Speech recognition system of intelligent mouse

Abstract

The invention discloses a voice recognition system of an intelligent mouse, and belongs to the technical field of data processing, voice data input by a user is collected through a voice collection sensor arranged in the intelligent mouse, then a voice spectrogram corresponding to the voice data is recognized through a CNN-CTC algorithm, and a voice recognition result is obtained; and finally, the specified document template is filled with the voice recognition result, so that mouse character data input can be realized, the functionality of the mouse is enhanced, the number of peripherals carried by a user is reduced, and the use convenience is enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of data processing, and in particular relates to a voice recognition system for an intelligent mouse. Background Art

[0002] With the continuous development of science and technology, computer input devices are becoming increasingly diverse. Mouse is a common computer input device that allows users to control the position of the cursor (usually an arrow or pointer) on the screen by moving the mouse on a plane. Although the mouse can perform various click operations, it is inconvenient to type outdoors or without a keyboard, whether it is a tablet or a laptop, and the external keyboard is not easy to carry, which causes certain inconveniences during operation. Although the mouse is easy to carry, the existing mouse does not have the function of typing or inputting text, so even if you carry a mouse, you cannot effectively complete the input of text. Summary of the invention

[0003] The invention provides a voice recognition system for an intelligent mouse, which is used to solve the problem that the mouse in the prior art has a single function and cannot effectively provide a data input function.

[0004] A voice recognition system for an intelligent mouse, comprising: a voice data acquisition module, a voice data preprocessing module, a voice recognition module and an automatic loading module;

[0005] The voice data acquisition module is used to acquire the voice input by the user through the voice acquisition sensor arranged in the intelligent mouse to obtain voice data;

[0006] The speech data preprocessing module is used to preprocess the speech data to obtain a speech spectrogram corresponding to the speech data;

[0007] The speech recognition module is used to recognize the speech spectrogram corresponding to the speech data using the CNN-CTC algorithm to obtain a speech recognition result;

[0008] The automatic loading module is used to fill the text information corresponding to the voice data input by the user into a preset document template based on the voice recognition result.

[0009] Further, the speech data is preprocessed to obtain a speech spectrogram corresponding to the speech data, including:

[0010] The speech data is framed and windowed, the frame length and frame shift are set to 25ms and 10ms respectively, and the window function adopts a Hamming window to obtain a multi-frame sub-speech signal;

[0011] After converting each frame of sub-speech signal into the frequency domain through fast Fourier transform, the energy spectrum density is calculated to obtain the characteristics of each frame of sub-speech signal;

[0012] The features of the sub-speech signals of each frame are concatenated and normalized to obtain a speech spectrogram.

[0013] Furthermore, the CNN-CTC algorithm is used to recognize the speech spectrogram corresponding to the speech data to obtain a speech recognition result, including:

[0014] The normalized speech spectrogram is used as the input of the CNN algorithm to obtain the posterior probability matrix output by the CNN algorithm;

[0015] The posterior probability matrix output by the CNN algorithm is used as the input of the CTC algorithm to perform optimal path decoding to obtain the speech recognition result.

[0016] Furthermore, before using the CNN algorithm, it also includes:

[0017] Initialize the hyperparameters of the CNN algorithm to obtain multiple different parameter individuals; wherein the parameter individuals include all the hyperparameters to be optimized of the CNN algorithm;

[0018] Obtain the cross entropy loss function value corresponding to each parameter individual, and determine the optimal parameter individual according to the cross entropy loss function value corresponding to the parameter individual;

[0019] Performing a spiral collaborative search on all parameter individuals around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search;

[0020] The parameter individuals after the spiral collaborative search are made to perform information fusion search on the information of other parameter individuals to obtain the parameter individuals after the information fusion search;

[0021] According to the cross entropy loss function value corresponding to each parameter individual, an interpolation controllable search is performed on the parameter individual after the information fusion search to obtain the parameter individual after the interpolation controllable search;

[0022] Determine whether the current number of training times has reached the maximum number of training times. If so, re-determine the optimal parameter individual based on the parameter individual after the interpolation controllable search, and use the hyperparameters contained in the optimal parameter individual as the final parameters of the CNN algorithm, and end the training process, otherwise return to the step of determining the optimal parameter individual.

[0023] Further, all parameter individuals are subjected to a spiral collaborative search around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search, including:

[0024] Based on the current number of training times, the spiral search factor is determined as:

[0025]

[0026] Among them, ξ k represents the spiral search factor corresponding to the kth parameter individual, represents the kth parameter individual in the tth training process, represents the optimal parameter individual, r1 represents a random number between (0,1), π represents the circumference of a circle, e represents a natural constant, cos represents the cosine function, T represents the preset maximum number of training times, and r2 represents a random number uniformly distributed within [0,1];

[0027] Based on the position of the optimal parameter individual in the solution space, all parameter individuals are subjected to a spiral collaborative search according to the spiral search factor, and the parameter individuals after the spiral collaborative search are obtained as follows:

[0028]

[0029] in, represents the parameter individuals after the kth spiral collaborative search, k = 1, 2, ..., K, K represents the total number of parameter individuals.

[0030] Furthermore, the parameter individuals after the spiral collaborative search are made to perform information fusion search on the information of other parameter individuals to obtain parameter individuals after the information fusion search, including:

[0031] Based on the current number of training times, the first information fusion factor and the second information fusion factor are determined as follows:

[0032]

[0033] Among them, α represents the first information fusion factor, α max represents the upper limit of the first information fusion factor, α min represents the lower limit of the first information fusion factor, T represents the preset maximum number of training times; β represents the second information fusion factor, e represents the natural constant, λ represents the information fusion coefficient, η represents the information fusion range control factor, and dist mn represents the Euclidean distance between the parameter individual after the m-th spiral collaborative search and the parameter individual after the m-th spiral collaborative search;

[0034] According to the first information fusion factor and the second information fusion factor, information fusion search is performed on the information of other parameter individuals, and the parameter individuals after information fusion search are obtained as follows:

[0035]

[0036] in, represents the d-th dimension parameter of the parameter individual after the n-th spiral collaborative search in the t-th training process, n = 1, 2, ..., K, K represents the total number of parameter individuals, d = 1, 2, ..., D, D represents the total dimension of the hyperparameters in the parameter individual, represents the d-th dimension parameter of the parameter individual after the n-th information fusion search, represents the d-th dimension parameter of the parameter individual after the m-th spiral collaborative search in the t-th training process, X max,d represents the upper limit of the hyperparameter of the dth dimension, X min,d Represents the lower limit of the hyperparameter of the dth dimension.

[0037] Furthermore, according to the cross entropy loss function value corresponding to each parameter individual, an interpolation controllable search is performed on the parameter individual after the information fusion search to obtain the parameter individual after the interpolation controllable search, including:

[0038] Determine the adaptive mutation probability based on the current number of training times;

[0039] A decision factor between (0,1) is randomly generated, and it is determined whether the decision factor is less than the adaptive mutation probability. If so, a controllable mutation strategy is used to perform mutation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search. Otherwise, an interpolation search strategy is used to perform interpolation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search.

[0040] Furthermore, based on the current number of training times, the adaptive mutation probability is determined as:

[0041]

[0042] in, represents the mutation probability of the parameter individual after the i-th information fusion search in the t-th training process, θ max Indicates the preset maximum mutation probability, θ min represents the preset minimum mutation probability, and T represents the preset maximum number of training times.

[0043] Furthermore, a controllable mutation strategy is used to perform mutation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search, including:

[0044] For any parameter individual after information fusion search, a first random parameter individual and a second random parameter individual are randomly determined for it;

[0045] Based on the first random parameter individual and the second random parameter individual, a mutation search is performed on the parameter individual after the information fusion search, and the mutation search individual is determined to be:

[0046]

[0047] in, represents the parameter individual after the jth information fusion search during the tth training process, represents the mutation search individual corresponding to the parameter individual after the jth information fusion search, τ represents the mutation search step, c represents the mutation search step perturbation factor, and c = 0.8*(1-t / T), represents the first random parameter individual, represents the second random parameter individual;

[0048] Determine whether the cross entropy loss function value corresponding to the mutation search individual corresponding to the parameter individual after the j-th information fusion search is less than the cross entropy loss function value corresponding to the parameter individual after the j-th information fusion search. If so, use the mutation search individual to replace the corresponding parameter individual after the j-th information fusion search, otherwise retain the original parameter individual after the j-th information fusion search.

[0049] Furthermore, an interpolation search strategy is used to perform interpolation search on the parameter individuals after the information fusion search, and the parameter individuals after the interpolation controllable search are obtained, including:

[0050] For any parameter individual after information fusion search, the longhorn beetle whisker search algorithm is used to search the parameter individual to obtain the first search individual corresponding to the longhorn beetle left whisker position and the second search individual corresponding to the longhorn beetle right whisker position;

[0051] According to the optimal parameter individual, the first search individual and the second search individual, an interpolation search is performed on the parameter individual after the information fusion search, and the interpolation search individual is obtained as follows:

[0052]

[0053] in, represents the d-th dimension parameter of the s-th interpolation search individual, It represents the d-th dimension parameter of the first search individual corresponding to the parameter individual after the s-th information fusion search. It represents the d-th dimension parameter of the second search individual corresponding to the parameter individual after the s-th information fusion search. represents the d-th dimension parameter of the optimal parameter individual, represents the fitness value corresponding to the second search individual, represents the fitness value corresponding to the first search individual, It represents the fitness value corresponding to the optimal parameter individual, which is obtained by adding the cross entropy loss function value to a constant and taking the inverse;

[0054] Determine whether the cross entropy loss function value of the s-th interpolation search individual is less than the cross entropy loss function value of the parameter individual after the s-th information fusion search. If so, use the s-th interpolation search individual to replace the parameter individual after the s-th information fusion search to obtain the parameter individual after the interpolation controllable search. Otherwise, directly use the parameter individual after the s-th information fusion search as the parameter individual after the interpolation controllable search.

[0055] The present invention provides a voice recognition system for an intelligent mouse, which collects voice data input by a user through a voice collection sensor arranged in the intelligent mouse, then uses a CNN-CTC algorithm to recognize a voice spectrogram corresponding to the voice data to obtain a voice recognition result, and finally fills the voice recognition result into a specified document template, thereby realizing the input of mouse text data, which not only enhances the functionality of the mouse, but also reduces the number of peripherals carried by the user, thereby enhancing convenience of use. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0057] Figure 1 A structural schematic diagram of a speech recognition system for an intelligent mouse provided by an embodiment of the present invention.

[0058] Figure 2 A training flowchart of a CNN algorithm provided in an embodiment of the present invention.

[0059] Among them, 101-speech data acquisition module, 102-speech data preprocessing module, 103-speech recognition module, 104-automatic loading module.

[0060] The above drawings have shown clear embodiments of the present invention, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present invention in any way, but to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0061] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0062] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0063] like Figure 1 As shown, the embodiment of the present invention provides a speech recognition system for an intelligent mouse, comprising: a speech data acquisition module 101, a speech data preprocessing module 102, a speech recognition module 103 and an automatic loading module 104;

[0064] The voice data acquisition module 101 is used to acquire the voice input by the user through the voice acquisition sensor arranged in the intelligent mouse to obtain voice data;

[0065] For example, the voice collection sensor may be a microphone disposed in the intelligent mouse, and voice data may be obtained after the user's voice is collected by the microphone.

[0066] The speech data preprocessing module 102 is used to preprocess the speech data to obtain a speech spectrogram corresponding to the speech data;

[0067] Optionally, speech data usually contains some other noise, which leads to low recognition accuracy. Therefore, MFCC (Mel-Fequency Cepstral Coefficients) features of the speech data can be extracted first, and then concatenated to obtain a speech spectrogram corresponding to the speech data.

[0068] The speech recognition module 103 is used to recognize the speech spectrogram corresponding to the speech data using the CNN-CTC algorithm to obtain a speech recognition result;

[0069] CTC (Connectionist Temporal Classification) uses a learning algorithm that directly calculates the overall output posterior probability of the sequence, and can complete the conversion of a speech signal into a corresponding word sequence or other text unit sequence with only one neural network. Under the CTC end-to-end framework, convolutional neural networks have greater application potential. CTC can splice all feature vectors of the entire speech signal into a large feature map as input. Through the stacking of CNN (Convolutional Neural Network) convolutional layers and pooling layers, the "receptive field" of the output layer neurons is improved as the number of layers increases, so that the final output of the network can capture more context information between speech frames; secondly, in the selection of modeling units, since CTC introduces "blank" labels, it is not necessary to strictly divide the category of each speech frame, so larger granularity modeling units such as syllables and Chinese characters can be selected. Moreover, the convolution and downsampling operations unique to CNN, while compressing the input feature map, also integrate and filter the speech frame data, and can extract more essential features from the speech signal of the large-granularity modeling unit containing more complex information. Therefore, the embodiment of the present invention adopts the CNN-CTC algorithm to recognize the speech spectrogram corresponding to the speech data to improve the speech recognition speed.

[0070] The automatic loading module 104 is used to fill the text information corresponding to the voice data input by the user into a preset document template based on the voice recognition result.

[0071] The preset document templates can be txt documents, word documents, and other documents that can be filled with text, thereby enriching the functions of the smart mouse and improving the convenience of data input.

[0072] The present invention provides a voice recognition system for an intelligent mouse, which collects voice data input by a user through a voice collection sensor arranged in the intelligent mouse, then uses a CNN-CTC algorithm to recognize a voice spectrogram corresponding to the voice data to obtain a voice recognition result, and finally fills the voice recognition result into a specified document template, thereby realizing the input of mouse text data, which not only enhances the functionality of the mouse, but also reduces the number of peripherals carried by the user, thereby enhancing convenience of use.

[0073] In an embodiment of the present invention, preprocessing the speech data to obtain a speech spectrogram corresponding to the speech data includes:

[0074] The speech data is framed and windowed, the frame length and frame shift are set to 25ms and 10ms respectively, and the window function adopts a Hamming window to obtain a multi-frame sub-speech signal;

[0075] Optionally, before framing, the speech data may be pre-emphasized to increase the accuracy of speech data recognition.

[0076] After converting each frame of sub-speech signal into the frequency domain through fast Fourier transform, the energy spectrum density is calculated to obtain the characteristics of each frame of sub-speech signal;

[0077] The features of the sub-speech signals of each frame are concatenated and normalized to obtain a speech spectrogram.

[0078] It is worth noting that in addition to the above method for obtaining speech spectrogram, other methods can also be used to obtain speech spectrogram. For example, the FBank feature map corresponding to the speech data can be obtained to obtain the speech spectrogram.

[0079] In the embodiment of the present invention, the CNN-CTC algorithm is used to recognize the speech spectrogram corresponding to the speech data to obtain a speech recognition result, including:

[0080] The normalized speech spectrogram is used as the input of the CNN algorithm to obtain the posterior probability matrix output by the CNN algorithm;

[0081] The posterior probability matrix output by the CNN algorithm is used as the input of the CTC algorithm to perform optimal path decoding to obtain the speech recognition result.

[0082] The workflow of the CNN-CTC algorithm is as follows: first, the feature extraction and compression of the input signal can be realized through convolution and pooling operations to obtain higher-level information in the speech; then the high-dimensional feature map output by multiple convolution kernels is flattened in the frequency dimension to obtain a feature matrix that integrates the information of the speech signal in the frequency domain and time domain, and then input it into the fully connected neural network, and obtain the posterior probability matrix of the output unit through the softmax operation. Then the posterior probability matrix is ​​used as the input of the CTC algorithm to obtain the speech recognition result.

[0083] Traditional CNN algorithms generally use the Adam optimizer for training. Although the Adam optimizer has a faster training speed, its training effect is poor and it is easy to fall into the local optimum, resulting in poor final speech recognition effect. Therefore, an embodiment of the present invention provides an improvement strategy, which is as follows.

[0084] like Figure 2 As shown, before using the CNN algorithm, it also includes:

[0085] S201, initializing the hyperparameters of the CNN algorithm to obtain a plurality of different parameter individuals; wherein the parameter individuals include all the hyperparameters to be optimized of the CNN algorithm;

[0086] S202, obtaining the cross entropy loss function value corresponding to each parameter individual, and determining the optimal parameter individual according to the cross entropy loss function value corresponding to the parameter individual;

[0087] S203, performing a spiral collaborative search on all parameter individuals around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search;

[0088] S204, enabling the parameter individual after the spiral collaborative search to perform information fusion search on the information of other parameter individuals to obtain the parameter individual after the information fusion search;

[0089] S205, performing interpolation controllable search on the parameter individuals after the information fusion search according to the cross entropy loss function value corresponding to each parameter individual, to obtain the parameter individuals after the interpolation controllable search;

[0090] S206. Determine whether the current number of training times has reached the maximum number of training times. If so, re-determine the optimal parameter individual based on the parameter individual after the interpolation controllable search, and use the hyperparameters contained in the optimal parameter individual as the final parameters of the CNN algorithm, and end the training process, otherwise return to the step of determining the optimal parameter individual.

[0091] In the embodiment of the present invention, all parameter individuals are subjected to a spiral collaborative search around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search, including:

[0092] Based on the current number of training times, the spiral search factor is determined as:

[0093]

[0094] Among them, ξ k represents the spiral search factor corresponding to the kth parameter individual, represents the kth parameter individual in the tth training process, represents the optimal parameter individual, r1 represents a random number between (0,1), π represents the circumference of a circle, e represents a natural constant, cos represents the cosine function, T represents the preset maximum number of training times, and r2 represents a random number uniformly distributed within [0,1];

[0095] Based on the position of the optimal parameter individual in the solution space, all parameter individuals are subjected to a spiral collaborative search according to the spiral search factor, and the parameter individuals after the spiral collaborative search are obtained as follows:

[0096]

[0097] in, represents the parameter individuals after the kth spiral collaborative search, k = 1, 2, ..., K, K represents the total number of parameter individuals.

[0098] The spiral collaborative search provided by the embodiment of the present invention can make all parameter individuals search around a position within an optimal area, and have stronger search capabilities in the early stage, and more refined search in the later stage of the algorithm. And the search position is constantly shifted, so as to realize the search of more areas, so that the algorithm has a stronger comprehensive search capability.

[0099] In the embodiment of the present invention, the parameter individuals after the spiral collaborative search are made to perform information fusion search on the information of other parameter individuals to obtain the parameter individuals after the information fusion search, including:

[0100] Based on the current number of training times, the first information fusion factor and the second information fusion factor are determined as follows:

[0101]

[0102] Among them, α represents the first information fusion factor, α max represents the upper limit of the first information fusion factor, α min represents the lower limit of the first information fusion factor, T represents the preset maximum number of training times; β represents the second information fusion factor, e represents the natural constant, λ represents the information fusion coefficient, η represents the information fusion range control factor, and dist mn represents the Euclidean distance between the parameter individual after the m-th spiral collaborative search and the parameter individual after the m-th spiral collaborative search;

[0103] According to the first information fusion factor and the second information fusion factor, information fusion search is performed on the information of other parameter individuals, and the parameter individuals after information fusion search are obtained as follows:

[0104]

[0105] in, represents the d-th dimension parameter of the parameter individual after the n-th spiral collaborative search in the t-th training process, n = 1, 2, ..., K, K represents the total number of parameter individuals, d = 1, 2, ..., D, D represents the total dimension of the hyperparameters in the parameter individual, represents the d-th dimension parameter of the parameter individual after the n-th information fusion search, represents the d-th dimension parameter of the parameter individual after the m-th spiral collaborative search in the t-th training process, X max,d represents the upper limit of the hyperparameter of the dth dimension, X min,d Represents the lower limit of the hyperparameter of the dth dimension.

[0106] The information fusion search provided by the embodiment of the present invention can enable a parameter individual to fuse the information of other parameter individuals and boundary information during the search process, thereby realizing the development of more unexplored areas, thereby avoiding missing the optimal solution and improving the solution search capability.

[0107] In an embodiment of the present invention, according to the cross entropy loss function value corresponding to each parameter individual, an interpolation controllable search is performed on the parameter individual after the information fusion search to obtain the parameter individual after the interpolation controllable search, including:

[0108] Determine the adaptive mutation probability based on the current number of training times;

[0109] A decision factor between (0,1) is randomly generated, and it is determined whether the decision factor is less than the adaptive mutation probability. If so, a controllable mutation strategy is used to perform mutation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search. Otherwise, an interpolation search strategy is used to perform interpolation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search.

[0110] In the embodiment of the present invention, based on the current number of training times, the adaptive mutation probability is determined as:

[0111]

[0112] in, represents the mutation probability of the parameter individual after the i-th information fusion search in the t-th training process, θ max Represents the preset maximum mutation probability, θ min represents the preset minimum mutation probability, and T represents the preset maximum number of training times.

[0113] In the embodiment of the present invention, a controllable mutation strategy is used to perform mutation search on parameter individuals after information fusion search to obtain parameter individuals after interpolation controllable search, including:

[0114] For any parameter individual after information fusion search, a first random parameter individual and a second random parameter individual are randomly determined for it;

[0115] Optionally, the first random parameter individual and the second random parameter individual may be randomly selected from parameter individuals after information fusion search, or may be randomly generated individuals.

[0116] Based on the first random parameter individual and the second random parameter individual, a mutation search is performed on the parameter individual after the information fusion search, and the mutation search individual is determined to be:

[0117]

[0118] in, represents the parameter individual after the jth information fusion search during the tth training process, represents the mutation search individual corresponding to the parameter individual after the jth information fusion search, τ represents the mutation search step, c represents the mutation search step perturbation factor, and c = 0.8*(1-t / T), represents the first random parameter individual, represents the second random parameter individual;

[0119] Determine whether the cross entropy loss function value corresponding to the mutation search individual corresponding to the parameter individual after the j-th information fusion search is less than the cross entropy loss function value corresponding to the parameter individual after the j-th information fusion search. If so, use the mutation search individual to replace the corresponding parameter individual after the j-th information fusion search, otherwise retain the original parameter individual after the j-th information fusion search.

[0120] The variation search provided by the embodiment of the present invention can greatly improve the global search capability of the algorithm, thereby effectively preventing the algorithm from falling into a local optimum.

[0121] In the embodiment of the present invention, an interpolation search strategy is used to perform an interpolation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search, including:

[0122] For any parameter individual after information fusion search, the longhorn beetle whisker search algorithm is used to search the parameter individual to obtain the first search individual corresponding to the longhorn beetle left whisker position and the second search individual corresponding to the longhorn beetle right whisker position;

[0123] According to the optimal parameter individual, the first search individual and the second search individual, an interpolation search is performed on the parameter individual after the information fusion search, and the interpolation search individual is obtained as follows:

[0124]

[0125] in, represents the d-th dimension parameter of the s-th interpolation search individual, It represents the d-th dimension parameter of the first search individual corresponding to the parameter individual after the s-th information fusion search. It represents the d-th dimension parameter of the second search individual corresponding to the parameter individual after the s-th information fusion search. represents the d-th dimension parameter of the optimal parameter individual, represents the fitness value corresponding to the second search individual, represents the fitness value corresponding to the first search individual, It represents the fitness value corresponding to the optimal parameter individual, which is obtained by adding the cross entropy loss function value to a constant and taking the inverse;

[0126] Determine whether the cross entropy loss function value of the s-th interpolation search individual is less than the cross entropy loss function value of the parameter individual after the s-th information fusion search. If so, use the s-th interpolation search individual to replace the parameter individual after the s-th information fusion search to obtain the parameter individual after the interpolation controllable search. Otherwise, directly use the parameter individual after the s-th information fusion search as the parameter individual after the interpolation controllable search.

[0127] The interpolation search provided in the embodiment of the present invention can enable the algorithm to fuse the optimal position and the position in its neighborhood for fusion search, which is more conducive to searching for a better solution. As the algorithm proceeds, the mutation probability gradually decreases, and the search shifts from the global optimal to the local optimal search, effectively improving the search capability of the algorithm and ultimately improving the accuracy of speech recognition.

[0128] Optionally, you can also perform out-of-bounds processing on hyperparameters after each search.

[0129] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed by the present invention. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A speech recognition system for an intelligent mouse, characterized in that: include: Voice data acquisition module, voice data preprocessing module, voice recognition module and automatic loading module; The voice data acquisition module is used to acquire the voice input by the user through the voice acquisition sensor arranged in the intelligent mouse to obtain voice data; The speech data preprocessing module is used to preprocess the speech data to obtain a speech spectrogram corresponding to the speech data; The speech recognition module is used to recognize the speech spectrogram corresponding to the speech data using the CNN-CTC algorithm to obtain a speech recognition result; The automatic loading module is used to fill the text information corresponding to the voice data input by the user into a preset document template based on the voice recognition result.

2. The voice recognition system for an intelligent mouse according to claim 1, characterized in that: Preprocessing the speech data to obtain a speech spectrogram corresponding to the speech data includes: The speech data is framed and windowed, the frame length and frame shift are set to 25ms and 10ms respectively, and the window function adopts a Hamming window to obtain a multi-frame sub-speech signal; After converting each frame of sub-speech signal into the frequency domain through fast Fourier transform, the energy spectrum density is calculated to obtain the characteristics of each frame of sub-speech signal; The features of the sub-speech signals of each frame are concatenated and normalized to obtain a speech spectrogram.

3. The speech recognition system of the intelligent mouse according to claim 2, characterized in that: The CNN-CTC algorithm is used to recognize the speech spectrogram corresponding to the speech data to obtain a speech recognition result, including: The normalized speech spectrogram is used as the input of the CNN algorithm to obtain the posterior probability matrix output by the CNN algorithm; The posterior probability matrix output by the CNN algorithm is used as the input of the CTC algorithm to perform optimal path decoding to obtain the speech recognition result.

4. The voice recognition system for an intelligent mouse according to claim 3, characterized in that: Before using the CNN algorithm, it also includes: Initialize the hyperparameters of the CNN algorithm to obtain multiple different parameter individuals; wherein the parameter individuals include all the hyperparameters to be optimized of the CNN algorithm; Obtain the cross entropy loss function value corresponding to each parameter individual, and determine the optimal parameter individual according to the cross entropy loss function value corresponding to the parameter individual; Performing a spiral collaborative search on all parameter individuals around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search; The parameter individuals after the spiral collaborative search are made to perform information fusion search on the information of other parameter individuals to obtain the parameter individuals after the information fusion search; According to the cross entropy loss function value corresponding to each parameter individual, an interpolation controllable search is performed on the parameter individual after the information fusion search to obtain the parameter individual after the interpolation controllable search; Determine whether the current number of training times has reached the maximum number of training times. If so, re-determine the optimal parameter individual based on the parameter individual after the interpolation controllable search, and use the hyperparameters contained in the optimal parameter individual as the final parameters of the CNN algorithm, and end the training process, otherwise return to the step of determining the optimal parameter individual.

5. The voice recognition system for an intelligent mouse according to claim 4, characterized in that: Performing a spiral collaborative search on all parameter individuals around the position of the optimal parameter individual in the solution space to obtain parameter individuals after the spiral collaborative search includes: Based on the current number of training times, the spiral search factor is determined as: Among them, ξ k represents the spiral search factor corresponding to the kth parameter individual, represents the kth parameter individual in the tth training process, represents the optimal parameter individual, r1 represents a random number between (0,1), π represents the circumference of a circle, e represents a natural constant, cos represents the cosine function, T represents the preset maximum number of training times, and r2 represents a random number uniformly distributed within [0,1]; Based on the position of the optimal parameter individual in the solution space, all parameter individuals are subjected to a spiral collaborative search according to the spiral search factor, and the parameter individuals after the spiral collaborative search are obtained as follows: in, represents the parameter individuals after the kth spiral collaborative search, k = 1, 2, ..., K, K represents the total number of parameter individuals.

6. The voice recognition system for an intelligent mouse according to claim 5, characterized in that: The parameter individuals after the spiral collaborative search are made to perform information fusion search on the information of other parameter individuals to obtain the parameter individuals after the information fusion search, including: Based on the current number of training times, the first information fusion factor and the second information fusion factor are determined as follows: Among them, α represents the first information fusion factor, α max represents the upper limit of the first information fusion factor, α min represents the lower limit of the first information fusion factor, T represents the preset maximum number of training times; β represents the second information fusion factor, e represents the natural constant, λ represents the information fusion coefficient, η represents the information fusion range control factor, and dist mn represents the Euclidean distance between the parameter individual after the m-th spiral collaborative search and the parameter individual after the m-th spiral collaborative search; According to the first information fusion factor and the second information fusion factor, information fusion search is performed on the information of other parameter individuals, and the parameter individuals after information fusion search are obtained as follows: in, represents the d-th dimension parameter of the parameter individual after the n-th spiral collaborative search in the t-th training process, n = 1, 2, ..., K, K represents the total number of parameter individuals, d = 1, 2, ..., D, D represents the total dimension of the hyperparameters in the parameter individual, represents the d-th dimension parameter of the parameter individual after the n-th information fusion search, represents the d-th dimension parameter of the parameter individual after the m-th spiral collaborative search in the t-th training process, X max,d represents the upper limit of the hyperparameter of the dth dimension, X min,d Represents the lower limit of the hyperparameter of the dth dimension.

7. The voice recognition system for an intelligent mouse according to claim 6, characterized in that: According to the cross entropy loss function value corresponding to each parameter individual, an interpolation controllable search is performed on the parameter individual after the information fusion search to obtain the parameter individual after the interpolation controllable search, including: Determine the adaptive mutation probability based on the current number of training times; A decision factor between (0,1) is randomly generated, and it is determined whether the decision factor is less than the adaptive mutation probability. If so, a controllable mutation strategy is used to perform mutation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search. Otherwise, an interpolation search strategy is used to perform interpolation search on the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search.

8. The voice recognition system for an intelligent mouse according to claim 7, characterized in that: Based on the current number of training times, the adaptive mutation probability is determined as: in, represents the mutation probability of the parameter individual after the i-th information fusion search in the t-th training process, θ max Indicates the preset maximum mutation probability, θ min represents the preset minimum mutation probability, and T represents the preset maximum number of training times.

9. The voice recognition system of the intelligent mouse according to claim 7, characterized in that: The controllable mutation strategy is used to perform mutation search on the parameter individuals after information fusion search to obtain the parameter individuals after interpolation controllable search, including: For any parameter individual after information fusion search, a first random parameter individual and a second random parameter individual are randomly determined for it; Based on the first random parameter individual and the second random parameter individual, a mutation search is performed on the parameter individual after the information fusion search, and the mutation search individual is determined to be: in, represents the parameter individual after the jth information fusion search during the tth training process, represents the mutation search individual corresponding to the parameter individual after the jth information fusion search, τ represents the mutation search step, c represents the mutation search step perturbation factor, and c = 0.8*(1-t / T), represents the first random parameter individual, represents the second random parameter individual; Determine whether the cross entropy loss function value corresponding to the mutation search individual corresponding to the parameter individual after the j-th information fusion search is less than the cross entropy loss function value corresponding to the parameter individual after the j-th information fusion search. If so, use the mutation search individual to replace the corresponding parameter individual after the j-th information fusion search, otherwise retain the original parameter individual after the j-th information fusion search.

10. The voice recognition system for an intelligent mouse according to claim 7, characterized in that: The interpolation search strategy is used to interpolate the parameter individuals after the information fusion search to obtain the parameter individuals after the interpolation controllable search, including: For any parameter individual after information fusion search, the longhorn beetle whisker search algorithm is used to search the parameter individual to obtain the first search individual corresponding to the longhorn beetle left whisker position and the second search individual corresponding to the longhorn beetle right whisker position; According to the optimal parameter individual, the first search individual and the second search individual, an interpolation search is performed on the parameter individual after the information fusion search, and the interpolation search individual is obtained as follows: in, represents the d-th dimension parameter of the s-th interpolation search individual, It represents the d-th dimension parameter of the first search individual corresponding to the parameter individual after the s-th information fusion search. It represents the d-th dimension parameter of the second search individual corresponding to the parameter individual after the s-th information fusion search. represents the d-th dimension parameter of the optimal parameter individual, represents the fitness value corresponding to the second search individual, represents the fitness value corresponding to the first search individual, It represents the fitness value corresponding to the optimal parameter individual, which is obtained by adding the cross entropy loss function value to a constant and taking the inverse; Determine whether the cross entropy loss function value of the s-th interpolation search individual is less than the cross entropy loss function value of the parameter individual after the s-th information fusion search. If so, use the s-th interpolation search individual to replace the parameter individual after the s-th information fusion search to obtain the parameter individual after the interpolation controllable search. Otherwise, directly use the parameter individual after the s-th information fusion search as the parameter individual after the interpolation controllable search.

Images