Braking system and method based on voice recognition

Abstract

The invention relates to the technical field of voice recognition, in particular to a braking system and method based on voice recognition, and the system comprises a voice collection module which is used for collecting voice data in a vehicle to obtain voice signals, and the voice collection module comprises a plurality of microphones arranged in the vehicle; the voice recognition module is used for recognizing an emergency braking password of a driver; according to the braking system, the sound of a driver and the emergency braking password serve as triggering conditions, when the vehicle meets an emergency situation, the driver sends out the same emergency braking password twice or more in a short time in a tension state, an emergency braking instruction is triggered, and when the driver steps on a brake, pressure feedback exists on a brake pad, so that the emergency braking command is triggered. And when the driver does not step on the brake, no pressure feedback exists on the brake pad, and the vehicle executes the emergency braking instruction for self-braking, so that the problem that the driver takes the accelerator as the brake is avoided, and the driving safety of the vehicle is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and more specifically, to a braking system and method based on speech recognition. Background Art

[0002] With the vigorous development of the new energy vehicle field in recent years, there will be more and more car users in the future, and the probability of safety accidents will also increase. Many drivers with poor psychological quality may mistake the accelerator for the brake when encountering an emergency, causing the vehicle to accelerate again, resulting in serious traffic accidents. Therefore, the present invention provides a braking system and method based on voice recognition to solve the above problem. Summary of the invention

[0003] The object of the present invention is to provide a braking system and method based on voice recognition to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A braking system based on voice recognition, comprising:

[0006] A voice collection module, used to collect sound data in the car to obtain voice signals, the voice collection module includes a plurality of microphones arranged in the car;

[0007] A speech recognition module is used to recognize the driver's emergency braking command. The speech recognition module includes speech feature extraction and timbre and voiceprint recognition. The speech feature extraction is to extract Mel frequency cepstral coefficient features, i.e. MFCC features, from speech signals. The timbre and voiceprint recognition are recognized using a neural network model.

[0008] The voice confirmation module is used to analyze the volume, speech speed, pitch and number of emergency braking commands to confirm whether the driver is in a nervous state. When the driver is in a nervous state and issues the same emergency braking command more than twice, the emergency braking command is triggered;

[0009] The brake confirmation module includes a pressure sensor arranged on the brake pad. The brake confirmation module is used to receive the vehicle emergency braking command and confirm whether to execute it. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command to brake automatically.

[0010] A preferred technical solution of the present application: the number of the microphones is 2 to 4, and the microphones are arranged on the steering wheel or the instrument panel.

[0011] A preferred technical solution of the present application: the emergency braking command includes brake, emergency brake and stop, the tone recognition is used to confirm whether it is the driver's voice, and the voiceprint recognition is used to confirm whether the driver issues the emergency braking command.

[0012] A preferred technical solution of the present application: the neural network model includes an input layer, a convolution layer, a pooling layer, a flattening layer, a recurrent layer, a fully connected layer and an output layer.

[0013] A preferred technical solution of the present application: the input data of the input layer is Among them, T is the number of time steps, F is the feature dimension, is a real number set, i.e., MFCC feature;

[0014] The convolution layer performs convolution operation on the input data X through a series of convolution kernels. Suppose the size of convolution kernel K is k×

[0015] F, where k is the height of the convolution kernel and F is the width of the convolution kernel. If it is the same as the input feature dimension, the mathematical expression of the convolution operation is:

[0016]

[0017] Among them, Y i,j The result of the convolution operation is located in the i-th row and j-th column. It means summing along the height direction of the convolution kernel. represents the sum along the width direction of the convolution kernel, m represents the index of the convolution kernel in the height direction, n represents the index of the convolution kernel in the width direction, X i+m,j+n The input data of the input layer is located in the i+mth row and j+nth column, K m,n represents the value of the convolution kernel at position (m,n), b y represents the bias term;

[0018] The pooling layer is used to reduce the dimension of the feature map. The pooling operations include maximum pooling and average pooling. Maximum pooling is used for processing. Suppose the pooling window size is p, the step size is s, and the output result Y of the convolution layer is i,j As the input of the pooling layer, the mathematical expression of the maximum pooling is:

[0019] Z i,j =max(Y i·s:i·s+p,j )

[0020] Among them, the output Y of the pooling layer i·s:i·s+p,j By the output result Y of the convolution layer i,j Apply the pooling operation to obtain Z i,j To maximize the output of the pooling layer, the result is located in the i-th row and j-th column. Z is the two-dimensional feature map output by the pooling layer, and its size is Hp ×W p ;

[0021] The flattening layer is used to flatten the two-dimensional feature map Z of the pooling layer into a one-dimensional vector z so as to be input into the recurrent layer, where the vector size of z is H p ×W p ;

[0022] The recurrent layer captures temporal dependencies through the input gate, forget gate, cell state update, output gate, and the hidden state mechanism of the current time step. The mathematical model of the recurrent layer is as follows:

[0023] Input Gate:

[0024] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0025] Forget Gate:

[0026] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0027] Cell status update:

[0028]

[0029] Output Gate:

[0030] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0031] Hidden state at the current time step:

[0032] h t =o t ⊙tanh(C t )

[0033] Among them, the input x of the recurrent layer t is the feature vector of the current time step, i.e., the one-dimensional vector z of the input flattened layer, and the output i of the input gate t Indicates the importance of the input information at the current time step, and the output of the forget gate f t Indicates the degree to which the old information is forgotten at the current time step, The new candidate cell state represents the new information of the current time step, C tis the cell state at the current time step, combining the old cell state with the new candidate cell state, C t-1 is the cell state of the previous time step, ⊙ represents element-by-element multiplication, and the output o of the output gate t Indicates the degree of output information at the current time step, and the hidden state h at the current time step t Represents the output information of the current time step, tanh is the hyperbolic tangent activation function with an output value between -1 and 1, σ is the sigmoid activation function with an output value between 0 and 1, and W i , W f , W C and W o are weight matrices, b i , b f , b C and b o are bias terms. The weight matrix and bias term correspond to the input gate, forget gate, cell state update and output gate respectively. t-1 is the hidden state of the previous time step, [h t-1 ,x t ] means concatenating the hidden state of the previous time step with the input feature vector of the current time step;

[0034] The fully connected layer is used to flatten the output z of the layer or the output h of the loop layer. t Further processing is finally mapped to the output space required for the classification task. Let the input of the fully connected layer be H, where H is the output z of the flattened layer or the output h of the loop layer. t , then the mathematical model of the fully connected layer is:

[0035] O = softmax(W O ·H+b O )

[0036] Among them, O is the output of the fully connected layer, W O is the weight matrix, b O is a bias term, and softmax is a normalized exponential function used to convert the output of the recurrent layer into a probability distribution;

[0037] The output O of the fully connected layer represents the probability of the extracted MFCC features belonging to the driver. The final classification result is determined by selecting the category with the highest probability and is expressed as:

[0038]

[0039] Among them, O i represents the i-th output O of the fully connected layer, is the highest probability of the i-th output O of the fully connected layer;

[0040] The model is trained using the cross entropy loss function, and the cross entropy loss function L is expressed as:

[0041]

[0042] Where N is the number of samples, y i is the true label, y i The value of is 1 or 0;

[0043] The output layer is based on the label y i The value of is used to determine whether the voice in the car is from the driver and whether the driver issues an emergency braking command.

[0044] A preferred technical solution of the present application: the timbre recognition is performed by a local model of a neural network model, the local model does not include a recurrent layer, the input H of the fully connected layer is the output z of the flattened layer, and when the label y i =1, it means that the driver's voice is recognized in the extracted MFCC features. i =0 means that there is no driver's timbre in the extracted MFCC features, and therefore no driver's voice;

[0045] The voiceprint recognition is performed through the overall model of the neural network model, and the input H of the fully connected layer is the output h of the loop layer. t , when the label y i =1 means that the extracted MFCC features contain the emergency braking command of the driver’s voice. i =0 indicates that there is no emergency braking command with driver's voice in the extracted MFCC features.

[0046] A preferred technical solution of the present application: the voice confirmation module extracts the volume, speech speed, pitch and frequency features in the voice signal for analysis, and the process is as follows:

[0047] Suppose the input speech signal is v(τ), where τ is the time variable and v(τ) represents the speech signal value at the time point, and then extract the following features:

[0048] Volume V:

[0049]

[0050] Where V is the time length, represents the integration operation of the function. The upper and lower limits of indicate that the range of integration is from 0 to D, and dτ represents an infinitesimal change in the variable τ;

[0051] Speech rate R:

[0052]

[0053] Among them, Q is the number of syllables;

[0054] Tone B:

[0055]

[0056] Among them, is the frequency, is the Fourier transform of the speech signal v(τ);

[0057] The feature normalization is expressed as:

[0058]

[0059] Among them, A is the original feature, that is, the volume, speaking rate and pitch features extracted from the speech signal, M is the mean of the feature, S is the standard deviation of the feature, A′ is the standardized feature, and the volume V′, speaking rate R′ and pitch B′ are obtained after feature standardization;

[0060] The logistic regression model is used to determine whether the driver is in a state of tension. The extracted feature vector β = [V′, R′, B′], the label of the tension state y β ∈{0,1}, where y β =0 indicates the normal state, where y β =1 indicates a tense state, and the mathematical model of the logistic regression model is:

[0061]

[0062] Among them, P(y=1|β) means that under the condition of a given feature vector β, the target variable y β The probability of taking the value as 1, w is the weight vector, w T is the transpose of w, u is the bias term, and e is the base of the natural logarithm;

[0063] Classification rules for logistic regression model:

[0064]

[0065] When β =1, the driver is in a state of tension. β =0, the driver is in normal state.

[0066] A preferred technical solution of the present application: the voice confirmation module is also provided with a counter, the initial value of the counter is zero, when the correct emergency brake port is identified and the driver is in a nervous state, the value of the counter is increased until the value reaches two or above, triggering the emergency braking command, when the wrong emergency brake port is identified or the driver is in a normal state, the value of the counter is cleared.

[0067] The present application also provides a braking method based on voice recognition, comprising the following steps:

[0068] S1. The emergency braking command with the driver's own voice is recorded into the braking system. When the vehicle encounters an emergency, the driver issues the same emergency braking command twice or more in a short period of time under a tense state, and the braking system uses this as a trigger condition;

[0069] S2, the braking system collects the sound data in the car through the voice collection module to obtain the voice signal, and then recognizes the timbre and voiceprint in the voice signal through the voice recognition module, first identifying whether the timbre is the driver's voice, and then identifying whether the voiceprint is the emergency braking command;

[0070] S3, analyzing the volume, speech speed and tone of the emergency brake command in the voice signal through the voice confirmation module to confirm whether the driver is in a nervous state, and then confirming whether the number of emergency brake commands reaches two or more;

[0071] S4. When the triggering conditions are met, the emergency braking command is triggered. The vehicle emergency braking command is received through the brake confirmation module. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command to brake automatically.

[0072] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0073] The braking system of the present invention uses the driver's voice and the emergency braking command as trigger conditions. When the vehicle encounters an emergency situation, the driver issues the same emergency braking command twice or more in a short period of time in a tense state, triggering the emergency braking command. When the driver steps on the brakes, there is pressure feedback on the brake pads and the emergency braking command stops. When the driver does not step on the brakes, there is no pressure feedback on the brake pads and the vehicle will execute the emergency braking command to brake by itself, avoiding the problem of the driver mistaking the accelerator for the brake and improving the safety of vehicle driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0075] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. The present invention is further described below in combination with the embodiments.

[0076] See also Figure 1 , an embodiment of the present application provides a braking system based on voice recognition, comprising:

[0077] A voice collection module, used to collect sound data in the car to obtain voice signals, the voice collection module includes multiple microphones arranged in the car;

[0078] Specifically, the number of microphones is 2 to 4, and the microphones are arranged on the steering wheel or the instrument panel, so that the microphones can comprehensively and clearly collect the sound data in the vehicle.

[0079] The speech recognition module is used to recognize the driver's emergency braking command. Specifically, the emergency braking command includes brake, emergency brake and stop. The timbre recognition is used to confirm whether it is the driver's voice. The voiceprint recognition is used to confirm whether the driver has issued the emergency braking command. The speech recognition module includes speech feature extraction and timbre and voiceprint recognition. The speech feature extraction is to extract the Mel frequency cepstral coefficient feature from the speech signal, that is, MFCC feature. The timbre and voiceprint recognition are recognized by using a neural network model;

[0080] Among them, the neural network model includes input layer, convolution layer, pooling layer, flattening layer, recurrent layer, fully connected layer and output layer;

[0081] The input data of the input layer is Among them, T is the number of time steps, F is the feature dimension, is a real number set, i.e., MFCC feature;

[0082] The convolution layer performs convolution operations on the input data X through a series of convolution kernels. Assume that the size of the convolution kernel K is k×F, where k is the height of the convolution kernel and F is the width of the convolution kernel, which is the same as the feature dimension of the input. The mathematical expression of the convolution operation is:

[0083]

[0084] Among them, Y i,j The result of the convolution operation is located in the i-th row and j-th column. It means summing along the height direction of the convolution kernel. represents the sum along the width direction of the convolution kernel, m represents the index of the convolution kernel in the height direction, n represents the index of the convolution kernel in the width direction, X i+m,j+n The input data of the input layer is located in the i+mth row and j+nth column, K m,n represents the value of the convolution kernel at position (m,n), b y represents the bias term;

[0085] The pooling layer is used to reduce the dimension of the feature map. The pooling operations include maximum pooling and average pooling. Maximum pooling is used for processing. Suppose the pooling window size is p, the step size is s, and the output result of the convolution layer is Y i,j As the input of the pooling layer, the mathematical expression of the maximum pooling is:

[0086] Z i,j =max(Y i·s:i·s+p,j )

[0087] Among them, the output Y of the pooling layer i·s:i·s+p,j By the output result Y of the convolution layer i,j Apply the pooling operation to obtain Z i,j To maximize the output of the pooling layer, the result is located in the i-th row and j-th column. Z is the two-dimensional feature map output by the pooling layer, and its size is H p ×W p ;

[0088] Specifically, the height H of the two-dimensional feature map Z output by the pooling layer p and width W p Calculated by the following formula:

[0089]

[0090] The flattening layer is used to flatten the two-dimensional feature map Z of the pooling layer into a one-dimensional vector z for input to the recurrent layer, where the vector size of z is H p ×W p ;

[0091] The recurrent layer captures temporal dependencies through the input gate, forget gate, cell state update, output gate, and the hidden state mechanism of the current time step. The mathematical model of the recurrent layer is as follows:

[0092] Input Gate:

[0093] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0094] Forget Gate:

[0095] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0096] Cell status update:

[0097]

[0098] Output Gate:

[0099] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0100] Hidden state at the current time step:

[0101] h t =o t ⊙tanh(C t )

[0102] Among them, the input x of the recurrent layer t is the feature vector of the current time step, i.e., the one-dimensional vector z of the input flattened layer, and the output i of the input gate t Indicates the importance of the input information at the current time step, and the output of the forget gate f t Indicates the degree to which the old information is forgotten at the current time step, The new candidate cell state represents the new information of the current time step, C t is the cell state at the current time step, combining the old cell state with the new candidate cell state, C t-1 is the cell state of the previous time step, ⊙ represents element-by-element multiplication, and the output o of the output gate t Indicates the degree of output information at the current time step, and the hidden state h at the current time step t Represents the output information of the current time step, tanh is the hyperbolic tangent activation function with an output value between -1 and 1, σ is the sigmoid activation function with an output value between 0 and 1, and W i , W f , W C and W o are weight matrices, b i , b f , b C and b o are bias terms. The weight matrix and bias term correspond to the input gate, forget gate, cell state update and output gate respectively. t-1 is the hidden state of the previous time step, [h t-1 ,x t ] means concatenating the hidden state of the previous time step with the input feature vector of the current time step;

[0103] The fully connected layer is used to flatten the output z of the layer or the output h of the loop layer. t Further processing is finally mapped to the output space required for the classification task. Let the input of the fully connected layer be H, where H is the output z of the flattened layer or the output h of the loop layer. t , then the mathematical model of the fully connected layer is:

[0104] O = softmax(W O ·H+b O )

[0105] Among them, O is the output of the fully connected layer, W O is the weight matrix, b O is a bias term, and softmax is a normalized exponential function used to convert the output of the recurrent layer into a probability distribution;

[0106] The output O of the fully connected layer represents the probability of the extracted MFCC features belonging to the driver. The final classification result is determined by selecting the category with the highest probability and is expressed as:

[0107]

[0108] Among them, O i represents the i-th output O of the fully connected layer, is the highest probability of the i-th output O of the fully connected layer;

[0109] The model is trained using the cross entropy loss function, and the cross entropy loss function L is expressed as:

[0110]

[0111] Where N is the number of samples, y i is the true label, y i The value of is 1 or 0;

[0112] The output layer is based on the label y i The value of determines whether the voice in the car is from the driver and whether the driver issues an emergency brake command;

[0113] Specifically, timbre recognition is performed through a local model of the neural network model. The local model does not include a recurrent layer. The input H of the fully connected layer is the output z of the flattened layer. When the label y i =1, it means that the driver's voice is recognized in the extracted MFCC features. i =0 means that there is no driver's timbre in the extracted MFCC features, and therefore no driver's voice;

[0114] Voiceprint recognition is performed through the overall model of the neural network model. The input H of the fully connected layer is the output h of the loop layer. t , when the label y i =1 means that the extracted MFCC features contain the emergency braking command of the driver’s voice. i =0 indicates that there is no emergency braking command with driver's voice in the extracted MFCC features.

[0115] The voice confirmation module is used to analyze the volume, speech speed, pitch and number of emergency braking commands to confirm whether the driver is in a nervous state. When the driver is in a nervous state and issues the same emergency braking command more than twice, the emergency braking command is triggered;

[0116] The voice confirmation module extracts the volume, speech speed, pitch and frequency features from the voice signal for analysis. The process is as follows:

[0117] Suppose the input speech signal is v(τ), where τ is the time variable and v(τ) represents the speech signal value at the time point, and then extract the following features:

[0118] Volume V:

[0119]

[0120] Where V is the time length, represents the integration operation of the function. The upper and lower limits of indicate that the range of integration is from 0 to D, and dτ represents an infinitesimal change in the variable τ;

[0121] Speech rate R:

[0122]

[0123] Among them, Q is the number of syllables;

[0124] Tone B:

[0125]

[0126] Among them, is the frequency, is the Fourier transform of the speech signal v(τ);

[0127] The feature normalization is expressed as:

[0128]

[0129] Among them, A is the original feature, that is, the volume, speaking rate and pitch features extracted from the speech signal, M is the mean of the feature, S is the standard deviation of the feature, A′ is the standardized feature, and the volume V′, speaking rate R′ and pitch B′ are obtained after feature standardization;

[0130] The logistic regression model is used to determine whether the driver is in a state of tension. The extracted feature vector β = [V′, R′, B′], the label of the tension state y β ∈{0,1}, where y β =0 indicates the normal state, where y β =1 indicates a tense state, and the mathematical model of the logistic regression model is:

[0131]

[0132] Among them, P(y=1|β) means that under the condition of a given feature vector β, the target variable y β The probability of taking the value as 1, w is the weight vector, w T is the transpose of w, u is the bias term, and e is the base of the natural logarithm;

[0133] Classification rules for logistic regression model:

[0134]

[0135] When β =1, the driver is in a state of tension. β =0, the driver is in normal state;

[0136] Specifically, the voice confirmation module is also provided with a counter, the initial value of which is zero. When the correct emergency brake port is identified and the driver is in a nervous state, the value of the counter is increased until the value reaches two or above, triggering the emergency braking command. When the wrong emergency brake port is identified or the driver is in a normal state, the value of the counter is cleared.

[0137] The brake confirmation module includes a pressure sensor arranged on the brake pad. The brake confirmation module is used to receive the vehicle emergency braking command and confirm whether to execute it. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command and brake automatically.

[0138] This embodiment provides a braking method based on voice recognition, comprising the following steps:

[0139] S1. The emergency braking command with the driver's own voice is recorded into the braking system. When the vehicle encounters an emergency, the driver issues the same emergency braking command twice or more in a short period of time under a tense state, and the braking system uses this as a trigger condition;

[0140] S2, the braking system collects the sound data in the car through the voice collection module to obtain the voice signal, and then recognizes the timbre and voiceprint in the voice signal through the voice recognition module, first identifying whether the timbre is the driver's voice, and then identifying whether the voiceprint is the emergency braking command;

[0141] S3, analyzing the volume, speech speed and tone of the emergency brake command in the voice signal through the voice confirmation module to confirm whether the driver is in a nervous state, and then confirming whether the number of emergency brake commands reaches two or more;

[0142] S4. When the triggering conditions are met, the emergency braking command is triggered. The vehicle emergency braking command is received through the brake confirmation module. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command to brake automatically.

[0143] In summary, according to the present invention, when the vehicle encounters an emergency situation, the driver issues the same emergency braking command twice or more in a short period of time under a tense state, triggering the emergency braking instruction. When the driver steps on the brake, there is pressure feedback on the brake pad and the emergency braking instruction stops. When the driver does not step on the brake, there is no pressure feedback on the brake pad and the vehicle will execute the emergency braking instruction to brake automatically.

[0144] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.

[0145] In addition, it should be understood that although this specification is described according to implementation modes, not every implementation mode contains only independent technical solutions. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should take the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A braking system based on voice recognition, characterized in that: include: A voice collection module, used to collect sound data in the car to obtain voice signals, the voice collection module includes a plurality of microphones arranged in the car; A speech recognition module is used to recognize the driver's emergency braking command. The speech recognition module includes speech feature extraction and timbre and voiceprint recognition. The speech feature extraction is to extract Mel frequency cepstral coefficient features, i.e. MFCC features, from speech signals. The timbre and voiceprint recognition are recognized using a neural network model. The voice confirmation module is used to analyze the volume, speech speed, pitch and number of emergency braking commands to confirm whether the driver is in a nervous state. When the driver is in a nervous state and issues the same emergency braking command more than twice, the emergency braking command is triggered; The brake confirmation module includes a pressure sensor arranged on the brake pad. The brake confirmation module is used to receive the vehicle emergency braking command and confirm whether to execute it. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command to brake automatically.

2. A braking system based on voice recognition according to claim 1, characterized in that: The number of the microphones is 2 to 4, and the microphones are arranged on a steering wheel or a dashboard.

3. A braking system based on voice recognition according to claim 1 or 2, characterized in that: The emergency brake command includes brake, emergency brake and stop, the tone recognition is used to confirm whether it is the driver's voice, and the voiceprint recognition is used to confirm whether the driver issues an emergency brake command.

4. A braking system based on voice recognition according to claim 3, characterized in that: The neural network model includes an input layer, a convolution layer, a pooling layer, a flattening layer, a recurrent layer, a fully connected layer and an output layer.

5. A braking system based on voice recognition according to claim 4, characterized in that: The input data of the input layer is Among them, T is the number of time steps, F is the feature dimension, is a real number set, i.e., MFCC feature; The convolution layer performs convolution operation on the input data X through a series of convolution kernels. Assume that the size of the convolution kernel K is k×F, where k is the height of the convolution kernel and F is the width of the convolution kernel, which is the same as the feature dimension of the input. The mathematical expression of the convolution operation is: Among them, Y i，j The result of the convolution operation is located in the i-th row and j-th column. It means summing along the height direction of the convolution kernel. represents the sum along the width direction of the convolution kernel, m represents the index of the convolution kernel in the height direction, n represents the index of the convolution kernel in the width direction, X i+m，j+n The input data of the input layer is located in the i+mth row and j+nth column, K m，n represents the value of the convolution kernel at position (m, n), b y represents the bias term; The pooling layer is used to reduce the dimension of the feature map. The pooling operations include maximum pooling and average pooling. Maximum pooling is used for processing. Suppose the pooling window size is p, the step size is s, and the output result Y of the convolution layer is i，j As the input of the pooling layer, the mathematical expression of the maximum pooling is: Z i，j =max(Y i·s：i·s+p，j ) Among them, the output Y of the pooling layer i·s：i·s+p，j By the output result Y of the convolution layer i，j Apply the pooling operation to obtain Z i，j To maximize the output of the pooling layer, the result is located in the i-th row and j-th column. Z is the two-dimensional feature map output by the pooling layer, and its size is H p ×W p ; The flattening layer is used to flatten the two-dimensional feature map Z of the pooling layer into a one-dimensional vector z so as to be input into the recurrent layer, where the vector size of z is H p ×W p ; The recurrent layer captures temporal dependencies through the input gate, forget gate, cell state update, output gate, and the hidden state mechanism of the current time step. The mathematical model of the recurrent layer is as follows: Input Gate: i t =σ(W i ·[h t-1 ,x t ]+b i ) Forget Gate: f t =σ(W f ·[h t-1 ,x t ]+b f ) Cell status update: Output Gate: the t =σ(W o ·[h t-1 ,x t ]+b i ) Hidden state at the current time step: h t =o t ⊙tanh(C t ) Among them, the input x of the recurrent layer t is the feature vector of the current time step, i.e., the one-dimensional vector z of the input flattened layer, and the output i of the input gate t Indicates the importance of the input information at the current time step, and the output of the forget gate f t Indicates the degree to which the old information is forgotten at the current time step, The new candidate cell state represents the new information of the current time step, C t is the cell state at the current time step, combining the old cell state with the new candidate cell state, C t-1 is the cell state of the previous time step, ⊙ represents element-by-element multiplication, and the output o of the output gate t Indicates the degree of output information at the current time step, and the hidden state h at the current time step t Represents the output information of the current time step, tanh is the hyperbolic tangent activation function with an output value between -1 and 1, σ is the sigmoid activation function with an output value between 0 and 1, and W i , W f , W C and W o are weight matrices, b i 、b f 、b C and b o are bias terms. The weight matrix and bias term correspond to the input gate, forget gate, cell state update and output gate respectively. t-1 is the hidden state of the previous time step, [h t-1 , x t ] means concatenating the hidden state of the previous time step with the input feature vector of the current time step; The fully connected layer is used to flatten the output z of the layer or the output h of the loop layer. t Further processing is finally mapped to the output space required for the classification task. Let the input of the fully connected layer be H, where H is the output z of the flattened layer or the output h of the loop layer. t , then the mathematical model of the fully connected layer is: O=softmax(W O ·H+b O ) Among them, O is the output of the fully connected layer, W O is the weight matrix, b O is a bias term, and softmax is a normalized exponential function used to convert the output of the recurrent layer into a probability distribution; The output O of the fully connected layer represents the probability of the extracted MFCC features belonging to the driver. The final classification result is determined by selecting the category with the highest probability and is expressed as: Among them, O i represents the i-th output O of the fully connected layer, is the highest probability of the i-th output O of the fully connected layer; The model is trained using the cross entropy loss function, and the cross entropy loss function L is expressed as: Where N is the number of samples, y i is the true label, y i The value of is 1 or 0; The output layer is based on the label y i The value of is used to determine whether the voice in the car is from the driver and whether the driver issues an emergency braking command.

6. A braking system based on voice recognition according to claim 5, characterized in that: The timbre recognition is performed by a local model of a neural network model, which does not include a recurrent layer. The input H of the fully connected layer is the output z of the flattened layer. When the label y i =1, it means that the driver's voice is recognized in the extracted MFCC features. i =0 means that there is no driver's timbre in the extracted MFCC features, and therefore no driver's voice; The voiceprint recognition is performed through the overall model of the neural network model, and the input H of the fully connected layer is the output h of the loop layer. t , when the label y i =1 means that the extracted MFCC features contain the emergency braking command of the driver’s voice. i =0 indicates that there is no emergency braking command with driver's voice in the extracted MFCC features.

7. A braking system based on voice recognition according to claim 1 or 6, characterized in that: The voice confirmation module extracts the volume, speech speed, pitch and frequency features in the voice signal for analysis, and the process is as follows: Suppose the input speech signal is v(τ), where τ is the time variable and v(τ) represents the speech signal value at the time point, and then extract the following features: Volume V: Where V is the time length, represents the integration operation of the function. The upper and lower limits of indicate that the range of integration is from 0 to D, and dτ represents an infinitesimal change in the variable τ; Speech rate R: Among them, Q is the number of syllables; Tone B: Among them, is the frequency, is the Fourier transform of the speech signal v(τ); The feature normalization is expressed as: Among them, A is the original feature, that is, the volume, speaking rate and pitch features extracted from the speech signal, M is the mean of the feature, S is the standard deviation of the feature, A′ is the standardized feature, and the volume V′, speaking rate R′ and pitch B′ are obtained after feature standardization; The logistic regression model is used to determine whether the driver is in a state of tension. The extracted feature vector β = [V′, R′, B′], the label of the tension state y β ∈{0,1), where y β =0 indicates the normal state, where y β =1 indicates a tense state, and the mathematical model of the logistic regression model is: Among them, P(y=1|β) means that under the condition of a given feature vector β, the target variable y β The probability of taking the value as 1, w is the weight vector, w T is the transpose of w, u is the bias term, and e is the base of the natural logarithm; Classification rules for logistic regression model: When β =1, the driver is in a state of tension. β =0, the driver is in normal state.

8. A braking system based on voice recognition according to claim 7, characterized in that: The voice confirmation module is also provided with a counter, the initial value of which is zero. When the correct emergency brake port is identified and the driver is in a nervous state, the value of the counter is increased until the value reaches two or more, triggering the emergency brake command. When the wrong emergency brake port is identified or the driver is in a normal state, the value of the counter is cleared.

9. A braking method based on voice recognition, characterized in that: The following steps are involved: S1. The emergency braking command with the driver's own voice is recorded into the braking system. When the vehicle encounters an emergency, the driver issues the same emergency braking command twice or more in a short period of time under a tense state, and the braking system uses this as a trigger condition; S2, the braking system collects the sound data in the car through the voice collection module to obtain the voice signal, and then recognizes the timbre and voiceprint in the voice signal through the voice recognition module, first identifying whether the timbre is the driver's voice, and then identifying whether the voiceprint is the emergency braking command; S3, analyzing the volume, speech speed and tone of the emergency brake command in the voice signal through the voice confirmation module to confirm whether the driver is in a nervous state, and then confirming whether the number of emergency brake commands reaches two or more; S4. When the triggering conditions are met, the emergency braking command is triggered. The vehicle emergency braking command is received through the brake confirmation module. When there is pressure feedback on the brake pad, the emergency braking command stops. When there is no pressure feedback on the brake pad, the vehicle will execute the emergency braking command to brake automatically.

Images