Artificial intelligence heat supply customer service system based on voice recognition

Abstract

The invention relates to the technical field of voice recognition, and provides an artificial intelligence heat supply customer service system based on voice recognition, and the system comprises a heat supply problem classification module which carries out the keyword recognition and classification of the voice needed by a user through a lightweight voice classification model based on a clustering algorithm, and generates a heat supply consultation MFCC and a heat supply appeal MFCC; the heat supply consultation reply module extracts a heat supply semantic problem from the heat supply consultation MFCC through a heavy-weight speech semantic recognition model based on a fused convolutional neural network and a long-short-term memory network, performs information base retrieval on the heat supply semantic problem, determines and sends reply information; and the heat supply appeal storage module performs text transcription on the heat supply appeal MFCC through a voice recognition model to generate an appeal content text and stores the appeal content text. According to the method, the computing power resources of the voice recognition model are allocated in the customer service system in combination with the real situation of the required voice of the user, and rapid and high-accuracy reply of the heat supply customer service is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to an artificial intelligence heating customer service system based on speech recognition. Background Art

[0002] In the current era of rapid technological development, speech recognition technology has shown great application potential in various fields and has become one of the hot topics in artificial intelligence research. With the increasing diversification of human-computer voice interaction methods, the time for speech recognition to be applied to the sub-field of heating customer service is gradually ripe.

[0003] At present, with the enhancement of environmental awareness and the adjustment of energy structure, heating policies are also constantly being adjusted and optimized. When the heating season arrives, users' demand for consulting new heating policies and reporting heating repairs to heating customer service increases rapidly. Most of these demands are similar, so there are realistic conditions for classifying and responding to demands through voice recognition technology.

[0004] Therefore, how to achieve lightweight classification of user heating problems and heavyweight, high-accuracy speech semantic recognition responses through an artificial intelligence model based on speech recognition is a technical problem that needs to be solved. Summary of the invention

[0005] To this end, the present invention provides an artificial intelligence heating customer service system based on speech recognition, which classifies the demand voices spoken by users guided by the intelligent heating customer service through a lightweight speech classification model based on a clustering algorithm, and extracts accurate heating semantic questions from the heating consultation MFCC through a heavyweight speech semantic recognition model based on a fusion convolutional neural network and a long short-term memory network, thereby allocating the computing power resources of the speech recognition model in the customer service system in combination with the actual situation of the user's demand voice, and achieving fast and high-accuracy responses from the heating customer service.

[0006] To achieve the above objectives, the present invention proposes an artificial intelligence heating customer service system based on speech recognition, comprising:

[0007] The heating problem classification module is used to perform keyword recognition and classification on the user's demand voice through a lightweight speech classification model based on a clustering algorithm to generate heating consultation MFCC and heating demand MFCC;

[0008] The heating consultation reply module is used to extract the heating semantic problem from the heating consultation MFCC through a heavyweight speech semantic recognition model based on a fusion convolutional neural network and a long short-term memory network, perform information database retrieval to determine the heating semantic problem, and send a reply message;

[0009] The heating demand storage module is used to transcribe the heating demand MFCC into text through a speech recognition model to generate and store the demand content text.

[0010] Furthermore, the lightweight speech classification model is constructed based on the K-Means clustering algorithm, and the heating problem classification module includes a preprocessing unit, a centroid setting unit and a clustering calculation module;

[0011] The pre-processing unit is used to convert the demand speech into a plurality of reply MFCC segments;

[0012] The centroid setting unit is used to set the initial centroid of the lightweight speech classification model based on the K-Means clustering algorithm according to multiple keywords of the heating speech sample library;

[0013] The clustering calculation module is used to perform clustering calculation and classification on the reply MFCC segments through the lightweight speech classification model to generate the heating consultation MFCC and the heating demand MFCC.

[0014] Further, the pre-processing unit includes a dynamic sampling sub-unit, a dynamic pre-emphasis sub-unit and a conversion sub-unit;

[0015] The dynamic sampling subunit is used to collect multiple voice quality sampling values ​​of the required voice at multiple time points;

[0016] The dynamic pre-emphasis subunit is used to calculate an emphasis coefficient according to a plurality of the speech quality sampling values, and to enhance the required speech by setting a filter algorithm of the emphasis coefficient to generate an enhanced required speech;

[0017] The conversion subunit is used to convert the enhanced demand speech into a plurality of the reply MFCC segments.

[0018] Furthermore, the filter algorithm includes a Hamming window function filtering algorithm and a high frequency enhancement algorithm;

[0019] The dynamic pre-emphasis subunit is used to calculate the window function length and the emphasis coefficient according to the speech quality sampling value, and adjust the amplitude-frequency response of the required speech by setting the Hamming window function filtering algorithm of the window function length, and adjust the high-frequency component of the required speech by setting the high-frequency enhancement algorithm of the emphasis coefficient to generate the enhanced required speech.

[0020] In the above scheme, by adjusting the parameters of the lightweight voice classification model based on the K-Means clustering algorithm and its preprocessing method for the reply voice, the optimal effect of applying the lightweight voice classification model to artificial intelligence heating customer service is achieved.

[0021] Further, the heavyweight speech semantic recognition model includes a convolutional neural network feature optimizer, a long short-term memory network optimizer and a heating problem classifier;

[0022] The heating consultation reply module is used to extract primary speech features from the heating consultation MFCC through the convolutional neural network feature optimizer, extract high-level temporal features from the primary speech features through the long short-term memory network optimizer, and extract the heating semantic problem from the high-level temporal features through the heating problem classifier.

[0023] Furthermore, the heating problem classifier includes a fully connected layer and a Softmax activation function layer;

[0024] The heating consultation and response module is used to generate a classification map of the heating problem through the fully connected layer using the time series high-level features, and classify the classification map through the Softmax activation function layer to generate the heating semantic problem.

[0025] Further, the long short-term memory network optimizer includes a hybrid attention mechanism;

[0026] The heating consultation response module is used to integrate and update the reset gate, update gate, candidate hidden state and hidden state of the long short-term memory network through the hybrid attention mechanism.

[0027] Further, the heating consultation reply module includes a question expansion unit and a question search unit;

[0028] The question expansion unit is used to select the heating semantic question from the heating similar question library through a search engine to expand the heating semantic question;

[0029] The question search unit is used to search the information base through the expanded heating semantic question to determine the reply information.

[0030] In the above scheme, a powerful deep learning model is formed by improving the heavyweight speech semantic recognition model of the fusion neural network and the long short-term memory network, which combines the advantages of the convolutional neural network in feature extraction and the strengths of the long short-term memory network in processing time series data, so that the output of the clustering algorithm can achieve the best recognition rate and processing speed with high accuracy.

[0031] Furthermore, the speech transcription model is a speech recognition library based on grammatical constraints of the heating language.

[0032] Further, the keywords include heating policy consultation related words and heating fault repair progress related words corresponding to the heating consultation MFCC, and heating fault repair application related words and heating complaint related words corresponding to the heating appeal MFCC;

[0033] The heating business inquiry module is also used to confirm the classification result of the heating problem classification module for the keyword to the user by voice.

[0034] In the above solution, the segmented application of artificial intelligence customer service in the field of heating customer service is realized, which greatly facilitates users to obtain corresponding voice information and reduces the cost of customer service companies.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. A lightweight speech classification model based on a clustering algorithm is used to classify the demand voices spoken by users guided by the intelligent heating customer service. The heating consultation MFCC is extracted with accurate heating semantic questions through a heavyweight speech semantic recognition model based on a fusion of convolutional neural networks and long short-term memory networks. The computing power resources of the speech recognition model are allocated in the customer service system in combination with the actual situation of the user's demand voice, and the heating customer service can respond quickly and accurately.

[0037] 2. By adjusting the parameters of the lightweight voice classification model based on the K-Means clustering algorithm and its preprocessing method for the reply voice, the optimal effect of applying the lightweight voice classification model to artificial intelligence heating customer service is achieved.

[0038] 3. A powerful deep learning model is realized by improving the heavyweight speech semantic recognition model that integrates neural networks and long short-term memory networks. It combines the advantages of convolutional neural networks in feature extraction and the strengths of long short-term memory networks in processing time series data, so that the output of the clustering algorithm can achieve the best recognition rate and processing speed with high accuracy.

[0039] 4. It has realized the segmented application of artificial intelligence customer service in the field of heating customer service, which greatly facilitates users to obtain corresponding voice information and reduces the cost of customer service companies. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a structural diagram of an artificial intelligence heating customer service system based on speech recognition according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the overall process of an artificial intelligence heating customer service system based on speech recognition according to an embodiment of the present invention;

[0042] Figure 3This is a schematic diagram of the flow of a heating consultation reply module of an artificial intelligence heating customer service system based on speech recognition according to an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the overall model structure of an artificial intelligence heating customer service system based on speech recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the objects and advantages of the present invention more clearly understood, the present invention is further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0045] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the protection scope of the present invention.

[0046] It should be noted that, in the description of the present invention, terms such as "up", "down", "left", "right", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is merely for the convenience of description and does not indicate or imply that the device or element must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present invention.

[0047] In addition, it should be noted that in the description of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0048] like Figures 1 to 4 As shown, the present invention provides an artificial intelligence heating customer service system based on speech recognition, which classifies the demand voices spoken by users guided by the intelligent heating customer service through a lightweight speech classification model based on a clustering algorithm, and extracts accurate heating semantic questions from the heating consultation MFCC through a heavyweight speech semantic recognition model based on a fusion of convolutional neural networks and long short-term memory networks, thereby allocating the computing power resources of the speech recognition model in the customer service system in combination with the actual situation of the user's demand voice, and achieving fast and high-accuracy replies from the heating customer service.

[0049] like Figures 1 to 4As shown, this embodiment proposes an artificial intelligence heating customer service system based on speech recognition, including: a heating problem classification module, which is used to perform keyword recognition and classification on the user's demand voice through a lightweight speech classification model based on a clustering algorithm to generate a heating consultation MFCC and a heating appeal MFCC; a heating consultation reply module, which is used to extract heating semantic problems from the heating consultation MFCC through a heavyweight speech semantic recognition model based on a fusion convolutional neural network and a long short-term memory network, and perform information database retrieval to determine the heating semantic problems and send reply information; a heating appeal storage module, which is used to transcribe the heating appeal MFCC into text through a speech recognition model to generate and store the appeal content text.

[0050] It is understandable that if Figure 4 As shown in the figure, the heavyweight speech semantic recognition model based on the fusion of convolutional neural networks (CNN) and long short-term memory networks (LSTM) constitutes a powerful deep learning model, which combines the advantages of convolutional neural networks in feature extraction and the advantages of long short-term memory networks in processing time series data. However, as the network weight increases, the amount of calculation also increases, resulting in the heavyweight speech semantic recognition model being used to recognize the entire speech. There are problems of large computing power requirements and high computing costs. Therefore, this embodiment uses a lightweight speech classification model of a clustering algorithm to identify, classify and preliminarily process the heating consultation MFCC for speech semantic recognition, which greatly reduces the computing power requirements of the semantic model. At the same time, the results of the lightweight speech classification model based on the clustering algorithm can be used for the processing of the heavyweight semantic model, reducing the computational complexity of the overall model, and realizing the intelligent selection of the heating consultation MFCC (Mel-Frequency Cepstral Coefficients (Mel-frequency cepstral coefficients) are used for heating semantic problem recognition, which effectively reduces the processing time of heavyweight semantic models and ensures that the model has accurate recognition performance and can be easily recognized. At the same time, in the tests of using clustering algorithm alone, using fusion convolutional neural network and long short-term memory network algorithm alone, using fusion convolutional neural network algorithm alone, using long short-term memory network algorithm alone, and using clustering algorithm and fusion convolutional neural network and long short-term memory network algorithm together, the recognition rate and processing speed are the best, and the error matrix is ​​the smallest, indicating high accuracy.

[0051] Preferably, the demand voice is the user's reply voice to the question voice of the intelligent customer service system, so that the user can be guided to better express the demand through the interaction between the intelligent customer service and the user.

[0052] It should be noted that the inquiry voice corresponds to the keyword as a fixed question used to guide the user to express needs close to the keyword. For example, the heating service inquiry module sends a voice to the user at the user end asking whether he needs to know the start and end time of heating. If the user answers yes, the heating time policy of the user's location in the user information database is directly informed to the user. If the user answers that he wants to know the start / end time of heating in a certain place, the heating problem classification module classifies it into the heating consultation MFCC based on the keywords of heating start and start time. The heating consultation reply module classifies the certain place in the heating consultation MFCC into the closest geographical location based on the characteristic information such as province, county, prefecture-level city, and district. Considering the personal pronunciation habits and voice quality, the accuracy of direct voice transcription to text is not high enough. The heating time policy of the geographical location is informed to the user, and the user is asked whether he is satisfied. If the user is not satisfied, the heating service inquiry module repeats a new inquiry voice that is different from the first time. Therefore, it is realized to build an intelligent heating customer service that can autonomously answer common heating questions.

[0053] Furthermore, if Figure 2 As shown, the lightweight speech classification model is constructed based on the K-Means clustering algorithm, and the heating problem classification module includes a preprocessing unit, a centroid setting unit and a clustering calculation module;

[0054] The pre-processing unit is used to convert the demand speech into a plurality of reply MFCC segments;

[0055] The centroid setting unit is used to set the initial centroids of the plurality of lightweight speech classification models according to the plurality of keywords of the heating speech sample library;

[0056] The clustering calculation module is used to perform clustering calculation and classification on the reply MFCC segments through the lightweight speech classification model to generate the heating consultation MFCC and the heating demand MFCC.

[0057] Furthermore, if Figure 2 As shown, the keywords include heating policy consultation related words and heating fault repair progress related words corresponding to the heating consultation MFCC, as well as heating fault repair application related words and heating complaint related words corresponding to the heating appeal MFCC; the heating business inquiry module is also used to confirm the classification results of the heating problem classification module for the keywords to the user by voice.

[0058] Specifically, Figure 2As shown, the preprocessing unit as a whole is used to perform sampling, quantization and encoding, and the sampling of the required speech with a set sampling frequency is performed through the Nyquist sampling theorem to generate an analog audio signal, quantization includes fast Fourier transforming the analog audio signal to obtain a spectrum, and encoding includes processing the spectrum through a group of Mel filters to generate a reply MFCC (Mel-Frequency Cepstral Coefficients) fragment. Among them, the quantization process adds a dynamic pre-emphasis algorithm based on the required speech quality to enhance the high-frequency components and compensate for the microphone characteristics. Preferably, the sampling frequency of the sampling process is set to 31.5kHz.

[0059] Specifically, the clustering calculation process of the clustering calculation module is a conventional process, including the following pseudocode content: when the cluster assignment result of any point changes: for each data point in the data set, for each centroid: calculate the distance between the centroid and the data point and assign the data point to the cluster closest to it; for each cluster, calculate the mean of all points in the cluster, and use the mean as the centroid.

[0060] It can be understood that the initial centroid point k of the conventional K-Means clustering algorithm is randomly selected, while the initial centroid point k is set according to a plurality of related words in this embodiment.

[0061] Furthermore, if Figure 2 As shown, the preprocessing unit includes a dynamic sampling subunit, a dynamic pre-emphasis subunit and a conversion subunit; the dynamic sampling subunit is used to collect multiple voice quality sampling values ​​of the required voice at multiple time points; the dynamic pre-emphasis subunit is used to calculate the emphasis coefficient according to the multiple voice quality sampling values, and enhance the required voice by setting the filter algorithm of the emphasis coefficient to generate enhanced required voice; the conversion subunit is used to convert the enhanced required voice into multiple reply MFCC segments.

[0062] Furthermore, if Figure 2 As shown, the filter algorithm includes a Hamming window function filtering algorithm and a high-frequency enhancement algorithm; the dynamic pre-emphasis subunit is used to calculate the window function length and the emphasis coefficient according to the speech quality sampling value, and adjust the amplitude-frequency response of the required speech by setting the Hamming window function filtering algorithm with the window function length, and adjust the high-frequency component of the required speech by setting the high-frequency enhancement algorithm with the emphasis coefficient to generate the enhanced required speech.

[0063] Specifically, the improved Hamming window function of this example can be defined as:

[0064] s′(n)=s(n)-αs(n-1)

[0065] In the formula, s′(n), s(n), s(n-1) are respectively the required speech signal with emphasized high frequency of the n-th frame sample, the input speech signal of the n-th frame sample and the input speech signal of the n-1-th frame sample, and α is the emphasis coefficient, preferably between 0.86 and 1.05.

[0066] The calculation process of the weighting coefficient is:

[0067] α=|w n+2 (n)-w n+1 (n)|∩|w n+1 (n)-w n (n)|∩|w n (n)-w n-1 (n)|∩|w n-1 (n)-w n-2 (n)|

[0068] In the formula, w n+2 (n), w n+1 (n), w n (n), w n-1 (n), w n-2 (n) represents the speech quality sampling values ​​of the n+2th frame, the n+1th frame, the nth frame, the n-1th frame and the n-2th frame respectively, and α is the emphasis coefficient.

[0069] It can be understood that in this example, the dynamic value of the emphasis coefficient is implemented to set different high-frequency emphasis levels according to different voice qualities.

[0070] Specifically, the Hamming window function filtering algorithm of this example is a conventional process, and only the window function length is dynamically valued. The relationship between the window function length and the speech quality sampling value is:

[0071]

[0072] In the formula, w n+2 (n), w n+1 (n), w n (n), w n-1 (n), w n-2 (n) represents the speech quality sampling values ​​of the n+2th frame, the n+1th frame, the nth frame, the n-1th frame and the n-2th frame respectively, and N is the window function length. It can be understood that a too long window function length will lead to spectrum leakage and aliasing, and a too short window function length will distort the high-frequency components in the signal. Therefore, by dynamically adjusting the window function length, the input sampling value is made more accurate.

[0073] In the above scheme, by adjusting the parameters of the lightweight voice classification model based on the K-Means clustering algorithm and its preprocessing method for the reply voice, the optimal effect of applying the lightweight voice classification model to artificial intelligence heating customer service is achieved.

[0074] Furthermore, if Figure 3 and 4 As shown, the heavyweight speech semantic recognition model includes a convolutional neural network feature optimizer, a long short-term memory network optimizer and a heating problem classifier;

[0075] The heating consultation reply module is used to extract primary speech features from the heating consultation MFCC through the convolutional neural network feature optimizer, extract high-level temporal features from the primary speech features through the long short-term memory network optimizer, and extract the heating semantic problem from the high-level temporal features through the heating problem classifier.

[0076] Furthermore, if Figure 3 and 4 As shown, the heating problem classifier includes a fully connected layer and a Softmax activation function layer; the heating consultation response module is used to generate a classification map of the heating problem through the fully connected layer for the time series high-level features, and classify the classification map through the Softmax activation function layer to generate the heating semantic problem.

[0077] Furthermore, if Figure 3 and 4 As shown, the long short-term memory network optimizer includes a hybrid attention mechanism; the heating consultation response module is used to integrate and update the forget gate, input gate, output gate, memory unit and hidden state of the long short-term memory network through the hybrid attention mechanism.

[0078] Specifically, the convolutional neural network includes: two groups of 7*7 convolution kernels and Relu activation function convolution operations to extract features; a maximum pooling layer to downsample the output of the convolution layer, reduce the number of parameters and the amount of calculation; a convolutional fully connected layer to connect the outputs of the convolution layer and the pooling layer. It can be understood that the convolutional neural network can perform a certain degree of time domain feature extraction and better frequency domain feature extraction to better understand the characteristics of the speech signal.

[0079] It is understandable that in speech recognition tasks, convolutional neural networks are combined with long short-term memory networks to fully utilize the advantages of convolutional neural networks in feature extraction and the ability of long short-term memory networks in temporal modeling. This combined model can effectively recognize and classify speech signals.

[0080] Specifically, Figure 4As shown, the long short-term memory network includes a forget gate z t , input gate r t 、Memory unit c t and the hidden state h t ; The long short-term memory network improves the linear transformation of the input vector x and the hidden state h on the basis of the original gated recurrent unit: the hidden state and the gate matrix are both represented by ~. First, define the input variable X passing through the input gate, T represents the time step, and the input vector x at the tth time step t , N represents the number of variables, and the corresponding weight Where d represents the number of neurons in the layer. During the operation, x t Each variable in Each column is element-wise multiplied (dot product) to obtain the output matrix o x , o x ∈R n×d For the hidden state The linear transformation of The corresponding weight matrix and Perform a standard matrix multiplication operation to obtain the output matrix o h Through the cyclic accumulation of time steps, the overall hidden state The construction method preserves the independent characteristics of each input variable at all time steps, forming a complete hidden state representation. can be viewed as consisting of independent hidden states at each time step The sequence composed of in The hidden state output by the LSTM network will be further integrated and feature extracted by the subsequent hybrid attention mechanism.

[0081] Specifically, the variable-by-variable time attention mechanism dynamically allocates the weights of the influence of different time steps on the overall sequence representation. The variable-by-variable time attention mechanism calculates the relative contribution of the hidden state of each time step in the sequence through a weighted sum operation. Based on the global feature representation generated by the variable-by-variable time attention mechanism, the model further applies the feature attention mechanism to dynamically adjust the relative contribution of each input feature in the sequence. The feature attention mechanism concatenates the time attention output with the hidden state of the current time step through a linear transformation, and calculates the attention weight of each feature.

[0082] Furthermore, the heating consultation and response module includes a question expansion unit and a question search unit; the question expansion unit is used to select the expanded heating semantic question from the heating similar question library through a search engine; the question search unit is used to search the information library through the expanded heating semantic question to determine the reply information.

[0083] Preferably, the search engine is SQL Server, which integrates full-text search and query optimization mechanisms and is particularly suitable for cloud servers equipped with speech recognition models.

[0084] In the above scheme, a powerful deep learning model is formed by improving the heavyweight speech semantic recognition model of the fusion neural network and the long short-term memory network, which combines the advantages of the convolutional neural network in feature extraction and the strengths of the long short-term memory network in processing time series data, so that the output of the clustering algorithm can achieve the best recognition rate and processing speed with high accuracy.

[0085] Furthermore, the speech transcription model is a speech recognition library based on grammatical constraints of the heating language.

[0086] Preferably, the speech recognition library is Vosk, which is a lightweight open source speech recognition library based on Kald i and TensorFlow, provides real-time speech recognition function, supports multiple dialect recognition, maintains small resource usage, and is particularly suitable for cloud servers equipped with speech recognition models.

[0087] Furthermore, the keywords include heating policy consultation related words and heating fault repair progress related words corresponding to the heating consultation MFCC, as well as heating fault repair application related words and heating complaint related words corresponding to the heating appeal MFCC; the heating business inquiry module is also used to confirm to the user by voice the classification results of the heating problem classification module for the keywords.

[0088] Specifically, the heating policy consultation related words include: heating policy words, heating time words, heating cost words, which are used to solve users' questions about heating / heating process. The heating fault repair progress related words include repair process words, which are used to solve users' heating and heating maintenance progress, such as the maintenance progress of outdoor pipes, boilers and other facilities, and the door-to-door plan progress of indoor maintenance personnel. The heating fault repair application related words include: repair request words, heating pipeline problem words, heating effect words, etc., which are used to automatically generate equipment damage repair applications, which facilitates the intelligent customer service to record and count heating facility damage and repair requests. The heating complaint related words are used to record user complaints about heating services, such as insufficient indoor heating temperature, etc.

[0089] In the above solution, the segmented application of artificial intelligence customer service in the field of heating customer service is realized, which greatly facilitates users to obtain corresponding voice information and reduces the cost of customer service companies.

[0090] In this embodiment, the light-weight speech classification model based on the clustering algorithm is used to classify the demand speech guided by the user by the intelligent heating customer service, and the heating consultation MFCC is extracted by the heavy-weight speech semantic recognition model based on the fusion convolutional neural network and the long short-term memory network to extract the accurate heating semantic problem, so as to realize the allocation of the computing power resources of the speech recognition model in the customer service system in combination with the actual situation of the user's demand speech, and realize the fast and high-accuracy reply of the heating customer service. By adjusting the parameters of the light-weight speech classification model based on the K-Means clustering algorithm and its pre-processing method for the reply speech, the best effect of the light-weight speech classification model applied to the artificial intelligence heating customer service is achieved. It is realized that the heavy-weight speech semantic recognition model formed by improving the fusion neural network and the long short-term memory network constitutes a powerful deep learning model, which combines the advantages of the convolutional neural network in feature extraction and the advantages of the long short-term memory network in processing time series data, so that it is used to process the output of the clustering algorithm to achieve the best recognition rate and processing speed, and the accuracy is high. It realizes the subdivision application of artificial intelligence customer service in the field of heating customer service, which greatly facilitates users to obtain corresponding voice information and reduces the cost of customer service enterprises.

[0091] So far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present invention.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An artificial intelligence heating customer service system based on voice recognition, characterized in that: include: The heating problem classification module is used to perform keyword recognition and classification on the user's demand voice through a lightweight speech classification model based on a clustering algorithm to generate heating consultation MFCC and heating demand MFCC; The heating consultation reply module is used to extract the heating semantic problem from the heating consultation MFCC through a heavyweight speech semantic recognition model based on a fusion convolutional neural network and a long short-term memory network, perform information database retrieval to determine the heating semantic problem, and send a reply message; The heating demand storage module is used to transcribe the heating demand MFCC into text through a speech recognition model to generate and store the demand content text.

2. The artificial intelligence heating customer service system based on voice recognition according to claim 1 is characterized in that: The heating problem classification module includes a preprocessing unit, a centroid setting unit and a clustering calculation module; The pre-processing unit is used to convert the demand speech into a plurality of reply MFCC segments; The centroid setting unit is used to set the initial centroid of the lightweight speech classification model based on the K-Means clustering algorithm according to multiple keywords of the heating speech sample library; The clustering calculation module is used to perform clustering calculation and classification on the reply MFCC segments through the lightweight speech classification model to generate the heating consultation MFCC and the heating demand MFCC.

3. The artificial intelligence heating customer service system based on voice recognition according to claim 2 is characterized in that: The pre-processing unit includes a dynamic sampling sub-unit, a dynamic pre-emphasis sub-unit and a conversion sub-unit; The dynamic sampling subunit is used to collect multiple voice quality sampling values ​​of the required voice at multiple time points; The dynamic pre-emphasis subunit is used to calculate an emphasis coefficient according to a plurality of the speech quality sampling values, and to enhance the required speech by setting a filter algorithm of the emphasis coefficient to generate an enhanced required speech; The conversion subunit is used to convert the enhanced demand speech into a plurality of the reply MFCC segments.

4. The artificial intelligence heating customer service system based on voice recognition according to claim 3 is characterized in that: The filter algorithm includes a Hamming window function filtering algorithm and a high frequency enhancement algorithm; The dynamic pre-emphasis subunit is used to calculate the window function length and the emphasis coefficient according to the speech quality sampling value, and adjust the amplitude-frequency response of the required speech by setting the Hamming window function filtering algorithm of the window function length, and adjust the high-frequency component of the required speech by setting the high-frequency enhancement algorithm of the emphasis coefficient to generate the enhanced required speech.

5. The artificial intelligence heating customer service system based on voice recognition according to claim 1 is characterized in that: The heavyweight speech semantic recognition model includes a convolutional neural network feature optimizer, a long short-term memory network optimizer and a heating problem classifier; The heating consultation reply module is used to extract primary speech features from the heating consultation MFCC through the convolutional neural network feature optimizer, extract high-level temporal features from the primary speech features through the long short-term memory network optimizer, and extract the heating semantic problem from the high-level temporal features through the heating problem classifier.

6. The artificial intelligence heating customer service system based on voice recognition according to claim 5 is characterized in that: The heating problem classifier includes a fully connected layer and a Softmax activation function layer; The heating consultation and response module is used to generate a classification map of the heating problem through the fully connected layer using the time series high-level features, and classify the classification map through the Softmax activation function layer to generate the heating semantic problem.

7. The artificial intelligence heating customer service system based on voice recognition according to claim 5 is characterized in that: The LSTM optimizer includes a hybrid attention mechanism; The heating consultation response module is used to integrate and update the reset gate, update gate, candidate hidden state and hidden state of the long short-term memory network through the hybrid attention mechanism.

8. The artificial intelligence heating customer service system based on voice recognition according to claim 1 is characterized in that: The heating consultation and reply module includes a question expansion unit and a question search unit; The question expansion unit is used to select the heating semantic question from the heating similar question library through a search engine to expand the heating semantic question; The question search unit is used to search the information base through the expanded heating semantic question to determine the reply information.

9. The artificial intelligence heating customer service system based on voice recognition according to claim 1 is characterized in that: The speech transcription model is a speech recognition library based on grammatical constraints of the heating language.

10. The artificial intelligence heating customer service system based on voice recognition according to claim 1 is characterized in that: The keywords include heating policy consultation related words and heating fault repair progress related words corresponding to the heating consultation MFCC, and heating fault repair application related words and heating complaint related words corresponding to the heating appeal MFCC; The heating business inquiry module is also used to confirm the classification result of the heating problem classification module for the keyword to the user by voice.

Images