Intelligent voice recognition and response system based on artificial intelligence

Abstract

The invention discloses an intelligent voice recognition and response system based on artificial intelligence, and belongs to the technical field of voice recognition and response. Comprising a voice recognition unit, a target question standardization module, a question and response feedback module and a response obtaining module, and the voice recognition unit obtains voice data and converts the voice data into text data; the voice recognition unit is connected to the target problem standardization module; the target question standardization module comprises a target question obtaining unit, the target question obtaining unit is connected to a word library conversion unit, and the word library conversion unit is connected with a synonym and standard word association library; the word library conversion unit outputs a calibration target problem; the target question obtaining unit obtains text data, and the text data is converted through the lexicon conversion unit. The intelligent voice recognition and response system based on artificial intelligence can intelligently recognize voice, and can quickly obtain accurate response voice according to the recognized text.

Description

Technical Field

[0001] The present invention specifically relates to an intelligent speech recognition and response system based on artificial intelligence, and belongs to the technical field of speech recognition and response. Background Art

[0002] Intelligent speech recognition is a technology that uses computers to automatically convert speech signals into text. There are many methods for speech recognition, such as China Patent Authorization Announcement No.: CN118230722B, which discloses an AI-based intelligent speech recognition method and system, which uses deep learning network models such as convolutional neural networks (CNN) and graph convolutional neural networks (GCN) to process and analyze speech signals to be recognized by using their powerful feature extraction capabilities and structured data processing capabilities, extracting local detailed waveform features and global contextual semantic association features of speech signals, and introducing an adaptive attention mechanism to integrate contextual information and visualize the distribution of important features, thereby ensuring the accuracy of semantic recognition; another example is China Patent Publication No.: CN118 136022A, discloses an intelligent speech recognition system and method, firstly, collects the signal of the speech to be recognized to obtain a speech analog signal, pre-processes the collected speech analog signal, and then uses the energy iteration cycle method to perform endpoint detection to obtain the speech digital signal to be recognized, and then extracts relevant feature information from the digital signal to be recognized; the speech signal to be recognized and its feature vector are passed through a network formed by an acoustic model component, a vocalization dictionary component, and a language model component to obtain all the matching recognition texts and the probability relationship between individual characters or words, and finally uses a decoder module to perform Viterbi decoding on the obtained data to obtain the final speech recognition text output; however, in the prior art, the speech recognition efficiency is general, and the response speech cannot be obtained quickly and accurately. Summary of the invention

[0003] To solve the above problems, the present invention proposes an intelligent speech recognition and response system based on artificial intelligence, which can intelligently recognize speech and quickly obtain accurate response speech according to the recognized text.

[0004] The intelligent speech recognition and response system based on artificial intelligence of the present invention comprises: A speech recognition unit, wherein the speech recognition unit acquires speech data and converts the speech data into text data; A target question standardization module, wherein the speech recognition unit is connected to the target question standardization module; the target question standardization module includes a target question acquisition unit, wherein the target question acquisition unit is connected to a word library conversion unit, wherein the word library conversion unit is respectively connected to a synonym and a standard word association library; the word library conversion unit outputs a calibrated target question; the target question acquisition unit acquires text data, and then converts the text data through the word library conversion unit to obtain a calibrated target question; the word library conversion unit calibrates each phrase of the text data, and during calibration, the phrase is calibrated through a synonym and a standard word association library, and when there are synonyms in the phrase, the synonyms are converted into standard words, thereby generating a calibrated target question; An answer acquisition module, the answer acquisition module includes a target factor acquisition module connected to the vocabulary conversion unit, the target factor acquisition module is connected to the index question generation module, the index question generation module is connected to the scoring unit, the index question library and the answer generation unit, the scoring unit is connected to the scoring library; the answer generation unit is connected to the index question and the answer database; the index question generation module outputs the index question, and the answer generation unit outputs the answer data; The target factor acquisition module acquires the calibration target question and sends the calibration target question to the index question generation module. The index question generation module compares the calibration target question with the index question database and performs correlation statistics through the scoring unit to obtain one or more index questions with the greatest correlation. At the same time, the answer generation unit uses the index question as an index to retrieve the index question and the answer database to obtain answer data. The question and answer feedback module, the output end of the answer acquisition module is connected to the question and answer feedback module, and the question and answer feedback module is connected to the voice library; the answer acquisition module indexes the question and answer data, and uses the indexed question and answer data as an index to retrieve and obtain voice data from the voice library; and outputs the voice data.

[0005] Furthermore, the voice data includes index question voice and answer data voice. The index question voice is sent to the interactive end of the answer interaction system. When the interactive end confirms, the answer data voice corresponding to the index question is sent to the interactive end; when the interactive end denies, the index question with the second largest correlation is pushed to the interactive end of the answer interaction system again, and confirmed again through the answer interaction system; until the interactive end confirms, the answer data voice corresponding to the index question is sent to the interactive end; when the statistical index questions are denied, the manual system is connected; when in use, when the index question is obtained, the answer data voice is automatically generated and sent directly to the interactive end; or after the index question interaction is adopted, the answer data voice is sent to the interactive end.

[0006] Furthermore, the target factor acquisition module acquires the calibration target question and extracts the target phrase unit from the calibration target question. ,get ; Index question library ; When the index question generation module works, the target phrase unit is obtained and index question bank , successively through right For comparison, when Existence ;Will Send to temporary storage ;in, Target phrase unit The last phrase of Target phrase unit The total number of phrases; for Any word group, ; Index Question Bank The last index question is, Index Question Bank The total number of index issues; Next, calculate the reservoir The various index issues in The intersection of = ∩ Perform intersection calculation, where For temporary storage Middle index problem, then the scoring unit obtains the intersection , and for the intersection Eliminate the error index problem series, specifically: scoring unit statistical intersection The number of each series in , then calculate the proportion of the series , ; ; Set this group of sequences as error index problem sequences; After eliminating error index problem sequences, obtain the counting sequence library Then Each set of sequences in is scored, and the scoring process is: ;in, Counting column library The last phrase of a sequence in For the scoring library Corresponding weight factors; complete the counting series library in sequence Calculate each set of series to get the counting series library Then, sort the data set according to the value, select one or more data from large to small, and reversely search the selected data to index the question library ; Get the index question result library; Then, output the index question from the index question result library; When two or more data overlap in the selected data, perform secondary selection of index questions, where data overlap means that the two data are the same, or the difference between the two data is less than the set threshold; When secondary selection of index questions, first count the number of phrases of the index question corresponding to the index question result library , then calculate the result ratio , ; and for two or more Compare the values ​​and select the one with the larger value. The index problem corresponding to the value is used as the index problem of the final output.

[0007] Furthermore, when the index problem is selected for the second time, according to The values ​​are arranged from large to small to get the output priority. The index question with the highest value ranking is the highest priority and is sent to the answer generation unit first. When extracting the first index question, the weight score is used for selection. When extracting the second index question, the intersection of the two is counted. The number of a set of sequences in The proportion of the number of phrases in the corresponding index question is output according to the proportion.

[0008] Furthermore, when extracting phrases from the calibration target question, the business dictionary is connected to extract phrases associated with the business from the calibration target question to obtain ; Extract phrases from the calibration target problem through the business dictionary library, obtain the business-related phrases in the calibration target problem, and establish a phrase set S.

[0009] Furthermore, the speech recognition unit includes a speech acquisition unit, a speech signal preprocessing unit, a feature extraction unit, a speech recognition unit and a text output unit connected in sequence; the speech acquisition unit acquires speech analog data and sends it to the speech signal preprocessing unit for preprocessing, and the preprocessing process is as follows: first, the speech analog data is converted into a digital signal group, and then the digital signal group is pre-emphasized, and the high-frequency part of the digital signal group is enhanced. The pre-emphasis is performed using a first-order high-pass filter, as follows:

[0010] is the digital signal after pre-emphasis, The first A digital signal, is the pre-emphasis coefficient, ranging from 0.95 to 0.99; Next, the pre-emphasized digital signal is subjected to frame division and windowing calculation, and the digital signal group is evenly divided into multiple frames, and each frame has a value range of 5 to 15 ms; an overlapping portion is set between adjacent frames; during the frame division and windowing calculation, a window function of a Hamming window or a Hanning window is used to obtain a windowed digital signal group; Next, the digital signal group output by the speech signal preprocessing unit is subjected to feature extraction by a feature extraction unit; the feature extraction is as follows: when extracting features, Mel cepstral coefficients are used for extraction, specifically: firstly, the data output by the speech signal preprocessing unit is subjected to fast Fourier transform, which is converted to the frequency domain in order to observe its spectrum energy distribution; the energy spectrum corresponding to the obtained spectrum is obtained; then, the obtained energy spectrum is processed by a mel triangle filter, and the output data is subjected to logarithmic operation. After the operation is completed, the Mel cepstral coefficients are obtained by discrete cosine transform, that is, a feature vector set G corresponding to the digital signal group is obtained, ; For the Frame feature vector, is the total amount of feature vector sets; then, the feature vector set G is sent to the speech recognition unit; speech recognition is performed through the end-to-end speech recognition model in the speech recognition unit, and the end-to-end speech recognition model is a CTC / AED joint model or a Conformer-CTC model; the speech recognition unit outputs text data.

[0011] Compared with the prior art, the intelligent speech recognition and response system based on artificial intelligence of the present invention, after acquiring the speech data, converts the speech data into text data, obtains the accurate response text by calibrating and recognizing the text data, and uses the response text as an index to obtain the speech data and response speech of the indexed question; it can accurately acquire the accurate interaction data sent by the interaction end, and at the same time, can return the response speech data matching it. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the overall structure of the intelligent speech recognition and response system of the present invention.

[0013] Figure 2 The figure is a schematic diagram of the working process of the intelligent speech recognition and response system of the present invention.

[0014] Figure 3 The figure is a schematic diagram of the interactive workflow of the voice data and response interactive system of the present invention. DETAILED DESCRIPTION

[0015] like Figures 1 to 3The intelligent speech recognition and response system based on artificial intelligence shown includes: A speech recognition unit, wherein the speech recognition unit acquires speech data and converts the speech data into text data; A target question standardization module, wherein the speech recognition unit is connected to the target question standardization module; the target question standardization module includes a target question acquisition unit, wherein the target question acquisition unit is connected to a word library conversion unit, wherein the word library conversion unit is respectively connected to a synonym and a standard word association library; the word library conversion unit outputs a calibrated target question; the target question acquisition unit acquires text data, and then converts the text data through the word library conversion unit to obtain a calibrated target question; the word library conversion unit calibrates each phrase of the text data, and during calibration, the phrase is calibrated through a synonym and a standard word association library, and when there are synonyms in the phrase, the synonyms are converted into standard words, thereby generating a calibrated target question; An answer acquisition module, the answer acquisition module includes a target factor acquisition module connected to the vocabulary conversion unit, the target factor acquisition module is connected to the index question generation module, the index question generation module is connected to the scoring unit, the index question library and the answer generation unit, the scoring unit is connected to the scoring library; the answer generation unit is connected to the index question and the answer database; the index question generation module outputs the index question, and the answer generation unit outputs the answer data; The target factor acquisition module acquires the calibration target question and sends the calibration target question to the index question generation module. The index question generation module compares the calibration target question with the index question database and performs correlation statistics through the scoring unit to obtain one or more index questions with the greatest correlation. At the same time, the answer generation unit uses the index question as an index to retrieve the index question and the answer database to obtain answer data. The question and answer feedback module, the output end of the answer acquisition module is connected to the question and answer feedback module, and the question and answer feedback module is connected to the voice library; the answer acquisition module indexes the question and answer data, and uses the indexed question and answer data as an index to retrieve and obtain voice data from the voice library; and outputs the voice data.

[0016] The voice data includes index question voice and answer data voice. The index question voice is sent to the interactive end of the answer interaction system. When the interactive end confirms, the answer data voice corresponding to the index question is sent to the interactive end; when the interactive end denies, the index question with the second largest correlation is pushed to the interactive end of the answer interaction system again, and confirmed again through the answer interaction system; until the interactive end confirms, the answer data voice corresponding to the index question is sent to the interactive end; when the statistical index questions are denied, the manual system is connected; when in use, when the index question is obtained, the answer data voice is automatically generated and sent directly to the interactive end; or after the index question interaction is adopted, the answer data voice is sent to the interactive end.

[0017] The target factor acquisition module acquires the calibration target question and extracts the target phrase unit from the calibration target question. ,get ; Index question library ; When the index question generation module works, the target phrase unit is obtained and index question bank , successively through right For comparison, when Existence ;Will Send to temporary storage ;in, Target phrase unit The last phrase of Target phrase unit The total number of phrases; for Any word group, ; Index Question Bank The last index question is, Index Question Bank The total number of index issues; Next, calculate the reservoir The various index issues in The intersection of = ∩ Perform intersection calculation, where For temporary storage Middle index problem, then the scoring unit obtains the intersection , and for the intersection Eliminate the error index problem series, specifically: scoring unit statistical intersection The number of each series in , then calculate the proportion of the series , ; ; Set this group of sequences as error index problem sequences; After eliminating error index problem sequences, obtain the counting sequence library Then Each set of sequences in is scored, and the scoring process is: ;in, Counting column library The last phrase of a sequence in For the scoring library Corresponding weight factors; complete the counting series library in sequence Calculate each set of series to get the counting series library Then, sort the data set according to the value, select one or more data from large to small, and reversely search the selected data to index the question library ; Get the index question result library; Then, output the index question from the index question result library; When two or more data overlap in the selected data, perform secondary selection of index questions, where data overlap means that the two data are the same, or the difference between the two data is less than the set threshold; When secondary selection of index questions, first count the number of phrases of the index question corresponding to the index question result library , then calculate the result ratio , ; and for two or more Compare the values ​​and select the one with the larger value. The index problem corresponding to the value is used as the index problem of the final output.

[0018] When the index problem is selected again, according to The values ​​are arranged from large to small to get the output priority. The index question with the highest value ranking is the highest priority and is sent to the answer generation unit first. When extracting the first index question, the weight score is used for selection. When extracting the second index question, the intersection of the two is counted. The number of a set of sequences in The proportion of the number of phrases in the corresponding index question is output according to the proportion.

[0019] When extracting phrases from the calibration target problem, the business dictionary library is connected, and phrases associated with the business are extracted from the calibration target problem through the business dictionary library to obtain ; Extract phrases from the calibration target problem through the business dictionary library, obtain the business-related phrases in the calibration target problem, and establish a phrase set S.

[0020] The speech recognition unit includes a speech acquisition unit, a speech signal preprocessing unit, a feature extraction unit, a speech recognition unit and a text output unit connected in sequence; the speech acquisition unit acquires speech analog data and sends it to the speech signal preprocessing unit for preprocessing, and the preprocessing process is as follows: first, the speech analog data is converted into a digital signal group, and then the digital signal group is pre-emphasized, and the high-frequency part of the digital signal group is enhanced. The pre-emphasis is performed using a first-order high-pass filter, as follows:

[0021] is the digital signal after pre-emphasis, The first A digital signal, is the pre-emphasis coefficient, ranging from 0.95 to 0.99; Next, the pre-emphasized digital signal is subjected to frame division and windowing calculation, and the digital signal group is evenly divided into multiple frames, and each frame has a value range of 5 to 15 ms; an overlapping portion is set between adjacent frames; during the frame division and windowing calculation, a window function of a Hamming window or a Hanning window is used to obtain a windowed digital signal group; Next, the digital signal group output by the speech signal preprocessing unit is subjected to feature extraction by a feature extraction unit; the feature extraction is as follows: when extracting features, Mel cepstral coefficients are used for extraction, specifically: firstly, the data output by the speech signal preprocessing unit is subjected to fast Fourier transform, which is converted to the frequency domain in order to observe its spectrum energy distribution; the energy spectrum corresponding to the obtained spectrum is obtained; then, the obtained energy spectrum is processed by a mel triangle filter, and the output data is subjected to logarithmic operation. After the operation is completed, the Mel cepstral coefficients are obtained by discrete cosine transform, that is, a feature vector set G corresponding to the digital signal group is obtained, ; For the Frame feature vector, is the total number of feature vector sets; Next, the feature vector set G is sent to the speech recognition unit; speech recognition is performed through an end-to-end speech recognition model in the speech recognition unit, and the end-to-end speech recognition model is a CTC / AED joint model or a Conformer-CTC model; and the speech recognition unit outputs text data.

[0022] The above embodiments are only preferred implementations of the present invention, so any equivalent changes or modifications made according to the structures, features and principles described in the scope of application of the present invention are included in the scope of application of the present invention.

Claims

1. An intelligent speech recognition and response system based on artificial intelligence, characterized in that: include: A speech recognition unit, wherein the speech recognition unit acquires speech data and converts the speech data into text data; A target question standardization module, to which the speech recognition unit is connected; The target question standardization module includes a target question acquisition unit, the target question acquisition unit is connected to a word library conversion unit, the word library conversion unit is respectively connected to a synonym and a standard word association library; the word library conversion unit outputs a calibration target question; the target question acquisition unit acquires text data, and then converts the text data through the word library conversion unit to obtain a calibration target question; the word library conversion unit calibrates each phrase of the text data, and during calibration, the phrase is calibrated through a synonym and a standard word association library, and when there are synonyms in the phrase, the synonyms are converted into standard words, thereby generating a calibration target question; An answer acquisition module, the answer acquisition module includes a target factor acquisition module connected to the vocabulary conversion unit, the target factor acquisition module is connected to the index question generation module, the index question generation module is connected to the scoring unit, the index question library and the answer generation unit, the scoring unit is connected to the scoring library; the answer generation unit is connected to the index question and the answer database; the index question generation module outputs the index question, and the answer generation unit outputs the answer data; The target factor acquisition module acquires the calibration target question and sends the calibration target question to the index question generation module. The index question generation module compares the calibration target question with the index question database and performs correlation statistics through the scoring unit to obtain one or more index questions with the greatest correlation. At the same time, the answer generation unit uses the index question as an index to retrieve the index question and the answer database to obtain answer data. A question and answer feedback module, wherein the output end of the answer acquisition module is connected to the question and answer feedback module, and the question and answer feedback module is connected to the speech library; The answer acquisition module indexes the question and answer data, and uses the indexed question and answer data as an index to retrieve and acquire voice data from the voice library; and outputs the voice data.

2. The intelligent speech recognition and response system based on artificial intelligence according to claim 1, characterized in that: The voice data includes the index question voice and the answer data voice, and the index question voice is sent to the interactive end of the answer interaction system. When the interactive end confirms, the answer data voice corresponding to the index question is sent to the interactive end; when the interactive end denies, the index question with the second largest correlation is pushed to the interactive end of the answer interaction system again, and confirmed again through the answer interaction system; After getting confirmation from the interactive end, the answer data voice corresponding to the index question is sent to the interactive end; when all the index questions counted are denied, the manual system is connected.

3. The intelligent speech recognition and response system based on artificial intelligence according to claim 1, characterized in that: The target factor acquisition module acquires the calibration target question and extracts the target phrase unit from the calibration target question. ,get ; Index question library ; When the index question generation module works, the target phrase unit is obtained and index question bank , successively through right For comparison, when Existence ;Will Send to temporary storage ;in, Target phrase unit The last phrase of Target phrase unit The total number of phrases; for Any word group, ; Index Question Bank The last index question is, Index Question Bank The total number of index issues; Next, calculate the reservoir The various index issues in The intersection of = ∩ Perform intersection calculation, where For temporary storage Middle index problem, then the scoring unit obtains the intersection , and for the intersection Eliminate the error index problem series, specifically: scoring unit statistical intersection The number of each series in , then calculate the proportion of the series , ; ; Set this group of sequences as error index problem sequences; After eliminating error index problem sequences, obtain the counting sequence library Then Each set of sequences in is scored, and the scoring process is: ;in, Counting column library The last phrase of a sequence in For the scoring library Corresponding weight factors; complete the counting series library in sequence Calculate each set of series to get the counting series library Then, sort the data set according to the value, select one or more data from large to small, and reversely search the selected data to index the question library ; Get the index question result library; Then, output the index question from the index question result library; When two or more data overlap in the selected data, perform secondary selection of index questions, where data overlap means that the two data are the same, or the difference between the two data is less than the set threshold; When secondary selection of index questions, first count the number of phrases of the index question corresponding to the index question result library , then calculate the result ratio , ; and for two or more Compare the values ​​and select the one with the larger value. The index problem corresponding to the value is used as the index problem of the final output.

4. The intelligent speech recognition and response system based on artificial intelligence according to claim 3, characterized in that: When the index problem is selected again, according to The values ​​are arranged from large to small to get the output priority. The indexed question with the highest ranking value is sent to the answer generation unit first as the highest priority.

5. The intelligent speech recognition and response system based on artificial intelligence according to claim 3 is characterized in that: When extracting phrases from the calibration target problem, the business dictionary library is connected, and phrases associated with the business are extracted from the calibration target problem through the business dictionary library to obtain .

6. The intelligent speech recognition and response system based on artificial intelligence according to claim 1, characterized in that: The speech recognition unit comprises a speech acquisition unit, a speech signal preprocessing unit, a feature extraction unit, a speech recognition unit and a text output unit which are connected in sequence.

Images