMI screen are connected to the 24V switching power supply via the step-down module. Other USB devices or sensors can be powered through the controller's peripheral interface.

[0057] Model inference acceleration based on Open Neural Network Exchange (ONNX) is a high-performance inference framework launched by Microsoft. It supports inference execution of ONNX models across multiple platforms. Implemented in C++, it also supports other languages ​​such as Python, making it ideal for DNN model inference on embedded and other smart terminal platforms. Therefore, this section describes the development of a C++ computation library based on ONNX to accelerate the inference of an Automatic Speech Recognition (ASR) model.

[0058] Model format conversion. ONNX, as a universal model exchange format, enables the migration of system models between different frameworks or platforms. Its specific usage process is as follows: Figure 2As shown. This patent uses the Keras interface for ASR model training and TensorFlow for the computation backend, resulting in a model in the default HDF5 (Hierarchical Data Format Version 5) format. HDF5 is a hierarchical file format used to store large-scale complex data, capable of simultaneously saving the weight parameters and network structure of a deep learning model. This patent uses the "tf2onnx" function library to achieve HDF5 to OONNX format conversion, with the interface function being "tf2onnx.convert.from_function".

[0059] The configuration interface mainly consists of three parts: function buttons, prompts, and information display. Function buttons are the activation buttons for various functions of the intelligent voice assistant. Depending on the activation method, pressing or releasing a function button will send a command to the controller. The prompts interface displays interactive prompts and data results returned by the controller. For example, in an automatic speech recognition task, the controller sends the recognized text to the HMI and displays it on the information prompts interface. The information display contains the information that is always displayed on the configuration interface, including a title display, real-time temperature and humidity display, and time and date display.

[0060] Speech feature extraction based on MFCC (Mel-frequency cepstral coefficients)

[0061] The core idea of ​​MFCC: MFCC (Mel-Frequency Cepstral Coefficients) is a speech feature representation method based on the auditory characteristics of the human ear.

[0062] The collected speech data from the elderly were all converted to a sampling frequency of 16kHz. Because the high-frequency energy of the glottis excitation during phonation is relatively weak, we pre-emphasized the signal to compensate for the attenuation of the high-frequency components.

[0063] ;

[0064] in, It is the nth sampling point after pre-emphasis. This is the nth sampling point of the original speech signal. The pre-emphasis coefficient is set to 0.95 here.

[0065] Then, the long speech signal is segmented into multiple short time segments (i.e., frames) for subsequent signal processing and feature extraction.

[0066] ;

[0067] in, It is the duration of a single frame (unit: milliseconds). This represents the time offset between adjacent frames. The sampling frequency (unit: Hz) is 16000 here. This represents the number of sampling points per frame.

[0068] To reduce spectral leakage, suppress abrupt changes at frame edges, and smooth the spectrum, windowing is applied.

[0069] ;

[0070] in, The value of the window function for the nth sampling point is (0≤n≤N−1), where N is the window length, which is the number of sampling points corresponding to the frame length, and is equal to 400.

[0071] Fast Fourier Transform (FFT):

[0072] ;

[0073] Amplitude spectrum calculation:

[0074] ;

[0075] in, Let be the complex spectral value of the k-th frequency component. Let be the power spectral energy of the k-th frequency component, where j is the imaginary unit.

[0076] Frequency to Mel scale conversion:

[0077] ;

[0078] Where f is the linear frequency (unit: Hz), and the linear frequency corresponding to the k-th frequency point. for

[0079] ;

[0080] Explanation of the range of k: Due to the symmetry of FFT, only the first half needs to be calculated.

[0081] Design a Mel filter bank: Distribute M triangular filters evenly on the Mel scale (usually M = 20~40, here we take 30). The frequency response of the m-th filter. satisfy:

[0082] ;

[0083] in, Let m be the gain value of the m-th Mel filter at the k-th frequency point.

[0084] The center frequency of the m-th Mel filter (converted back to linear frequency)

[0085] Energy accumulation:

[0086] ;

[0087] in, This represents the energy integral value of the m-th Mel band.

[0088] Take the logarithm:

[0089] ;

[0090] Discrete Cosine Transform:

[0091] ;

[0092] in, Let n be the nth MFCC coefficient, where n is the index of the MFCC coefficient (usually n = 0, 1, ..., L−1, L ≤ M).

[0093] Calculate the first-order difference ΔMFCC:

[0094] ;

[0095] in, The first difference of the nth MFCC coefficient in the t-th frame. This is the time offset. This is the size of the difference window (usually W=2).

[0096] Calculate the second-order difference ΔΔMFCC:

[0097] ;

[0098] Data preparation:

[0099] Calculate the MFCC coefficients, mean, standard deviation, extreme values, and quantiles of the first and second differences for each sample, and then normalize them. There are various forms of normalization, which will not be elaborated here. The resulting dataset is: D = [MFCC-mean, MFCC-std, MFCC-min, MFCC-max, MFCC-q1, MFCC-q2, MFCC-q3, ΔMFCC-mean, ..., ΔMFCC-q3, ΔΔMFCC-mean, ..., ΔΔMFCC-q3], which correspond to the normalized MFCC coefficients, mean, standard deviation, extreme values, and quantiles of the first and second differences, respectively.

[0100] Dataset ,in Finding cluster partitions and cluster center Minimize the objective function:

[0101] ;

[0102] Where K is the preset number of clusters, For the sample set of the k-th cluster, Let k be the centroid (center point) of the k-th cluster. Let k=5.

[0103] For each sample Calculate its distance to all centroids and assign it to the nearest cluster:

[0104] ;

[0105] The superscript (t) represents the value at the t-th iteration.

[0106] Recalculate the centroid of each cluster so that it equals the mean of the samples in that cluster:

[0107] ;

[0108] By alternately optimizing cluster allocation and centroid position, the objective function is guaranteed to decrease monotonically after each iteration. This occurs when the centroid movement distance is less than a threshold. ϵ Or stop when the maximum number of iterations is reached:

[0109] ;

[0110] The score is calculated for every 1000 new data entries. When the score exceeds the threshold, re-clustering is triggered.

[0111] Establish a scoring function

[0112] ;

[0113] in: This represents the noise ratio of the current batch of data. If data points... To the centroid of its cluster The distance exceeds 3 ( If the distance within the cluster is the standard deviation, then it is determined to be a noise point.

[0114] = Where N is the current batch data size. This is an indicator function (1 if the condition is met, 0 otherwise).

[0115] This represents the average proportion of historical noise.

[0116] ;

[0117] T represents the number of historical batches. Let t be the noise ratio of the t-th batch.

[0118] The standard deviation of the historical noise ratio.

[0119] ;

[0120] Calculate the D value: D is the KS test statistic (DK). [0,1]), the calculation steps are as follows:

[0121] 1. Historical Reference Distribution: The cumulative distribution function of the distances from historical data to the centroid of its cluster. Cumulative distribution function ;

[0122] 2. Current Distribution: The cumulative distribution function that calculates the distance from the current batch of data to the centroid of its cluster. .

[0123] 3. KS statistic: D∈[0,1], and the larger the value, the more significant the distribution difference.

[0124] , This represents the 95th percentile of historical ratings.

[0125] Weight Sure:

[0126] Core idea: Principal component analysis (PCA) determines the relative importance of variables by finding the projection directions (principal components) that maximize the variance of the data. In the scoring function... In this study, the first principal component loading coefficient of PCA can be used to automatically assign weights α and β, so that the scoring function maximizes data variation and thus captures the combined effects of burst noise and distribution drift.

[0127] Constructing data matrix X

[0128] D and , forming matrix X.

[0129] ;

[0130] Calculate the covariance matrix C of X:

[0131] ;

[0132] in, Let Z be the variance. Let D be the variance. Let Z be the covariance of D.

[0133] Find the eigenvectors of the covariance matrix C. and eigenvalues :

[0134] ;

[0135] Select the eigenvector corresponding to the largest eigenvalue. That is, the direction of the first principal component. The weights can then be calculated. :

[0136] ;

[0137] A speech recognition system works by converting speech signals into text or instructions. The system includes modules such as feature extraction, acoustic models, language models, and decoders. During the training phase, the speech signal undergoes multiple processing steps, including feature extraction, acoustic model training, and language model training. During the recognition phase, the speech signal is also processed and compared with training samples to find the most similar result.

[0138] The entire workflow involves extracting acoustic features from the training corpus, processing the signal, training the acoustic model and language model, and inputting the acoustic model score and language model score into the decoding search. The decoding search then yields the recognition result. In the recognition stage, the language signal to be recognized undergoes the same processing to obtain acoustic features, which are compared with the features of the training samples to find the most similar signal as the recognition result. The speech recognition system mainly includes feature extraction, acoustic modeling, language modeling, and decoding search. The decoder includes the acoustic model, dictionary, and language model.

[0139] The following will further explain the components of the basic principles of speech recognition:

[0140] Feature extraction: Feature extraction converts the speech signal from the time domain to the frequency domain, transforming the original acoustic signal into a set of digital features, providing more effective features for the next step of speech recognition. Mel-frequency cepstral coefficients (MFCCs) are commonly used in feature extraction.

[0141] Acoustic Model: An acoustic model is a fundamental model in the speech recognition process, used to describe the statistical relationship between specific speech units and acoustic features. Acoustic models can be constructed based on different speech units and their acoustic features, such as phonemes, words, or syllables. The main function of an acoustic model is to estimate the probability of speech units corresponding to their acoustic features given a speech input.

[0142] Dictionaries: Dictionaries are used to assign phonemes in audio data to corresponding written expressions, that is, to convert the heard speech signal into text information that people can understand. For example, Chinese dictionaries establish a one-to-one mapping between pinyin and Chinese characters. Dictionaries store an index table between words and phonemes.

[0143] Language models are trained on large corpora of data, and their learning goal is to obtain an estimate of the generation probability of all legal sentences. Through statistical analysis of the target language text, language models can calculate the probability of each word's occurrence and the relationships between words, thereby helping speech recognition systems to more accurately understand and convert speech signals.

[0144] Decoding: The main task of decoding is to convert the transcribed speech signal into a readable text sequence. In the decoding process, firstly, an acoustic model is used to calculate the matching degree between the input speech signal and all phonemes, and the result is converted into a transcribed speech signal sequence. Then, a language model is used to calculate the probability of each phoneme in the text, as well as the probability distribution of sentences composed of all letters or words in the text. A search strategy is used to find the optimal matching route, which is then translated to obtain the final text result.

[0145] Each time re-clustering is triggered, a proportional sample is randomly selected from each new cluster. The sampled data is then manually converted into text. The sampled data and the text content are used as training data to train the speech recognition model, resulting in five speech recognition models with different parameters. Each cluster is then input into its corresponding model for speech recognition.

[0146] This invention discloses an intelligent elderly voice assistant system and method based on speech recognition, comprising: a feature extraction module, a dynamic clustering module, a monitoring module, a speech recognition module, and signal connections between the modules. The feature extraction module is used to extract speech features from the data; the dynamic clustering module divides data with different speech features into different clusters; the monitoring module is used to calculate a scoring function to determine whether to trigger re-clustering; and the speech recognition module is used for speech recognition.

[0147] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0148] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0149] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0150] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0152] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A voice recognition based smart elderly voice assistant system characterized by, include: The feature extraction module is used to extract MFCC features from the input speech signals of elderly users, including pre-emphasis, framing, windowing, fast Fourier transform, Mel filter bank energy calculation, logarithmic operation, discrete cosine transform and first-order and second-order difference processing, to generate a dataset containing MFCC coefficient statistics. The dynamic clustering module uses the k-means clustering algorithm to divide the speech data into multiple clusters and initiates re-clustering based on preset trigger conditions; The preset triggering conditions comprehensively evaluate the statistical differences between the noise ratio and the data distribution pattern through a scoring function. The triggering conditions for the dynamic clustering module include the calculation result of the scoring function exceeding a preset quantile of historical scoring values. The scoring function is defined as: wherein, is the noise proportion of the current batch data, is the average of the historical noise proportions, is the standard deviation of the historical noise proportions, D is the KS statistic; and α and β are the weight of the normalized value of the noise proportion and the KS statistic, respectively, The model training module is used to generate training data by proportionally sampling each new cluster after triggering re-clustering, and to train multiple speech recognition models based on the sampled data, with each model corresponding to the optimal parameters of a different cluster. The speech recognition module calls the corresponding speech recognition model according to the cluster to which the user's speech belongs, and the model is based on the ONNX format to accelerate inference; The method for calculating the noise ratio is as follows: For the current batch of data, if the distance from a data point to the centroid of its cluster exceeds 3 times the standard deviation of the distance within the cluster, it is determined to be a noise point. The noise ratio η = number of noise points / N, where N is the total amount of data in the current batch. The α and β are determined through the following steps: Construct a data matrix containing the noise ratio normalized value Z and the KS statistic D; Calculate the covariance matrix of the data matrix and extract the eigenvector corresponding to the largest eigenvalue; After normalizing the eigenvectors, we obtain α and β.

2. The intelligent elderly voice assistant system based on speech recognition according to claim 1, characterized in that, The MFCC feature extraction includes: Pre-emphasis processing uses a transfer function; The frame length is 25ms, the frame shift is 10ms, and the sampling frequency is 16kHz. The windowing process uses a Hamming window with a length of 400 sampling points; The Mel filter bank contains 30 triangular filters.

3. The intelligent elderly voice assistant system based on speech recognition according to claim 1, characterized in that, The system also includes hardware modules: The front-end human-computer interaction interface is used to receive user commands and display recognition results; The backend embedded controller deploys the feature extraction module, dynamic clustering module, and speech recognition module. Environmental sensors and IoT modules are used to collect environmental data and upload it to the cloud.

4. A method for intelligent elderly speech recognition based on dynamic clustering, based on the intelligent elderly voice assistant system based on speech recognition as described in any one of claims 1-3, characterized in that, Includes the following steps: MFCC features are extracted from the input speech signals of elderly users to generate a dataset containing MFCC coefficient statistics; The k-means clustering algorithm is used to divide the data into multiple clusters, and the noise ratio and data distribution difference are monitored based on the scoring function. When the trigger condition is met, re-clustering is initiated. For each new cluster after re-clustering, sample proportionally and use the sampled data to train multiple speech recognition models. The corresponding speech recognition model is invoked based on the cluster to which the user's voice belongs, and the model is based on the ONNX format to accelerate inference.

5. The intelligent elderly speech recognition method based on dynamic clustering according to claim 4, characterized in that, Model training includes: The sampled data was manually converted into text to generate the training set; Train acoustic and language models independently for each cluster and optimize model parameters; The recognition result is generated by combining the acoustic model score and the language model probability through a decoder.

6. The intelligent elderly speech recognition method based on dynamic clustering according to claim 5, characterized in that, The inference acceleration of the speech recognition model is achieved through the following steps: Convert the trained HDF5 format model to ONNX format; An embedded inference implementation is achieved by writing a C++ interface based on the ONNX Runtime framework.

Images