CN107808659A - Intelligent sound signal type recognition system device - Google Patents

Abstract

An intelligent voice signal pattern recognition system device, comprising a frame body 10, the frame body 10 is provided with a cavity, and a voice collection module 1, a voice recognition module 2, a central processing unit 3, a wireless signal Transceiver 4, display screen 8, memory 33, network module 31, memory card 32, loudspeaker 35 and power supply 9, voice acquisition module 1 includes microphone 11, wireless walkie-talkie 12 and fixed recorder 13, voice recognition module 2 includes voice Input unit 20, speech preprocessing unit 21, speech signal feature extraction unit 22, feature matching discrimination classification unit 23, speech signal is collected by speech acquisition module 1, the signal that gathers is processed by speech recognition module 2, and data signal is preserved by memory 33 , the visualization of the operation flow of the human-computer interaction and the output of the result is displayed by the display screen 8, so it is more convenient for people to recognize the voice signal.

Description

Intelligent voice signal pattern recognition system device

technical field

The invention discloses an intelligent voice signal pattern recognition system device, which belongs to the technical field of intelligent electronic products, and specifically is an intelligent voice signal pattern recognition system equipped with a voice collection module, a voice recognition module, a control system and a loudspeaker. device.

Background technique

In people's daily life, there are all kinds of speech signals, such as speech signals from people's communication, sounds from machine operation, sounds from playing music, sounds from car whistles, etc. The speech signals are almost Flooded around the entire living environment, sometimes people want to accurately learn and identify which objects in a group of voice signals are sent. In common sound signals, people can often distinguish what objects make different sounds, but when multiple objects make sounds at the same time, especially when multiple objects of the same kind make sounds at the same time, or the recording environment is noisy, it is difficult for people Distinguish which sound is made by which object, for example, in a group of recordings of a multi-person debate phenomenon, when there are many people speaking, it is difficult for people to distinguish which words are spoken by which debater by listening to the recording. Therefore, people usually need a device capable of recognizing speech.

Before the present invention, there were also some voice recognition products on the market, such as some voice input software, etc., but most of them recognized words or letters in the voice, or paired and recognized simple single voices. Talking with mobile phones and other products, the mobile phone can complete certain tasks after recognizing the semantics of the voice, such as simple tasks such as calling and searching, but it cannot distinguish the characteristics of the voice, and cannot accurately identify the origin of similar voices or the same words. Similar questions uttered by the subject. Therefore, it is not convenient for people to use flexibly.

Contents of the invention

In order to overcome the above-mentioned technical shortcomings, the object of the present invention is to provide an intelligent voice signal pattern recognition system device, which can easily identify and record voice signals and propose characteristic parameters, and perform intelligent pattern recognition on unknown voice signals by performing existing signals , classification and extraction.

In order to achieve the above-mentioned purpose, the technical scheme that the present invention takes is: comprise frame body 10, frame body 10 is provided with cavity, is provided with speech collection module 1, speech recognition module 2, central processing unit 3, wireless in frame body 10 Signal transceiving device 4, display screen 8, memory 33, network module 31, memory card 32, loudspeaker 35 and power supply 9, voice acquisition module 1 includes microphone 11, wireless walkie-talkie 12 and fixed recorder 13, and voice recognition module 2 includes Voice input unit 20, voice preprocessing unit 21, voice signal feature extraction unit 22, feature matching discrimination classification unit 23, voice signal is collected by voice collection module 1, the signal collected is processed by voice recognition module 2, and data signal is processed by memory 33 Preserve, the visualization of the output of the operation process of man-machine interaction and result is shown by display screen 8, loudspeaker 35 is set to carry out voice prompt and broadcast recognition result to operation step, and network module 31 is set to the present invention and Internet cloud platform are connected, The central processing unit 3 is set to the program control and data calculation of the whole system device, and the wireless signal transceiver 4 is set to receive and transmit the radio signals generated by the wireless walkie-talkie 12, the smart phone, and the network module 31 and connect the present invention with the Internet. For wireless connection, the memory card 32 is set to read the recorded external voice data into the database of the present invention.

The present invention has designed that the voice input unit 20 is set to include two types of "voice input mode" and "voice test mode", which can be provided by the microphone 11 provided by the voice collection module 1, the wireless walkie-talkie 12, the fixed recorder 13 and the smart phone. Input voice in any mode of the mobile phone. In the "voice input mode", the voice input unit 20 is set to only perform voice input on one person or one object at a time. It is characterized in that the input voice is a 5-30 second audio Signal, the present invention adopts the multi-state voice input strategy, it is characterized in that, can comprise the multi-state combined voice of normal speech, singing or high/middle/bass in the voice of input, display 8 real-time display voice waveform and complete progress bar, input Need to carry out data labeling after speech is finished, labeling method adopts manual manual labeling, as having collected Zhang San's sound, promptly remark in the dialog box that display screen 8 of the present invention shows: " Zhang San's voice ", preserve and get final product, input The voice is stored in the memory 33. Under the "voice test mode", the present invention collects and tests together with one or more input tools of the microphone 11 in the voice collection module 1, the wireless walkie-talkie 12, the fixed recorder 13 and the smart phone. Voice, the test voice collection process is real-time collection, without any restrictions on the number of people, objects and time.

The present invention designs, voice input unit 20 is set to be connected with voice collection module 1, microphone 11 is connected to voice collection module 1 by audio line, wireless walkie-talkie 12 is connected with voice collection module 1 by radio signal.

The present invention is designed, voice collection module 1 can also adopt smart mobile phone to carry out voice signal input, by matching connection with voice collection module 1 of the present invention with mobile phone, matching mode comprises bluetooth, infrared ray, WIFI and scan two-dimension code to connect, realize voice Recording is equivalent to using the mobile phone as a wireless microphone, which is more convenient for multi-group voice interaction.

The present invention has designed that the voice preprocessing unit 21 converts the voice signal collected by the voice acquisition module 1 into an electrical signal, that is, converts an analog signal into a digital signal, and then performs conventional signal processing, including environmental background noise elimination, signal framing, Filtering, pre-emphasis, windowing functions, endpoint detection, etc.

The present invention designs that the speech signal feature extraction unit 22 is set to extract the main feature parameters reflecting the essence of the speech from the original speech signal to form a feature vector x _i , x _i = (x _i1 , x _i2 ,...x _ij ,..., x _in ) ^T , x _ij represents the j-th speech feature value of the i-th object or individual, the feature parameter extraction method preferably adopts frequency cepstral coefficient method (MFCC), and spectral envelope method, LPC interpolation method can also be used , LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and the feature vector system obtained after feature extraction will be automatically saved in the pattern class database. All sound features of an object or person correspond to a pattern class. If input N After the speech of an individual or object, N pattern classes are obtained. If each pattern class has n feature parameters, an n-dimensional feature space can be formed, that is, the marked feature signal set can be recorded as D={(x ₁ , y ₁ ),(x ₂ ,y ₂ ),…( _xi ,y _i ),…,(x _N ,y _N )}, where x _i ∈ χ=R ⁿ , _xi represents the i-th input The speech feature signal of the object or person, y _i ∈ Y={1,2,…,N}, y _i represents the i-th person or object, N represents the number of the N-th person or object, and the marked speech feature data constitutes The pattern class database is stored in the memory 33 of the present invention.

The present invention has designed, and feature matching distinguishes classification unit 23 and is set to adopt intelligent multiclass classifier, and the learning algorithm of classifier is set to adopt improved neural network classification algorithm, by the speech feature signal set that has entered and marked as training data , let the network model learn the training data, obtain classification rules, and complete the training of the classifier; then use the trained classifier to intelligently classify and identify unknown test voice signals; after the test signal extracts features, the present invention will Automatically carry out feature matching, carry out feature matching with the feature parameters of the extracted test voice signal in real time and the sample voice feature parameters that have been entered and marked in the memory 33 of the present invention, and calculate the test voice signal and all the sample voice signals that have been entered. similarity, then the test voice signal is divided into the sample signal pattern category with the highest similarity, and finally the present invention outputs the recognition result to the outside world, a report similar to "this is the sound of XXX", for example, if the present invention has stored Zhang San's speech characteristic signal, when Zhang San spoke or sang to the present invention, the present invention can calculate Zhang San's test speech characteristic parameter automatically and the entry speech signal of Zhang San that has entered and marked the most similar, through Recognize and automatically output "This is Zhang San's voice".

The present invention has designed, the multi-layer artificial neural network structure that multiclass classifier adopts, and it is characterized in that, one end of network is defined as input layer, and the other end is defined as output layer, and the part in the middle of input layer and output layer is defined as hidden layer , the input layer is used to receive external input signals, re-send the input signals to all neurons in the hidden layer, and pass the results to the output layer after calculation by the hidden layer, the output layer receives signals from the hidden layer, and outputs after calculation As for the classification result, that is, the recognition result, the preferred number of hidden layers in the present invention is set to 1-200 layers.

The invention designs that the training process of the improved artificial neural network classification algorithm includes steps 1-7.

Step 1: Network initialization. According to the number of input voice signals, the algorithm database is continuously updated. When the voice signals of N objects are input, N pattern classes are formed, and the sample space (X, Y) is obtained. The i-th group of samples is (X _i , Y _i ), X _i represents the set of feature vectors extracted from the i-th object, and Y _i represents the marked i-th object; according to the system input and output sequence (X, Y), determine the network input layer node number n, the hidden layer The number of nodes l, the number of nodes in the output layer m, wherein the value of n is determined by the number of corresponding feature values in the input signal feature extraction, the value of m is determined by the number of stored speech pattern classes, and the reference value of l is The value of a ranges from 0 to 10, which is automatically calculated and determined by the model, and the connection weight ω _ij between the neurons of the input layer and the hidden layer and the connection weight between the neurons of the hidden layer and the output layer are initialized ω _jk , initialize hidden layer threshold a and output layer threshold b, given learning rate η and neuron activation function.

Step 2: Calculate the output of the hidden layer. According to the input change X, the connection weight ω _ij of the neurons in the input layer and the hidden layer, and the hidden layer threshold a, calculate the hidden layer output H; record the output of the jth hidden layer node as H _j , j=1,2,...,l, where l is the number of nodes in the hidden layer, f is the hidden layer activation function, and there are many kinds of activation functions, and the present invention preferably adopts f(x) ⁼ (1+e- ^x ) ^-1 .

Step 3: Compute the output of the output layer. According to the hidden layer output H, the connection weight ω _jk between the hidden layer and the output layer neurons, and the output layer threshold b, calculate the output layer output O, record the output of the kth output layer node as O _k , k=1,2,...,m, where m is the number of nodes in the output layer, b _k is the threshold of the kth node in the output layer, and H _j is the output value of the jth node in the hidden layer.

Step 4: Calculate the prediction error. According to the output O and the expected output Y (true value) obtained by the network prediction, calculate the total error e of the network prediction, e _k is the error generated by the kth output layer node,

Step 5: Update weights. Update the network connection weights ω _jk and ω _ij according to the total network prediction error e, ω _jk ⁺ ＝ω _jk +η·H _j ·E _k , where j=1,2,...,l, k=1,2,... , m, η is the learning rate, E _k represents the sensitivity of the network total error of the output layer node to the output layer network node k, where i=1,2,...,n, j=1,2,...,l.

Step 6: Threshold update. Update the hidden layer threshold a and the output layer threshold b according to the network prediction total error e, j=1,2,...,l; b _k ⁺ =b _k +η·E _k , k=1,2,...,m.

Step 7: Judging whether the algorithm iteration is convergent, if not, return to step 2, and the preferred minimum error of the present invention is 0.001 to end the iteration.

The present invention designs that the voice collection module 1 is built with a voice collection card for collecting and processing the collected voice signals.

The present invention has designed, and fixed recorder 13 adopts windproof type microphone.

The present invention designs, and display screen 8 adopts touch screen or LED display screen with backlight.

In the present invention, a plurality of fixed recorders 13 can be provided, arranged at the shell of the present invention, for increasing the intensity of voice recording.

The present invention has a long-term storage function for the voice signals that have been entered and marked. Any voice signal stored in the voice pattern database of the present invention can be retrieved at any time with unknown test voices for comparison and recognition.

The use procedure of the present invention is that first turn on the power switch 5, then the system runs automatically, the display screen 8 lights up and displays the operation interface, and people can select two functions of "voice input mode" and "voice test mode".

(1) When selecting voice input, central processing unit 3 can control voice input unit 20 to enter " voice input mode ", and display screen 8 and loudspeaker 35 can prompt simultaneously " now is voice input mode, please speak " similar hint, people Can input voice in any mode of microphone 11 provided by voice collection module 1, wireless walkie-talkie 12, fixed recorder 13 and smart phone; in order to ensure that the present invention can accurately identify and quantify the voice characteristics of the identified object, therefore in "voice In the "recording mode" stage, only one person or one object can be recorded at a time. Since the sound signal data emitted by the same person when speaking and singing will have certain characteristic deviations, in order to improve the accuracy of speech signal recognition, this The invention adopts a multi-state voice input strategy, that is, the recorded voice can include normal speech, singing, or high/middle/bass and other multi-state combined sounds. The recording time is 5 to 30 seconds, and the display 8 will display the real-time waveform of the voice And complete the progress bar, if the voice of recording is unsatisfactory can delete input again, need to carry out data mark after input voice completes, and marking method adopts artificial manual mark, as gathering the sound of Zhang San, namely the dialogue shown in display screen 8 of the present invention Remarks in the frame: "Zhang San's voice", just save it, and the input voice is stored in the memory 33 of the present invention.

(2) After the voice signal has been entered, the control system of the present invention automatically sends the marked voice signal into the voice preprocessing unit 21, and the voice preprocessing unit 21 converts the voice signal collected by the voice acquisition module 1 into an electrical signal, which is about to The analog signal is converted into a digital signal, and then undergoes conventional signal processing, including environmental background noise elimination, signal framing, filtering, pre-emphasis, window function and endpoint detection, etc.

(3) control system of the present invention sends into signal feature extraction unit 22 automatically to speech signal after preprocessing, and speech signal feature extraction unit 22 extracts the feature parameter reflecting speech essence from the speech signal after preprocessing, obtains feature vector x _i , the feature parameter extraction method preferably adopts frequency cepstrum coefficient method (MFCC), and can also use spectral envelope method, LPC interpolation method, LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and extract features The feature vector system obtained after that is automatically saved in the pattern category library. All the voice features of a person correspond to a pattern category. If the voices of N individuals are input, N pattern categories are obtained. Thereby obtain the database of one person's corresponding speech signal pattern class, all data are all stored in the memory 33 of the present invention, so far, speech signal input pattern content is finished.

(4) After the voice entry is complete, the voice test can be carried out. When carrying out the voice test, it is only necessary to select "voice test mode" in the operation interface of the display screen 8, and the central processing unit 3 can control the voice input unit 20 to enter " Voice test mode", display screen 8 and loudspeaker 35 can prompt "in voice test..." similar prompt at the same time, at this moment people do not get to do any operation, the present invention can pass microphone 11, wireless walkie-talkie 12 in voice collection module 1 , the fixed recorder 13 and one or more input tools of the smart phone collect the test voice together, and the test voice collection process is real-time collection without any time limit and number of people.

(5) voice data collected under "voice test mode", the system device of the present invention can automatically carry out preprocessing and feature extraction to the test voice signal, convert the voice test signal collected into an electrical signal, and perform conventional Signal feature extraction is performed after filtering, noise removal, window function and endpoint detection.

(6) After the test signal extracts features, the present invention can automatically carry out feature matching, and the feature parameter of the test voice signal of extraction is carried out feature matching with the marked sample voice feature parameter that has entered in memory 33 of the present invention in real time, and Calculate the similarity between the test voice signal and all the original voice signals that have been recorded, and divide the test voice signal into the pattern category with the highest similarity, and finally the present invention outputs to the outside world, "This is the sound of XXX" similar Report, for example, if the present invention has stored Zhang San's voice feature signal, when Zhang San speaks or sings to the present invention, the present invention will automatically output "this is Zhang San's voice" after identification.

When the present invention is tested in public places, because in the test environment, there may be multiple objects speaking at the same time in the same time period, that is, the voice signal collected is a signal of broadband aliasing. Errors in extraction, the strategy adopted by the present invention is to use intelligent algorithms to first match and identify the voice feature parameters of a single person speaking and then identify and store them, and then the system automatically screens and separates the voice signals when speaking together, Output recognition result at last and report " now is the common voice of Zhang San, Li Si, Wang Wu... " similar hint, and prompts to exist XX voice and fails to recognize, press power off key 6 when shutting down the system.

The present invention also designs that the system device can also output to people a list of recognition results in a multi-person communication environment, including the number of people or objects speaking on the spot under the test environment, and screen and play recordings from multiple people speaking at the same time Identify and separate out what each person is saying, while filtering out other people's voices and ambient sounds.

When the voice signal feature of the specimen not stored in the present invention appears in the test voice signal, the present invention will automatically record the unknown voice signal feature to remind people whether to mark and store the voice signal of the object.

Description of drawings

Fig. 1 is a structural schematic diagram of the present invention.

Fig. 2 is a system frame diagram of the present invention.

Fig. 3 is a schematic diagram of a multi-layer artificial neural network of the present invention.

Fig. 4 is a flow chart of the improved neural network classification algorithm of the speech signal of the present invention.

Detailed ways

Accompanying drawing 1 is an embodiment of the present invention, in conjunction with accompanying drawing 1～accompanying drawing 4 concrete description present embodiment, comprise frame body 10, frame body 10 is provided with cavity, is provided with voice acquisition module 1 in frame body 10 , voice recognition module 2, central processing unit 3, wireless signal transceiving device 4, display screen 8, memory 33, network module 31, memory card 32, loudspeaker 35 and power supply 9, voice acquisition module 1 includes microphone 11, wireless walkie-talkie 12 With fixed recorder 13, voice recognition module 2 includes voice input unit 20, voice preprocessing unit 21, voice signal feature extraction unit 22, feature matching discrimination classification unit 23, voice signal is collected by voice collection module 1, the signal that gathers Processed by the voice recognition module 2, the data signal is stored by the memory 33, the visualization of the operation process of human-computer interaction and the output of the result is displayed by the display screen 8, the speaker 35 is set to perform voice prompts and broadcast recognition results to the operation steps, and the network module 31 It is set to connect the present invention with the Internet cloud platform, the central processing unit 3 is set to program control and data calculation of the entire system device, and the wireless signal transceiver device 4 is set to the wireless interphone 12, smart phone, and network module 31. The radio signal is received, transmitted and the present invention is wirelessly connected to the Internet, and the memory card 32 is set to read the recorded external voice data into the database of the present invention.

In the present embodiment, the voice input unit 20 is set to include two types of "voice input mode" and "voice test mode", which can be provided by the microphone 11 provided by the voice collection module 1, the wireless walkie-talkie 12, the fixed recorder 13 and Smartphone input voice in any way, in the "voice input mode", the voice input unit 20 is set to only perform voice input to one person or one object at a time, and it is characterized in that the input voice is a section of 5-30 seconds Audio signal, the present invention adopts the multi-state voice input strategy, it is characterized in that, can comprise the multi-state combination voice of normal speech, singing or high/middle/bass in the voice of input, display 8 real-time display voice waveform and complete progress bar, Need to carry out data labeling after input voice finishes, labeling method adopts artificial manual labeling, as having collected Zhang San's sound, promptly remark in the dialog box that display screen 8 of the present invention shows: " Zhang San's voice ", save and get final product, input The voice of the voice is stored in the memory 33. Under the "voice test mode", the present invention collects together through one or more input tools of the microphone 11 in the voice collection module 1, the wireless walkie-talkie 12, the fixed recorder 13 and the smart phone Test voice, the test voice collection process is real-time collection, without any restrictions on the number of people, objects and time.

In this embodiment, the voice input unit 20 is configured to be connected to the voice collection module 1, the microphone 11 is connected to the voice collection module 1 through an audio cable, and the wireless walkie-talkie 12 is connected to the voice collection module 1 through a radio signal.

In this embodiment, the voice collection module 1 can also use a smart phone to input voice signals, by matching and connecting the mobile phone with the voice collection module 1 of the present invention, the matching methods include Bluetooth, infrared, WIFI and scanning two-dimensional codes to connect, to achieve Voice recording is equivalent to using the mobile phone as a wireless microphone, which is more convenient for multi-group voice interaction.

In this embodiment, the voice preprocessing unit 21 converts the voice signal collected by the voice collection module 1 into an electrical signal, that is, converts an analog signal into a digital signal, and then performs conventional signal processing, including environmental background noise elimination, signal framing , filtering, pre-emphasis, windowing function and endpoint detection, etc.

In this embodiment, the speech signal feature extraction unit 22 is configured to extract the main feature parameters reflecting the essence of the speech from the original speech signal to form a feature vector x _i , x _i = (x _i1 , x _i2 , ... x _ij , ... , x _in ) ^T , x _ij represents the j-th speech feature value of the i-th object or individual, the feature parameter extraction method preferably adopts frequency cepstral coefficient method (MFCC), and spectral envelope method, LPC interpolation can also be used method, LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and the feature vector system obtained after feature extraction will be automatically saved in the pattern class database. All sound features of an object or person correspond to a pattern class. After the voices of N people or objects, N pattern classes are obtained. If each pattern class has n feature parameters, it can form an n-dimensional feature space, that is, the marked feature signal set can be recorded as D={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…( _xi ,y _i ),…,(x _N ,y _N )}, where x _i ∈ χ=R ⁿ , x _i represents the i-th input The speech feature signal of an object or person, y _i ∈ Y={1,2,…,N}, y _i represents the i-th person or object, N represents the number of the N-th person or object, and the marked speech feature data A pattern database is formed and stored in the memory 33 of the present invention.

In the present embodiment, the feature matching discrimination classification unit 23 is set to adopt an intelligent multiclass classifier, and the learning algorithm of the classifier is set to adopt an improved neural network classification algorithm, and the speech feature signal set that has been entered and marked is used as a training data, let the network model learn the training data, obtain classification rules, and complete the training of the classifier; then use the trained classifier to intelligently classify and identify unknown test voice signals; after the test signal extracts features, the present invention Can automatically carry out feature matching, the feature parameter of the test voice signal of extraction is carried out feature matching with the sample voice feature parameter that has entered and marked in the memory 33 of the present invention in real time, and calculates the test voice signal and all the sample voice signals that have entered The similarity of the test voice signal is then divided into the sample signal pattern category with the highest similarity, and finally the present invention outputs the recognition result to the outside world, a report similar to "this is the sound of XXX", for example, if the present invention has Stored Zhang San's speech characteristic signal, when Zhang San speaks or sings to the present invention, the present invention can automatically calculate Zhang San's test speech characteristic parameter and the input speech signal of Zhang San that has entered and marked the most similar, After identification, automatically output "This is Zhang San's voice".

In this embodiment, the multi-layer artificial neural network structure adopted by the multi-class classifier is characterized in that one end of the network is defined as the input layer, the other end is defined as the output layer, and the part between the input layer and the output layer is defined as the hidden layer. Layer, the input layer is used to receive the input signal from the outside world, re-send the input signal to all neurons in the hidden layer, and pass the result to the output layer after calculation by the hidden layer, the output layer receives the signal from the hidden layer, and after calculation To output classification results, that is, recognition results, the number of hidden layers is preferably set to 1-200 in the present invention.

In the present embodiment, the process of the improved artificial neural network classification algorithm training is as follows:

Step 1: Network initialization. According to the number of input voice signals, the algorithm database is continuously updated. When the voice signals of N objects are input, N pattern classes are formed, and the sample space (X, Y) is obtained. The i-th group of samples is (X _i , Y _i ), X _i represents the set of feature vectors extracted from the i-th object, and Y _i represents the marked i-th object; according to the system input and output sequence (X, Y), determine the network input layer node number n, the hidden layer The number of nodes l, the number of nodes in the output layer m, wherein the value of n is determined by the number of corresponding feature values in the input signal feature extraction, the value of m is determined by the number of stored speech pattern classes, and the reference value of l is The value of a ranges from 0 to 10, which is automatically calculated and determined by the model, and the connection weight ω _ij between the neurons of the input layer and the hidden layer and the connection weight between the neurons of the hidden layer and the output layer are initialized ω _jk , initialize hidden layer threshold a and output layer threshold b, given learning rate η and neuron activation function.

Step 2: Calculate the output of the hidden layer. According to the input change X, the connection weight ω _ij of the neurons in the input layer and the hidden layer, and the hidden layer threshold a, calculate the hidden layer output H; record the output of the jth hidden layer node as H _j , j=1,2,...,l, where l is the number of nodes in the hidden layer, f is the hidden layer activation function, and there are many kinds of activation functions, and the present invention preferably adopts f(x) ⁼ (1+e- ^x ) ^-1 .

Step 3: Compute the output of the output layer. According to the hidden layer output H, the connection weight ω _jk between the hidden layer and the output layer neurons, and the output layer threshold b, calculate the output layer output O, record the output of the kth output layer node as O _k , k=1,2,...,m, where m is the number of nodes in the output layer, b _k is the threshold of the kth node in the output layer, and H _j is the output value of the jth node in the hidden layer.

Step 4: Calculate the prediction error. According to the output O and the expected output Y (true value) obtained by the network prediction, calculate the total error e of the network prediction, e _k is the error generated by the kth output layer node, Step 5: Update weights. Update the network connection weights ω _jk and ω _ij according to the total network prediction error e, ω _jk ⁺ ＝ω _jk +η·H _j ·E _k , where j=1,2,...,l, k=1,2,... , m, η is the learning rate, E _k represents the sensitivity of the network total error of the output layer node to the output layer network node k, Where i=1,2,...,n, j=1,2,...,l.

Step 6: Threshold update. Update the hidden layer threshold a and the output layer threshold b according to the network prediction total error e, j=1,2,...,l; b _k ⁺ =b _k +η·E _k , k=1,2,...,m.

Step 7: Judging whether the algorithm iteration is convergent, if not, return to step 2, and the preferred minimum error of the present invention is 0.001 to end the iteration.

In this embodiment, the voice collection module 1 has a built-in voice collection card for collecting and processing the collected voice signals.

In this embodiment, the fixed recorder 13 adopts a windproof microphone.

In this embodiment, the display screen 8 adopts a touch screen or an LED display screen with a backlight.

In this embodiment, there may be multiple fixed recorders 13 arranged at the casing of the present invention to increase the intensity of voice recording.

The present invention has a long-term storage function for the voice signals that have been entered and marked. Any voice signal stored in the voice pattern database of the present invention can be retrieved at any time with unknown test voices for comparison and recognition.

The use procedure of the present invention is that first turn on the power switch 5, then the system runs automatically, the display screen 8 lights up and displays the operation interface, and people can select two functions of "voice input mode" and "voice test mode".

(1) When selecting voice input, central processing unit 3 can control voice input unit 20 to enter " voice input mode ", and display screen 8 and loudspeaker 35 can prompt simultaneously " now is voice input mode, please speak " similar hint, people Can input voice in any mode of microphone 11 provided by voice collection module 1, wireless walkie-talkie 12, fixed recorder 13 and smart phone; in order to ensure that the present invention can accurately identify and quantify the voice characteristics of the identified object, therefore in "voice In the "recording mode" stage, only one person or one object can be recorded at a time. Since the sound signal data emitted by the same person when speaking and singing will have certain characteristic deviations, in order to improve the accuracy of speech signal recognition, this The invention adopts a multi-state voice input strategy, that is, the recorded voice can include normal speech, singing, or high/middle/bass and other multi-state combined sounds. The recording time is 5 to 30 seconds, and the display 8 will display the real-time waveform of the voice And complete the progress bar, if the voice of recording is unsatisfactory can delete input again, need to carry out data mark after input voice completes, and marking method adopts artificial manual mark, as gathering the sound of Zhang San, namely the dialogue shown in display screen 8 of the present invention Remarks in the frame: "Zhang San's voice", just save it, and the input voice is stored in the memory 33 of the present invention.

(2) After the voice signal has been entered, the control system of the present invention automatically sends the marked voice signal into the voice preprocessing unit 21, and the voice preprocessing unit 21 converts the voice signal collected by the voice acquisition module 1 into an electrical signal, which is about to The analog signal is converted into a digital signal, and then undergoes conventional signal processing, including environmental background noise elimination, signal framing, filtering, pre-emphasis, window function and endpoint detection, etc.

(3) control system of the present invention sends into signal feature extraction unit 22 automatically to speech signal after preprocessing, and speech signal feature extraction unit 22 extracts the feature parameter reflecting speech essence from the speech signal after preprocessing, obtains feature vector x _i , the feature parameter extraction method preferably adopts frequency cepstrum coefficient method (MFCC), and can also use spectral envelope method, LPC interpolation method, LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and extract features The feature vector system obtained after that is automatically saved in the pattern category library. All the voice features of a person correspond to a pattern category. If the voices of N individuals are input, N pattern categories are obtained. Thereby obtain the database of one person's corresponding speech signal pattern class, all data are all stored in the memory 33 of the present invention, so far, speech signal input pattern content is finished.

(4) After the voice entry is complete, the voice test can be carried out. When carrying out the voice test, it is only necessary to select "voice test mode" in the operation interface of the display screen 8, and the central processing unit 3 can control the voice input unit 20 to enter " Voice test mode", display screen 8 and loudspeaker 35 can prompt "in voice test..." similar prompt at the same time, at this moment people do not get to do any operation, the present invention can pass microphone 11, wireless walkie-talkie 12 in voice collection module 1 , the fixed recorder 13 and one or more input tools of the smart phone collect the test voice together, and the test voice collection process is real-time collection without any time limit and number of people.

(5) voice data collected under "voice test mode", the system device of the present invention can automatically carry out preprocessing and feature extraction to the test voice signal, convert the voice test signal collected into an electrical signal, and perform conventional Signal feature extraction is performed after filtering, noise removal, window function and endpoint detection.

(6) After the test signal extracts features, the present invention can automatically carry out feature matching, and the feature parameter of the test voice signal of extraction is carried out feature matching with the marked sample voice feature parameter that has entered in memory 33 of the present invention in real time, and Calculate the similarity between the test voice signal and all the original voice signals that have been recorded, and divide the test voice signal into the pattern category with the highest similarity, and finally the present invention outputs to the outside world, "This is the sound of XXX" similar Report, for example, if the present invention has stored Zhang San's voice feature signal, when Zhang San speaks or sings to the present invention, the present invention will automatically output "this is Zhang San's voice" after identification.

When the present invention is tested in public places, because in the test environment, there may be multiple objects speaking at the same time in the same time period, that is, the voice signal collected is a signal of broadband aliasing. Errors in extraction, the strategy adopted by the present invention is to use intelligent algorithms to first match and identify the voice feature parameters of a single person speaking and then identify and store them, and then the system automatically screens and separates the voice signals when speaking together, Output recognition result at last and report " now is the common voice of Zhang San, Li Si, Wang Wu... " similar hint, and prompts to exist XX voice and fails to recognize, press power off key 6 when shutting down the system.

In this embodiment, the system device can also output to people a list of recognition results in a multi-person communication environment, including the number of people or objects speaking on the spot in the test environment, and screen and play recordings from multiple people speaking at the same time Identify and separate out what each person is saying, while filtering out other people's voices and ambient sounds.

When the voice signal feature of the specimen not stored in the present invention appears in the test voice signal, the present invention will automatically record the unknown voice signal feature to remind people whether to mark and store the voice signal of the object.

In the technical field of intelligent voice signal pattern recognition system device; every frame body 10 is included, the frame body 10 is provided with a cavity, and a voice collection module 1, a voice recognition module 2, a central processing unit 3, a wireless Signal transceiving device 4, display screen 8, memory 33, network module 31, memory card 32, loudspeaker 35 and power supply 9, voice acquisition module 1 includes microphone 11, wireless walkie-talkie 12 and fixed recorder 13, and voice recognition module 2 includes Voice input unit 20, voice preprocessing unit 21, voice signal feature extraction unit 22, feature matching discrimination classification unit 23, voice signal is collected by voice collection module 1, the signal collected is processed by voice recognition module 2, and data signal is processed by memory 33 Preserve, the visualization of the output of the operation process of man-machine interaction and result is shown by display screen 8, loudspeaker 35 is set to carry out voice prompt and broadcast recognition result to operation step, and network module 31 is set to the present invention and Internet cloud platform are connected, The central processing unit 3 is set to the program control and data calculation of the whole system device, and the wireless signal transceiver 4 is set to receive and transmit the radio signals generated by the wireless walkie-talkie 12, the smart phone, and the network module 31 and connect the present invention with the Internet. Wireless connection, memory card 32 is set to the technical content that the recorded external voice data is read in the database of the present invention all within the scope of protection of the present invention, it should be pointed out that the scope of protection of the present invention should not be limited to appearance feature, the present invention The molding of the frame body 10 can be set to square, cylindrical, polygonal prism or other moldings similar to cabbage, watermelon, stone, etc., and all technical contents that are different in molding and substantially technical content are also the same as the present invention. At the same time, those skilled in the art make conventional obvious small improvements or small combinations on the basis of the content of the present invention, as long as the technical content is included in the scope of the content of the present invention. within the scope of protection of the invention.

Claims (8)

1. An intelligent speech signal pattern recognition system device; it is characterized in that: comprise frame (10), described frame (10) is provided with cavity, it is characterized in that, in frame (10) is provided with voice Acquisition module (1), voice recognition module (2), central processing unit (3), wireless signal transceiver (4), display screen (8), memory (33), network module (31), memory card (32) , loudspeaker (35) and power supply (9), voice acquisition module (1) includes microphone (11), wireless walkie-talkie (12) and fixed recorder (13), and voice recognition module (2) includes voice input unit (20 ), speech preprocessing unit (21), speech signal feature extraction unit (22), feature matching discrimination classification unit (23), speech signal is collected by speech collection module (1), and the signal collected is by speech recognition module (2) processing, the data signal is stored by the memory (33), the visualization of the operation process of human-computer interaction and the output of the result is displayed by the display screen (8), the loudspeaker (35) is set to perform voice prompts and broadcast recognition results to the operation steps, and the network module (31) is set to connect the present invention with Internet cloud platform, and central processing unit (3) is set to the program control of whole system device and data is operated, and wireless signal transceiving device (4) is set to wireless walkie-talkie (12) , Smartphone, the radio signal that network module (31) produces receives, transmits and the present invention is wirelessly connected with the Internet, and memory card (32) is set to read the recorded external voice data into the database of the present invention. 2. intelligent voice signal pattern recognition system device according to claim 1; It is characterized in that: voice input unit (20) is set to include " voice entry mode " and " voice test mode " two types, can collect by voice The microphone (11), wireless walkie-talkie (12), fixed recorder (13) and smart phone provided by the module (1) input voice in any mode. In the "voice input mode", the voice input unit (20) is set to Only one person or object can be entered into voice at a time. It is characterized in that the input voice is an audio signal of 5 to 30 seconds. The present invention adopts a multi-state voice input strategy. It is characterized in that the input voice can include For normal speaking, singing, or high/middle/bass multi-state combined voice, the display (8) displays the voice waveform and the completion progress bar in real time. After the input voice is completed, the data needs to be marked. The marking method adopts manual marking. On the display screen (8 ) Remarks in the dialog box shown: "the sound of XXX", save and get final product, the voice of entry is stored in memory (33), under " voice test mode ", pass the microphone (11) in the voice collection module (1) , wireless walkie-talkie (12), fixed recorder (13) and one or more of the input tools in the smart phone to collect the test voice together, the test voice collection process is real-time collection, without any restrictions on the number of people, objects and time, the smart phone The voice acquisition module (1) is set as a wireless matching connection, and the matching methods include Bluetooth, infrared, WIFI, and scanning a QR code to connect to realize voice recording, which is equivalent to using the mobile phone as a wireless microphone to realize multi-group voice interaction. 3. the intelligent voice signal pattern recognition system device according to claim 1; It is characterized in that: the voice preprocessing unit (21) converts the voice signal collected by the voice acquisition module (1) into an electrical signal, that is, converts an analog signal into The digital signal is then subjected to conventional signal processing, including ambient background noise removal, signal framing, filtering, pre-emphasis, windowing, and endpoint detection. 4. intelligent speech signal pattern recognition system device according to claim 1; It is characterized in that: speech signal feature extraction unit (22) is set to extract the main feature parameter that reflects speech essence from original speech signal, forms feature vector x _i , x _i =(x _i1 , x _i2 ,…x _ij ,…,x _in ) ^T , x _ij represents the jth speech feature value of the i-th object or individual, and the feature parameter extraction method adopts the frequency cepstral coefficient method (MFCC), can also use spectral envelope method, LPC interpolation method, LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and the feature vector system obtained after feature extraction will be automatically saved in the pattern database. All sound features of an object or person correspond to a pattern class. If the voices of N individuals or objects are input, N pattern classes are obtained. If each pattern class has n feature parameters, an n-dimensional feature space can be formed, namely The marked feature signal set can be written as D={(x ₁ ,y ₁ ),(x ₂ ,y ₂ ),…( _xi ,y _i ),…,(x _N ,y _N )}, where x _i ∈ χ=R ⁿ , x _i represents the speech feature signal of the i-th object or person entered, y _i ∈ Y={1,2,…,N}, y _i represents the i-th person or object, and N represents The digital number of the Nth person or object, and the marked voice feature data form a pattern database, and are stored in the memory (33) of the present invention. 5. intelligent speech signal pattern recognition system device according to claim 1; It is characterized in that: feature matching distinguishes classification unit (23) and is set to adopt intelligent multiclass classifier, and the learning algorithm of classifier is set to adopt improved neural The network classification algorithm uses the input and marked speech feature signal set as the training data, so that the network model can learn the training data, obtain the classification rules, and complete the training of the classifier; then use the trained classifier to test the unknown Speech signal is carried out intelligent classification and identification; After test signal extraction feature, the present invention can carry out feature matching automatically, the feature parameter of the test speech signal of extracting is in real time with the sample speech that has entered and marked in memory (33) of the present invention The characteristic parameter is carried out feature matching, and calculates the similarity of test speech signal and all the sample speech signals that have entered, then divides test speech signal into that sample signal mode category with the highest similarity with it, and finally the present invention outputs recognition to the outside world As a result, "this is the voice of XXX" similar reports. 6. intelligent speech signal pattern recognition system device according to claim 1; It is characterized in that: the multi-layer artificial neural network structure that multiclass classifier adopts, it is characterized in that, one end of network is defined as input layer, and the other end is defined as The output layer, the part between the input layer and the output layer is defined as the hidden layer, the input layer is used to receive the input signal from the outside world, resend the input signal to all neurons in the hidden layer, and pass the result after calculation by the hidden layer For the output layer, the output layer receives signals from the hidden layer, outputs the classification result after calculation, that is, the recognition result, and the number of layers of the preferred hidden layer in the present invention is set to 1-200 layers. 7. intelligent speech signal pattern recognition system device according to claim 1; It is characterized in that: the step of improving artificial neural network training is as follows: Step 1: Network initialization; according to the number of input voice signals, the algorithm database is continuously updated. When the voice signals of N objects are input, N pattern classes are formed, and the sample space (X, Y) is obtained. The i-th group of samples Namely (X _i , Y _i ), Xi _i represents the set of feature vectors extracted from the i-th object, and Y _i represents the marked i-th object; determine the network input layer structure according to the system input and output sequence (X, Y) The number of points n, the number of hidden layer nodes l, and the number of output layer nodes m, wherein the value of n is determined by the number of corresponding feature values in the feature extraction of the input signal, the value of m is determined by the number of stored speech pattern classes, and the reference value of l for The value of a ranges from 0 to 10, which is automatically calculated and determined by the model, and the connection weight ω _ij between the neurons of the input layer and the hidden layer and the connection weight between the neurons of the hidden layer and the output layer are initialized ω _jk , initialize hidden layer threshold a and output layer threshold b, given learning rate η and neuron activation function; Step 2: Calculate the output of the hidden layer; according to the input change X, the connection weight ω _ij of the neurons in the input layer and the hidden layer, and the hidden layer threshold a, calculate the output H of the hidden layer; record the jth hidden layer The output of the layer-containing node is H _j , j=1,2,...,l, where l is the number of nodes in the hidden layer, f is the hidden layer activation function, and there are many kinds of activation functions, and the present invention preferably adopts f(x) ⁼ (1+e- ^x ) ^-1 ; Step 3: Calculate the output of the output layer; according to the hidden layer output H, the connection weight ω _jk between the hidden layer and the output layer neurons, and the output layer threshold b, calculate the output layer output O, record the kth output The output of the layer node is O _k , k=1,2,...,m, where m is the number of nodes in the output layer, b _k is the threshold of the kth node in the output layer, and Hj is the output value of the _jth node in the hidden layer; Step 4: Calculate the prediction error; calculate the total error e of the network prediction according to the output O and the expected output Y (true value) obtained by the network prediction, e _k is the error generated by the kth output layer node, <mrow><msub><mi>e</mi><mi>k</mi></msub><mo>=</mo><mfrac><mn>1</mn><mn>2</mn></mfrac><msup><mrow><mo>(</mo><msub><mi>Y</mi><mi>k</mi></msub><mo>-</mo><msub><mi>O</mi><mi>k</mi></msub><mo>)</mo></mrow><mn>2</mn></msup><mo>;</mo></mrow> Step 5: Update the weights; update the network connection weights ω _jk and ω _ij according to the total network prediction error e, ω _jk ⁺ =ω _jk +η·H _j ·E _k , where j=1,2,...,l, k=1,2,...,m, η is the learning rate, E _k represents the sensitivity of the network total error of the output layer node to the output layer network node k, where i=1,2,...,n, j=1,2,...,l; Step 6: Threshold update; update the hidden layer threshold a and the output layer threshold b according to the network prediction total error e, j=1,2,...,l; b _k ⁺ =b _k +η·E _k , k=1,2,...,m; Step 7: Judging whether the algorithm iteration is convergent, if not, return to step 2, and the preferred minimum error of the present invention is 0.001 to end the iteration. 8. according to claim 1,2,3,4,5,6,7 any described intelligent voice signal pattern recognition system device; It is characterized in that: basic operating system flow process is set to: 1) Turn on the power switch (5), then the system runs automatically, the display screen (8) lights up and displays the operation interface, and people can select two functions of "voice input mode" and "voice test mode"; when selecting voice input, Central processing unit (3) can control voice input unit (20) to enter " voice recording mode ", and display screen (8) and loudspeaker (35) can prompt simultaneously " be voice recording mode now, please speak " similar prompting, and people can By the microphone (11) provided by the voice collection module (1), wireless walkie-talkie (12), fixed recorder (13) and intelligent mobile phone any mode input voice; In order to ensure that the present invention can accurately identify and quantify the identified object Voice characteristics, so in the "voice recording mode" stage, only one person or one object can be recorded at a time. Since the same person speaks and sings, the sound signal data will have certain characteristic deviations. Therefore, in order to improve the voice For the accuracy of signal recognition, the present invention adopts a multi-state voice input strategy, that is, the input voice can include normal speech, singing, or high/middle/bass and other states of multi-state combination sounds, and the recording time is 5 to 30 seconds. Display (8) can display voice real-time waveform and complete progress bar, if the recorded voice is unsatisfactory, it can be deleted and entered again. After the input voice is completed, data marking is required. The marking method adopts manual marking, such as collecting the voice of Zhang San, that is Remark in the dialog box that display screen (8) of the present invention shows: " Zhang San's voice ", save and get final product, the voice of typing is stored in memory (33) of the present invention; 2) After the voice signal has been entered, the control system of the present invention automatically sends the marked voice signal into the voice preprocessing unit (21), and the voice preprocessing unit (21) converts the voice signal collected by the voice acquisition module (1) It is an electrical signal, that is, the analog signal is converted into a digital signal, and then conventional signal processing is performed, including environmental background noise elimination, signal framing, filtering, pre-emphasis, windowing function and endpoint detection, etc.; 3) control system of the present invention sends into signal feature extraction unit (22) automatically with preprocessed speech signal, and speech signal feature extraction unit (22) extracts the feature parameter reflecting speech essence from the speech signal after preprocessing, The feature vector x _i is obtained, and the feature parameter extraction method adopts frequency cepstral coefficient method (MFCC), and can also use spectral envelope method, LPC interpolation method, LPC root-finding method, Hilbert transform method, etc. to obtain acoustic features, and extract The feature vector system obtained after the feature is automatically saved in the pattern category library. All the voice features of a person correspond to a pattern category. If the voices of N people are input, N pattern categories will be obtained. If each pattern category has n feature parameters , thereby obtain the database of one people's corresponding speech signal pattern class, all data are all stored in the memory (33) of the present invention, so far, speech signal input pattern content is finished; 4) After the voice input is completed, the voice test can be carried out. When performing the voice test, you only need to select "voice test mode" in the operation interface of the display screen (8), and the central processing unit (3) will control the voice input unit (20) Enter " voice test mode ", display screen (8) and loudspeaker (35) can prompt simultaneously " in voice test... " similar hint, at this moment people do not get to do any operation, the present invention can pass voice acquisition module The microphone (11), wireless walkie-talkie (12), fixed recorder (13) and one or more input tools in the smart phone in (1) collect the test voice together, and the test voice collection process is real-time collection without any time Restrictions and limitations on numbers of people; 5) For the voice data collected under the "voice test mode", the system device of the present invention will automatically perform preprocessing and feature extraction on the test voice signal, convert the collected voice test signal into an electrical signal, and perform conventional filtering , signal feature extraction after noise removal, window function and endpoint detection; 6) after the test signal extraction feature, the present invention can automatically carry out feature matching, the feature parameter of the test voice signal of extraction is carried out feature matching with the sample voice feature parameter that has entered in the memory (33) of the present invention in real time, And calculate the degree of similarity between the test voice signal and all recorded original voice signals, and classify the test voice signal into the highest pattern category with the highest similarity, and finally the present invention outputs to the outside world, "this is the sound of XXX" is similar For example, if the present invention has stored Zhang San's voice feature signal, when Zhang San speaks or sings to the present invention, the present invention will automatically output "this is Zhang San's voice" after recognition; 7) The system device can also output to people a list of recognition results in a multi-person communication environment, including the number of people or objects speaking on the spot in the test environment, as well as screening and playing the recognition and separation from the recordings of multiple people speaking at the same time. What each person said, while filtering out other people's voices and ambient sounds, when the voice signal features of the specimen that are not stored in the present invention appear in the test voice signal, the present invention will automatically record the unknown voice signal features to remind Whether people label and store the speech signal of the object.