Intelligent skin care product recommendation platform and method based on big data speech recognition technology

Abstract

The present invention discloses an intelligent skin care product recommendation platform and method based on big data speech recognition technology, comprising the following steps: S1: inputting basic information of the user; S2: collecting the user's speech signal and preprocessing the speech signal; S3: extracting speech features from the preprocessed speech signal to obtain a speech feature vector; S4: performing dialect recognition on the speech feature vector to extract the final speech feature vector; S5: based on the final speech feature vector, determining the skin care product label required by the user, and then matching it with the skin care product data on the market to obtain the skin care product data that needs to be trained; S6: constructing a regression model for predicting the skin care products suitable for the user, inputting the skin care product data that needs to be trained and training, and obtaining the recommended skin care product results. This application solves the problem of low recommendation accuracy of traditional skin care product recommendations by inventing a platform for quickly recommending skin care products suitable for users using speech recognition technology.

Description

Technical Field

[0001] The present invention relates to an intelligent skin care product recommendation platform and method based on big data voice recognition technology, and in particular to a platform and method for quickly recommending skin care products suitable for users based on the user's skin condition and needs through voice recognition technology. Background Art

[0002] With the improvement of living standards, people's demand for beauty and skin care is getting higher and higher, and the types of skin care products on the market are becoming more and more subdivided, which can produce different effects for different skin types. Faced with a wide variety of skin care products, it is difficult for users to make the right choice, which easily leads to dissatisfaction with the product effect after purchasing the product. Moreover, most traditional skin care product recommendation methods use manual recommendations, requiring customers to come to the counter for consultation or fill out questionnaires to understand the user's skin condition, which is time-consuming and labor-intensive and has low accuracy. With the development of mobile Internet, live streaming has become an important channel for e-commerce platforms to sell cosmetics. In addition, the interaction between skin care product recommendation platforms and users is becoming more and more intelligent, and voice interaction is gradually becoming the mainstream interaction method. Especially for middle-aged and elderly users, they are usually not good at typing, and the efficiency of handwritten text input is low, so they prefer to use voice interaction. However, many middle-aged and elderly users speak with a strong accent, and some even do not speak Mandarin at all.

[0003] The skin care product recommendation platforms and methods currently available on the market based on user skin type and needs are mainly based on collaborative filtering algorithms and content recommendation algorithms. When the platform's electronic customer service recommends products based on user data, the impact of the user's accent is not taken into account. When the user has a large accent problem, the recognition accuracy will be low, resulting in the mismatch between the skin care products recommended by the platform and the user. Therefore, the research and development of an intelligent skin care product recommendation platform based on big data speech recognition technology is of practical significance. Summary of the invention

[0004] In view of the above-mentioned defects of the prior art, the present invention proposes an intelligent skin care product recommendation platform and method based on big data speech recognition technology. The technical solution steps designed by the present invention include:

[0005] S1: Enter the user's basic information;

[0006] S2: Collecting a user's voice signal and preprocessing the voice signal;

[0007] S3: extracting speech features from the preprocessed speech signal to obtain a speech feature vector;

[0008] S4: performing dialect recognition on the speech feature vector to extract a final speech feature vector;

[0009] S5: Based on the final speech feature vector, determine the skin care product label required by the user, and then match it with the skin care product data on the market to obtain the skin care product data that needs to be trained;

[0010] S6: Construct a regression model for predicting skin care products suitable for the user, input the skin care product data that needs to be trained and perform training to obtain recommended skin care product results.

[0011] Preferably, S1 includes: when a user registers on the platform, the user enters basic information, and the basic information includes but is not limited to dialect, accent, skin quality, age, gender, and work environment information.

[0012] Preferably, the S2 includes: the preprocessing includes pre-emphasis processing, overlap frame processing and endpoint detection.

[0013] Preferably, S3 includes: the speech feature extraction includes extracting MFCC feature parameters in the speech signal, first performing Fourier transform on the speech signal, then performing spectrum filtering, obtaining frequency filtering features, and then performing discrete cosine transform to obtain MFCC features.

[0014] Preferably, the spectrum filtering includes: using a triangular bandpass filter, the formula is:

[0015]

[0016] In the formula, m is the number of the filter, and then calculate H m The logarithmic energy M is:

[0017]

[0018] Preferably, S4 includes: constructing a corresponding dialect speech corpus according to the dialect and accent information in the user's basic information, and constructing an acoustic model based on the dialect speech corpus, then inputting the speech feature vector for training, outputting the pronunciation state sequence of the language unit, and then decoding the pronunciation state sequence of the language unit to output the final speech feature vector.

[0019] Preferably, the construction of the dialect speech corpus includes: based on the dialect and accent information in the basic information of the user, based on four dialect databases including the Chinese dialect computer dictionary DOC, the Sino-Tibetan cognate database retrieval system, the Chinese dialect tone database, and the dialect speech and vocabulary database, using the Chinese dialect computer processing system to query the database, and screen out the dialect speech and vocabulary corresponding to the user's dialect and accent, and then integrate them to obtain the user's dialect speech corpus.

[0020] Preferably, the acoustic model includes: using a Gaussian mixture model-hidden Markov model, first initializing the dialect speech corpus to obtain a single-factor acoustic model, then performing triphone modeling, and finally training the single-factor acoustic model to output the result, and the formula is expressed as follows:

[0021] P(X|W)=∑P(x|w t )P(w t |w t-1 )

[0022] In the formula, P(x|w t ) is the Gaussian mixture model probability, P(w t |w t-1 ) is the hidden Markov model probability.

[0023] Preferably, S8 comprises: converting the skin care product data to be trained into a machine language formula to obtain input data for training:

[0024] D=(x 1 ,y 1 ), (x 2 ,y 2 )…(x N ,y N )

[0025] In the formula, x i is an m-dimensional vector, y takes values ​​in {0, 1};

[0026] Then fit a regression model to predict the skin care products suitable for the user:

[0027]

[0028] In the formula, x is the skin care product data to be trained, y is the final prediction result, ω is the weight, and ω takes values ​​between [0, 1]. When the value of ω is closer to 0, it means that the data is less suitable, and when it is closer to 1, it means that the data is more suitable;

[0029] Then the skin care product data x that needs to be trained is input into the equation. If the final result is greater than the value of the prediction straight line, the skin care product will be recommended. If it is less than the prediction value, the skin care product will not be recommended.

[0030] Preferably, the training process includes: during the training process, relevant data is exported, a data file is generated, and it is transmitted to a file server through a secure file transfer protocol. After the relevant data file is obtained, it is loaded into a self-built big database, and corresponding model training is performed on the data, and corresponding report data is generated, and finally returned to the Mysql database for data display to the user; the Mysql database is used to store the data in the model, including the user's basic information and the skin care product data that needs to be trained, and Elastic-Search is used to assist in storing information; the self-built big database includes four dialect databases, a dialect speech corpus and skin care product data on the market.

[0031] Beneficial effects:

[0032] 1. Compared with the traditional skin care product recommendation platform and method based on speech recognition, the present invention introduces Gaussian mixture model-hidden Markov model in speech recognition and focuses on identifying user dialect information, thereby reducing the problem of inaccurate skin care product recommendations due to dialects;

[0033] 2. Use big data analysis technology to build a regression model to predict suitable skin care products for users, which can quickly and accurately recommend suitable skin care products to users;

[0034] 3. The training process uniformly uses self-built big database and MySQL database to store information, information interaction and model training, ensuring data integrity, saving training steps and training space, thereby improving training efficiency and ensuring the accuracy of training results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a schematic diagram of the steps of a preferred embodiment of the intelligent skin care product recommendation platform and method based on big data speech recognition technology of the present invention;

[0036] Figure 2 It is a schematic diagram of the training process of a preferred embodiment of the intelligent skin care product recommendation platform and method based on big data speech recognition technology of the present invention. DETAILED DESCRIPTION

[0037] The embodiments of the present invention are described in detail below. The following embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation methods and specific operation processes are given, but the protection scope of the present invention is not limited to the following embodiments.

[0038] The present invention designs an intelligent skin care product recommendation platform and method based on big data speech recognition technology. The technical solution includes the following steps: Figure 1 As shown, specifically including:

[0039] S1: Enter the user's basic information;

[0040] S2: Collecting the user's voice signal and preprocessing the voice data;

[0041] S3: extracting speech features from the preprocessed speech signal to obtain a speech feature vector;

[0042] S4: performing dialect recognition on the speech feature vector to extract a final speech feature vector;

[0043] S5: Based on the final speech feature vector, determine the skin care product label required by the user, and then match it with the skin care product data on the market to obtain the skin care product data that needs to be trained; (Extract keyword labels from the final speech feature vector, such as "dry, oily, moisturizing, whitening, freckle removal, etc.", match the final keyword labels with the skin care product data on the market, and obtain a large amount of skin care product data that meets customer needs.)

[0044] S6: Construct a regression model for predicting skin care products suitable for the user, input the skin care product data that needs to be trained and perform training to obtain recommended skin care product results.

[0045] Specifically, in one embodiment, when the platform interacts with the user through voice, the platform will send voice questions about skin care product recommendations to the user, such as asking the user what type of skin care product (mask, cream, etc.) the user needs to buy, the type of brand and price range they prefer, etc.; the platform obtains the user's specific voice information requirements through preset questions, and then analyzes the user's response voice, and finally recommends the most suitable skin care products to the user based on the platform's recommendation algorithm and method.

[0046] Preferably, S1 includes: when the user registers on the platform, the user enters basic information and demand for skin care products, and the basic information includes but is not limited to dialect, accent, skin quality, age, gender and work environment information.

[0047] Preferably, S2 includes: pre-processing including pre-emphasis processing, overlap frame processing and endpoint detection.

[0048] Specifically, pre-emphasis processing: convert the voice signal into digital information, and then use a digital filter to improve the spectrum. The formula is as follows: H(z) = 1-az -1 ; In the formula, a is the pre-emphasis coefficient, which ranges from 0.9 to 1.0. It increases the low energy part of the signal. The enhanced voice signal can be evenly distributed in each frequency band. Because the ability of the collected voice signal in the high frequency band is generally weaker than that in the low frequency band, it is necessary to strengthen the energy of the high frequency band. To strengthen the ability of the high frequency band, a digital filter is used;

[0049] Overlapping frame processing: Introduce the window function to intercept the speech signal. The formula is as follows:

[0050] In order to make the frames coherent and smooth, overlapping frame processing is selected. When the speech signal is framed, the frequency band distributed around the zero point will become wider, causing some energy loss. Therefore, the window function is introduced to intercept the speech signal, but spectrum leakage will occur in the process, so the Hamming window is added to process the speech signal.

[0051] Endpoint detection: remove noise outside the speech signal, leaving valid data, improve the quality of speech recognition, reduce the waste of manpower and material resources, and use threshold judgment for endpoint detection.

[0052] Preferably, S3 includes: speech feature extraction includes extracting MFCC feature parameters in the speech signal, first performing Fourier transform on the speech signal, then performing spectrum filtering, obtaining frequency filtering features, and then performing discrete cosine transform to obtain MFCC features.

[0053] Specifically, MFCC feature parameters are widely used and are most consistent with human auditory perception. Therefore, the MFCC feature parameters in the speech signal are extracted (after obtaining the feature parameter vector, it is used for modeling training and finally predicting the word sequence of unknown speech). First, the speech signal is Fourier transformed, and the expression is:

[0054]

[0055] Then the spectrum is filtered using a set of triangular bandpass filters with the function:

[0056]

[0057] In the formula, m is the number of the filter, and then calculate H m The logarithmic energy M is as follows:

[0058]

[0059] Get the Mel frequency filter feature, perform discrete cosine transform on it, and get the MFCC feature. The formula is as follows:

[0060]

[0061] Preferably, S4 includes: constructing a corresponding dialect speech corpus according to the dialect and accent information in the user's basic information, and constructing an acoustic model based on the dialect speech corpus, then inputting speech features for training, outputting a pronunciation state sequence of a language unit, and then decoding the pronunciation state sequence of the language unit to output a final speech feature.

[0062] Preferably, constructing a dialect speech corpus includes: based on the dialect and accent information in the user's basic information, based on four dialect databases including a Chinese dialect computer dictionary DOC, a Sino-Tibetan cognate database retrieval system, a Chinese dialect tone database, and a dialect speech and vocabulary database, using a Chinese dialect computer processing system to query the database, and screen out the dialect speech and vocabulary corresponding to the user's dialect and accent, and then integrate them to obtain the user's dialect speech corpus.

[0063] Preferably, the dialects in Chinese are generally divided into ten major dialects, including Wu dialect, Min dialect, Yue dialect, Mandarin dialect, etc. However, even under the same dialect system, there are certain accent differences between dialects in different cities, districts, counties and even villages. The method for establishing an acoustic model of the present invention adopts a Gaussian mixture model (GMM)-hidden Markov model (HMM). First, the dialect speech corpus is initialized to obtain a single-factor acoustic model, and then three-phone modeling is performed. Finally, the single-factor acoustic model is trained and the result is output. The formula is expressed as follows:

[0064] P(X|W)=∑P(x|w t )P(w t |w t-1 )

[0065] In the formula, P(x|w t ) is the Gaussian mixture model probability, P(w t |w t-1 ) is the hidden Markov model probability.

[0066] Preferably, S8 includes: converting the skin care product data to be trained into a machine language formula to obtain input data for training:

[0067] D=(x 1 ,y 1 ), (x 2 ,y 2 )…(x N ,y N )

[0068] In the formula, x i is an m-dimensional vector, y takes values ​​in {0, 1};

[0069] Then fit a regression model to predict the skin care products suitable for the user:

[0070]

[0071] In the formula, x is the skin care product data to be trained, y is the final prediction result, ω is the weight, and ω takes values ​​between [0, 1]. When the value of ω is closer to 0, it means that the data is less suitable, and when it is closer to 1, it means that the data is more suitable;

[0072] Then input the skin care product data x that needs to be trained into the equation. If the final result is greater than the value of the prediction line, the skin care product will be recommended. If it is less than the predicted value, the skin care product will not be recommended.

[0073] Preferably, if Figure 2 As shown in the figure, the training process includes: during the training process, the relevant data is exported to generate data files, which are transmitted to the file server through the Secure File Transfer Protocol (SFTP). After obtaining the relevant data files, they are loaded into the self-built big database, and the data are trained with corresponding models, and the corresponding report data are generated. Finally, they are returned to the MySQL database for data display to the user; the MySQL database is used to store the data in the model, including the basic information of the user and the skin care product data that needs to be trained, and Elastic-Search is used to assist in storing information; the self-built big database includes four dialect databases, dialect speech corpus and skin care product data on the market.

[0074] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. An intelligent skin care product recommendation platform and method based on big data speech recognition technology, characterized in that: The following steps are involved: S1: Enter the user's basic information; S2: Collecting a user's voice signal and preprocessing the voice signal; S3: extracting speech features from the preprocessed speech signal to obtain a speech feature vector; S4: performing dialect recognition on the speech feature vector to extract a final speech feature vector; S5: Based on the final speech feature vector, determine the skin care product label required by the user, and then match it with the skin care product data on the market to obtain the skin care product data that needs to be trained; S6: Construct a regression model for predicting skin care products suitable for the user, input the skin care product data that needs to be trained and perform training to obtain recommended skin care product results.

2. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 1, characterized in that: The S1 includes: When a user registers on the platform, he or she enters basic information, including but not limited to dialect, accent, skin type, age, gender, and work environment information.

3. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 1, characterized in that: The S2 includes: The preprocessing includes pre-emphasis processing, overlap frame processing and endpoint detection.

4. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 1, characterized in that: The S3 includes: The speech feature extraction includes extracting MFCC feature parameters in the speech signal. First, the speech signal is subjected to Fourier transform, and then spectral filtering is performed to obtain frequency filtering features, and then discrete cosine transform is performed to obtain MFCC features.

5. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 4, characterized in that: The spectrum filtering includes: Using a triangular bandpass filter, the formula is: In the formula, m is the number of the filter, and then calculate H m The logarithmic energy M is:

6. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 1, characterized in that: The S4 includes: According to the dialect and accent information in the user's basic information, a corresponding dialect speech corpus is constructed, and an acoustic model is constructed based on the dialect speech corpus. Then, the speech feature vector is input for training, and a pronunciation state sequence of the language unit is output. Then, the pronunciation state sequence of the language unit is decoded and a final speech feature vector is output.

7. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 6, characterized in that: The construction of the dialect speech corpus comprises: According to the dialect and accent information in the user's basic information, based on four dialect databases including the Chinese dialect computer dictionary DOC, the Sino-Tibetan cognate database retrieval system, the Chinese dialect tone database, and the dialect speech and vocabulary database, the Chinese dialect computer processing system is used to query the database, and the dialect speech and vocabulary corresponding to the user's dialect and accent are screened out, and then integrated to obtain the user's dialect speech corpus.

8. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 6, characterized in that: The acoustic model includes: Using the Gaussian mixture model-hidden Markov model, the dialect speech corpus is first initialized to obtain a single-factor acoustic model, then three-phone modeling is performed, and finally the single-factor acoustic model is trained to output the result. The formula is expressed as follows: P(X|W)=∑P(x|w t )P(w t |w t-1 ) In the formula, P(x|w t ) is the Gaussian mixture model probability, P(w t |w t-1 ) is the hidden Markov model probability.

9. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to claim 1, characterized in that: The S6 includes: The skin care product data to be trained is converted into a machine language formula to obtain the training input data: D=(x 1 ,and 1 ),(x 2 ,and 2 )…(x N ,and N ) In the formula, x i is an m-dimensional vector, y takes values ​​in {0, 1}; Then fit a regression model to predict the skin care products suitable for the user: In the formula, x is the skin care product data to be trained, y is the final prediction result, ω is the weight, and ω takes values ​​between [0, 1]. When the value of ω is closer to 0, it means that the data is less suitable, and when it is closer to 1, it means that the data is more suitable; Then the skin care product data x that needs to be trained is input into the equation. If the final result is greater than the value of the prediction straight line, the skin care product will be recommended. If it is less than the prediction value, the skin care product will not be recommended.

10. The intelligent skin care product recommendation platform and method based on big data speech recognition technology according to any one of claims 1 to 9, characterized in that: The training process includes: During the training process, the relevant data is exported to generate data files, which are then transmitted to the file server through a secure file transfer protocol. After the relevant data files are obtained, they are loaded into a self-built big database, and corresponding model training is performed on the data, and corresponding report data is generated. Finally, the data is returned to the Mysql database for data display to the user; the Mysql database is used to store the data in the model, including the user's basic information and the skin care product data that needs to be trained, and Elastic-Search is used to assist in storing information; the self-built big database includes four dialect databases, a dialect speech corpus, and skin care product data on the market.

Images