A Speech Recognition Method Based on a Multi-Synaptic Connection Optical Pulse Neural Network

Abstract

The present invention relates to a speech recognition method based on a multi-synaptic connection optical pulse neural network, comprising: Step 1: obtaining a vowel data set as a training sample; Step 2: establishing a multi-synaptic optical pulse neural network model; Step 3: inputting the training sample into the optical pulse neural network model, and training the optical pulse neural network model by using a weight update algorithm to obtain a trained optical pulse neural network model; Step 4: implementing speech recognition by using the trained optical pulse neural network model. The speech recognition method based on the multi-synaptic connection optical pulse neural network of the present invention improves the accuracy of speech recognition and enhances the processing ability of complex speech data.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech recognition, and particularly relates to a speech recognition method based on a multi-synaptic connection optical pulse neural network. Background Art

[0002] Intelligent speech is known as the "pearl" on the crown of artificial intelligence. With the breakthrough of deep learning technology, the improvement of big data technology and natural language understanding ability, it has led to an industrial boom. In addition to traditional industries such as education, customer service, and telecommunications, it has also opened up new application fields for speech technology such as in-vehicle, home, medical, and intelligent hardware. With the rise of artificial intelligence, speech recognition technology based on artificial neural networks has moved from theory to market application. Speech recognition is essentially a pattern recognition based on speech feature parameters. Through model training and learning, the input speech can be classified according to a certain pattern.

[0003] As the core field of brain-inspired computing, spiking neural networks are known as the "third-generation neural network model". It can effectively simulate the dynamic system of continuous information transmission between biological neurons over time. Its spike information transmission mechanism can exactly make up for the defect of deep learning in simulating time characteristics.

[0004] However, in order to improve the calculation efficiency, traditional spiking neural network algorithms adopt extremely simplified spiking neuron models, which only contain one dimension and even lack the internal mechanism of generating spikes. They cannot effectively simulate the characteristics of spike transmission and refractory period in biological neural networks, reducing the complexity of the network. Secondly, in the traditional local learning algorithm for directly training spiking neural networks, it is easy to encounter the problem of parameter saturation, and the weights may tend to a very large or very small value and cannot converge stably. Thirdly, traditional spiking neural networks include inhibitory and excitatory neurons and synapses, and there are negative spike and negative weight information. However, in the optical domain, negative spike information cannot be effectively expressed, so it cannot be directly applied to optical pulse neural networks. In addition, in actual biological neural networks, there may be multiple pairs of connections between a pair of neurons, while existing optical pulse neural network algorithms only consider single-synaptic connections. The above-mentioned defects will all lead to a relatively low recognition accuracy of traditional spiking neural network algorithms in the process of speech recognition and cannot process complex speech data. Summary of the Invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a speech recognition method based on a multi-synaptic connection optical pulse neural network. By using an optical pulse neuron based on a laser model, an algorithm adapted to an optical hardware platform is provided. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] The present invention provides a speech recognition method based on a multi-synaptic optical pulse neural network, comprising:

[0007] Step 1: Obtain a vowel dataset as training samples;

[0008] Step 2: Establish a multi-synaptic optical pulse neural network model;

[0009] Step 3: Inputting the training samples into the optical pulse neural network model, and training the optical pulse neural network model using a weight update algorithm to obtain a trained optical pulse neural network model;

[0010] Step 4: Use the trained optical pulse neural network model to realize speech recognition.

[0011] In one embodiment of the present invention, step 2 includes:

[0012] Step 2.1: Determine the number of input neurons based on the characteristic dimension of the training sample and the preset number of Gaussian pulses;

[0013] Step 2.2: Determine the number of synapses in the multi-synaptic optical pulse neural network model based on the number of input neurons and the preset number of synapses from each input neuron to the output neuron;

[0014] Step 2.3: Establish the multi-synaptic optical pulse neural network model according to the number of the input neurons, the number of the output neurons and the number of synapses.

[0015] In one embodiment of the present invention, step 3 includes:

[0016] Step 3.1: Initialize the parameters of the multi-synaptic optical pulse neural network model;

[0017] Step 3.2: Gaussian encoding the training sample according to a preset number of Gaussian pulses;

[0018] Step 3.3: Input the Gaussian-encoded training sample into the optical pulse neural network model to obtain the model output result, which includes the number of pulses n generated by the output neuron. o , first pulse time t o , and the peak time t max ;

[0019] Step 3.4: adjusting the weights of the multi-synaptic optical pulse neural network model according to the true category of the training sample and the model output result;

[0020] Step 3.5: Repeat Step 3.3 - Step 3.4 until the preset number of training times is reached, and a trained optical pulse neural network model is obtained.

[0021] In an embodiment of the present invention, Step 3.4 includes:

[0022] Determine whether the model output result meets the constraint conditions. If it meets, execute the weight update algorithm to adjust the weights of the multi - synaptic optical pulse neural network model. If it does not meet, execute Step 3.5; where

[0023] The constraint conditions are:

[0024]

[0025] where n d represents the target number of pulses of the target neuron, d i represents the incoming delay of each synapse, t i represents the input time of the training sample, and T w represents the learning time window;

[0026] The weight update algorithm is:

[0027]

[0028] w(x + 1)=w(x)+Δw oi ;

[0029] where Δw oi represents the weight adjustment value, η represents the learning rate, Δw STDP represents adjusting the weight according to the STDP rule, and Δw aSTDP represents adjusting the weight according to the anti - STDP rule. w(x) represents the current weight value, and w(x + 1) represents the updated weight value.

[0030] In an embodiment of the present invention, in Step 3.4, if w(x + 1)<0, then let w(x + 1)=rand, where rand represents a positive random number.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] The speech recognition method based on a multi - synaptic connection optical pulse neural network of the present invention proposes a scheme to expand the feature dimension through Gaussian coding and multi - synapses for the problems of applying an optical pulse neural network to the field of speech recognition. It effectively improves the performance of the optical pulse neural network, randomly activates the synapses with negative weights to avoid input information loss, and can avoid over - fitting, improving the accuracy of speech recognition and enhancing the processing ability of complex speech data.

[0033] The above description is only an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the content of the specification. And in order to make the above and other objects, features and advantages of the present invention more obvious and understandable, the following specifically gives preferred embodiments and, in conjunction with the accompanying drawings, details are described as follows. Description of the Drawings

[0034] Figure 1 It is a schematic diagram of a speech recognition method based on a multi-synaptic connection optical pulse neural network provided by an embodiment of the present invention;

[0035] Figure 2 It is a schematic structural diagram of a multi-synaptic optical pulse neural network model provided by an embodiment of the present invention;

[0036] Figure 3 It is a schematic diagram of a training process of a multi-synaptic connection optical pulse neural network provided by an embodiment of the present invention;

[0037] Figure 4 It is a schematic diagram of the result of a simulation experiment provided by an embodiment of the present invention. Detailed Embodiments

[0038] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following combines the accompanying drawings and specific embodiments to detail a speech recognition method based on a multi-synaptic connection optical pulse neural network proposed according to the present invention.

[0039] The foregoing and other technical contents, features and effects of the present invention can be clearly presented in the following detailed description in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and specific understanding of the technical means and effects adopted by the present invention to achieve the predetermined purpose can be obtained. However, the accompanying drawings are only for reference and illustration, and are not used to limit the technical solution of the present invention.

[0040] Embodiment 1

[0041] Please refer to Figures 1 - 3 , Figure 1 which is a schematic diagram of a speech recognition method based on a multi-synaptic connection optical pulse neural network provided by an embodiment of the present invention, Figure 2 and which is a schematic structural diagram of a multi-synaptic optical pulse neural network model provided by an embodiment of the present invention; Figure 3 which is a schematic diagram of a training process of a multi-synaptic connection optical pulse neural network provided by an embodiment of the present invention. As shown in the figure, the speech recognition method based on a multi-synaptic connection optical pulse neural network in this embodiment includes:

[0042] Step 1: Obtain a vowel data set as a training sample;

[0043] Phonemes are the basic units of speech. Airflow resonates in the vocal tract, and the frequency of the resonance reflects the main characteristics of the vocal tract. The position, bandwidth, and amplitude of the formant determine the timbre of vowels. Vowel datasets can be used to train models for speech recognition.

[0044] In this embodiment, the Deterding vowel dataset can be used as a training sample for the multi-synaptic optical pulse neural network model. The dataset uses 10-dimensional logarithmic area ratio coefficients extracted based on linear prediction as speech feature parameters.

[0045] It should be noted that if the feature dimension of the training sample is small, it is not conducive to classification training. Therefore, in this embodiment, the feature dimension of the training sample is expanded through Gaussian coding.

[0046] Step 2: Establish a multi-synaptic optical pulse neural network model;

[0047] Specifically, step 2 includes:

[0048] Step 2.1: Determine the number of input neurons based on the feature dimension of the training sample and the preset number of Gaussian pulses;

[0049] Assume that each training sample has N f features, the preset number of Gaussian pulses is N g , then the feature dimension after Gaussian encoding is N F =N g ×N f , linearly map the Gaussian coded features to the pulse center moment, t s is the input time set of the sth training sample, etc. are the time corresponding to each input feature, in ns. Each feature is encoded as a pulse time of an input neuron, so the number of input neurons required is also N F .

[0050] Step 2.2: Determine the number of synapses in the optical pulse neural network model based on the number of input neurons and the preset number of synapses from each input neuron to the output neuron;

[0051] Specifically, for each input feature after encoding Set up multiple synaptic connections. Let the number of synapses from each input neuron to the output neuron be N s , then the total number of synapses required for the multi-synaptic optical pulse neural network model is N s ×N F .

[0052] Step 2.3: Establish a multi-synaptic optical pulse neural network model using the rate equation of the laser neuron according to the number of input neurons, the number of output neurons, and the number of synapses.

[0053] In this embodiment, the number of output neurons N o is the number of sample categories. Specifically, a multi-synaptic optical pulse neural network model can be established using the rate equation of the laser neuron. The structure of the multi-synaptic optical pulse neural network model is as Figure 2 shown.

[0054] Step 3: Input the training samples into the optical pulse neural network model, and use the weight update algorithm to train the optical pulse neural network model to obtain a trained optical pulse neural network model;

[0055] Specifically, Step 3 includes:

[0056] Step 3.1: Initialize the parameters of the multi-synaptic optical pulse neural network model;

[0057] In this embodiment, parameter initialization includes initializing the laser neuron parameters, setting the incoming delay d i and the weight w i of each synapse to random values. This process is equivalent to adding noise, which can effectively improve the performance of the test set; initialize the learning time window T w , and the learning rate η.

[0058] Step 3.2: Perform Gaussian encoding on the training samples according to the preset number of Gaussian pulses;

[0059] Step 3.3: Input the Gaussian-encoded training samples into the optical pulse neural network model to obtain the model output results. The model output results include the number of pulses n [[ID=3 and the weight w i of each synapse to random values. This process is equivalent to adding noise, which can effectively improve the performance of the test set; initialize the learning time window T w , and the learning rate η.

[0058] Step 3.2: Perform Gaussian encoding on the training samples according to the preset number of Gaussian pulses;

[0059] Step 3.3: Input the Gaussian-encoded training samples into the optical pulse neural network model to obtain the model output results. The model output results include the number of pulses n o generated by the output neurons, the time t o of the first pulse, and the time t max at the peak;

[0060] It should be noted that if no pulse is generated, let the number of pulses n o generated by the output neurons be 0, t o = T max , where T max is the maximum simulation time.

[0061] Step 3.4: Adjust the weights of the multi-synaptic optical pulse neural network model according to the true category of the training samples and the model output results;

[0062] In this embodiment, the weight update is performed by combining the Tempotron algorithm and the STDP rule. It should be noted that before the model training, the output neurons corresponding to the input sample categories c (c = 1, 2,..., C) need to be defined as the target neurons.

[0063] Specifically, step 3.4 includes:

[0064] Judge whether the model output result meets the constraint condition. If it meets, execute the weight update algorithm to adjust the weights of the multi-synaptic optical pulse neural network model. If it does not meet, execute step 3.5;

[0065] Among them, the constraint condition is:

[0066]

[0067] Among them, n d represents the target pulse number of the target neuron, d i represents the incoming delay of each synapse, t i represents the input time of the training sample, and T w represents the learning time window;

[0068] The weight update algorithm is:

[0069]

[0070] w(x + 1) = w(x) + Δw oi (3);

[0071] Among them, Δw oi represents the weight adjustment value, η represents the learning rate, Δw STDP represents adjusting the weight according to the STDP rule, and Δw aSTDP represents adjusting the weight according to the anti-STDP rule. w(x) represents the current weight value, and w(x + 1) represents the updated weight value.

[0072] Since in the optical pulse neural network, negative weight values and negative pulse information are difficult to directly represent in the optical domain, the weights need to be restricted during the training process. Therefore, in step 3.4, if w(x + 1) < 0, then let w(x + 1) = rand, where rand represents a positive random number.

[0073] In this embodiment, it is selected to set the weight value less than 0 to a relatively small random number, which is equivalent to reactivating the synapses that have stopped working. In this process, the setting of the random synaptic weights is also equivalent to adding noise during the training process, which can avoid overfitting to a certain extent and improve the performance of the network.

[0074] Step 3.5: Repeat Step 3.3 - Step 3.4 until the preset number of training times is reached, and a trained optical pulse neural network model is obtained.

[0075] Step 4: Implement speech recognition using the trained optical pulse neural network model.

[0076] In the process of using the trained optical pulse neural network model for speech recognition, input features of the same scale as the training samples need to be obtained through the same encoding method. In actual detection, in order to further improve the accuracy of the network, when judging whether the network is correctly classified, only the output neuron number with the smallest output pulse moment needs to be found. In the case where multiple output neurons generate pulses, as long as the neuron that generates the pulse earliest is the target neuron, it can be determined as a correct classification.

[0077] It should be noted that the speech recognition method based on the multi - synaptic connection optical pulse neural network in this embodiment is adapted to the optical hardware platform and can be implemented using optical pulse neurons based on the laser model, which has more advantages in the hardware platform. The speech recognition method based on the multi - synaptic connection optical pulse neural network in this embodiment proposes a scheme to expand the feature dimension through Gaussian coding and multi - synapses for the problems of applying the optical pulse neural network to the field of speech recognition, effectively improving the performance of the optical pulse neural network. By randomly activating the synapses with negative weights, it avoids the loss of input information and can avoid over - fitting, improving the speech recognition accuracy of the pulse neural network and enhancing the ability to process complex speech data.

[0078] Embodiment 2

[0079] This embodiment illustrates the effect of the speech recognition method based on the multi - synaptic connection optical pulse neural network in Embodiment 1 through simulation experiments.

[0080] Specifically, the Deterding vowel dataset is used for testing, and the dataset uses 10 - dimensional log area ratio coefficients extracted based on linear prediction as speech feature parameters. A pulse neuron model containing a saturable vertical - cavity surface - emitting semiconductor laser is used for training and inference.

[0081] Simulation experiment conditions:

[0082] The sample feature dimension of the Deterding vowel dataset is N f = 10. First, it is expanded through Gaussian coding, and N g = 4 is selected, and the number of synapses N s = 3 is selected. Four types of vowels are randomly selected for the experiment. Then the scale of the neural network is: 40 input neurons, 120 synapses, and 4 output neurons. In the simulation: the maximum simulation time is T max = 25ns, learning window Tw Set to 14 ns, the coding range is 6 - 15 ns, the maximum number of iterations is 200, and the learning rate η = 0.1.

[0083] Please refer to Figure 4 the schematic diagram of the results of the simulation experiment shown, where, in figure (a), N s = 1, the schematic diagram of the results of the simulation experiment where the samples are not Gaussian - coded, in figure (b), N s = 1, N g = 8, the schematic diagram of the results of the simulation experiment, in figure (c), N s = 3, the schematic diagram of the results of the simulation experiment where the samples are not Gaussian - coded, in figure (d), N s = 3, N g = 4, the schematic diagram of the results of the simulation experiment. It can be seen from the figure that by increasing the number of synapses and adopting Gaussian coding, the performance of the optical pulse neural network can be gradually improved.

[0084] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant is intended to cover non - exclusive inclusion, so that an article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the article or device including said element. "Connection" or "connected" and other similar words are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "up", "down", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as a limitation of the present invention.

[0085] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A speech recognition method based on a multi-synaptic optical pulse neural network, characterized in that: include: Step 1: Obtain a vowel dataset as training samples; Step 2: Establish a multi-synaptic optical pulse neural network model; Step 3: Inputting the training sample into the optical pulse neural network model, and training the optical pulse neural network model using a weight update algorithm to obtain a trained optical pulse neural network model; Step 3 includes: Step 3.1: Initialize the parameters of the multi-synaptic optical pulse neural network model; Step 3.2: Gaussian encoding the training sample according to a preset number of Gaussian pulses; Step 3.3: Input the Gaussian-encoded training sample into the optical pulse neural network model to obtain the model output result, which includes the number of pulses generated by the output neuron. , first pulse time , and the peak moment ; Step 3.4: adjusting the weights of the multi-synaptic optical pulse neural network model according to the true category of the training sample and the model output result; Step 3.4 includes: Determine whether the model output meets the constraint conditions. If so, execute the weight update algorithm to adjust the weights of the multi-synaptic optical pulse neural network model. If not, execute step 3.5; wherein, The constraints are: ； in, represents the target number of pulses of the target neuron, represents the incoming delay of each synapse, represents the input time of the training sample, represents the learning time window; The weight update algorithm is: ； ； in, Represents the weight adjustment value, represents the learning rate, Indicates that the weights are adjusted according to the STDP rule. Indicates that the weights are adjusted according to the anti-STDP rule. Indicates the current weight value, Represents the updated weight value; like , then let , Represents a positive random number; Step 3.5: Repeat steps 3.3 to 3.4 until the preset number of training times is reached to obtain a trained optical pulse neural network model; Step 4: Use the trained optical pulse neural network model to realize speech recognition.

2. The speech recognition method based on multi-synaptic optical pulse neural network according to claim 1, characterized in that: The step 2 includes: Step 2.1: Determine the number of input neurons based on the characteristic dimension of the training sample and the preset number of Gaussian pulses; Step 2.2: Determine the number of synapses in the multi-synaptic optical pulse neural network model based on the number of input neurons and the preset number of synapses from each input neuron to the output neuron; Step 2.3: Establish the multi-synaptic optical pulse neural network model according to the number of the input neurons, the number of the output neurons and the number of synapses.

Images