A method for improving speech recognition rate in professional fields

Through the method of level-first matching and frequency-first matching, combined with deep learning and speech noise reduction technology, the problem of high speech recognition error rate in professional fields is solved, automatic recognition and update of special vocabulary is realized, and recognition accuracy and system intelligence are improved. change.

CN114822545BActive Publication Date: 2024-11-08SOUTH CHINA UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210298454.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-25
Publication Date
2024-11-08
Estimated Expiration
2042-03-25

AI Technical Summary

Technical Problem

In the professional field, the existing speech recognition technology has a high recognition error rate due to homophones and noise interference, and requires manual establishment and update of special vocabulary databases, which is time-consuming and labor-intensive and difficult to meet professional needs.

Method used

The method of level-first matching and frequency-first matching is adopted. Local special vocabulary and professional terms are queued according to frequency differences through the differential frequency special vocabulary library, and the pinyin string is preferred. Combined with deep learning and speech noise reduction technology, special vocabulary is automatically recognized and updated.

Benefits of technology

It improves the accuracy of voice recognition in professional fields, reduces manual maintenance costs, reduces error rates, and realizes an intelligent voice recognition system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114822545B_ABST
    Figure CN114822545B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for improving the speech recognition rate in a professional field, which is mainly used for recognizing speech in a professional field or a specific industry. The professional field usually involves a large number of professional terms and special terms in various application departments of the professional field that combine local characteristics, such as equipment names containing location names, work section names and even professional names, so the speech recognition error rate is relatively high. The present invention proposes a secondary difference frequency principle, and automatically establishes a difference frequency special vocabulary library, which includes a first-level difference frequency sub-library for storing local special vocabulary and a second-level difference frequency sub-library for storing professional terms. Pinyin and text are matched with the difference frequency special vocabulary as the center, and an arbitrary position conversion mechanism is adopted. Through these steps, the accuracy of speech recognition can be improved, especially the special vocabulary of local professional departments can be recognized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition, and in particular to a method for improving speech recognition rate in a professional field. Background Art

[0002] In voice conversations in professional fields or specific industries, such as railways, civil aviation, banks, electricity and other departments, there are usually a lot of professional terms and place names, lines, station names, equipment names with specific numbers and even names of people that are unique to each local department. Due to the existence of a large number of homophones in the language, current voice recognition technology often recognizes these relatively uncommon special words as other common ordinary words, resulting in a high error rate and difficulty in meeting the needs of the profession or industry. Therefore, the main reason for the above problem is that the current voice recognition technology is based on frequency priority matching. When the voice is converted into pinyin, it will give priority to matching those common words and popular words that appear more frequently.

[0003] Some literature proposes adding dedicated vocabulary, but there are still three problems:

[0004] 1. The special vocabulary of local departments will not appear in the general vocabulary database. Professional or industry departments in each region need to manually establish them according to the needs of their own departments. Local personnel are required to screen special vocabulary one by one from a large amount of local text materials. In addition, these special vocabulary must be constantly updated, replaced and the refresh frequency must be counted, which is time-consuming and laborious. The daily tasks of these department personnel are heavy and it is difficult for them to split themselves.

[0005] 2. Currently, the matching starts with the first letter of a pinyin string and converts it into text one by one. Due to the interference of noise and the existence of homophones, some pinyins of important special words in the back may be combined with the previous pinyins in advance and converted into other phrases, resulting in errors. For example, the railway department originally wanted to start the vehicle at XXX station, but it turned out to start the vehicle at other stations, causing serious consequences. On the contrary, errors in ordinary words are relatively acceptable.

[0006] 3. Add special words and common popular words together in the vocabulary. Special words appear less frequently, but popular words will still be given priority when their pinyin is the same. Summary of the invention

[0007] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and propose a method for improving the speech recognition rate in professional fields, which can accurately recognize professional or industry speech and can recognize special vocabulary in the profession or industry at a relatively low cost, especially special vocabulary of local departments, such as local place names, lines, stations, power stations, specific equipment names and even personal names.

[0008] To achieve the above object, the technical solution provided by the present invention is: a method for improving the speech recognition rate in a professional field, comprising:

[0009] Level priority matching: after converting speech to pinyin, a pinyin string consisting of letters and tones is obtained, which is named A. In the process of converting A into text, it is preferentially matched with the pinyin of the first-level difference frequency sub-library vocabulary of the difference frequency special word library. If the match is successful, part of the pinyin of A is converted into text. If the match is unsuccessful, the second-level difference frequency sub-library is considered. The matching is to find the similarity between a part of the pinyin of A and the correct pinyin of a Chinese phrase or character, which is also referred to as the matching of pinyin and text or vocabulary; the difference frequency special word library is composed of local special words and their pinyin to form a first-level difference frequency sub-library, and professional terms and their pinyin to form a second-level difference frequency sub-library. The sound is composed of a secondary difference frequency sub-library; the queuing order of a word in the difference frequency sub-library is determined by the difference of its frequency, that is, the higher the frequency of the word in the special data, the higher the queuing order of the word, and the higher the frequency of the word in the general data, the lower the queuing order of the word; the word refers to a Chinese phrase, a Chinese phrase includes at least 2 Chinese characters, the local special vocabulary refers to the vocabulary used only in the local machine, local area network, a specific region, group or department, local special vocabulary and professional terms are collectively referred to as special vocabulary, and the others are called general vocabulary, and the difference in frequency is referred to as difference frequency;

[0010] Frequency priority matching: After the level priority matching is completed, the remaining pinyin of A is matched with the pinyin of common vocabulary. Non-specialized vocabulary with high frequency appearing in common data is matched first. Finally, the remaining pinyin is matched with the pinyin of a single Chinese character.

[0011] Further, the step of obtaining a pinyin string composed of letters and tones includes:

[0012] S1. Authority identification: Perform voiceprint recognition on the wake-up voice input by the user. If the user has the authority to use, proceed to step S2, otherwise, exit;

[0013] S2, speech noise reduction: perform speech noise reduction on user speech to reduce the interference of noise on recognition;

[0014] S3, voice to pinyin conversion: perform voice analysis and recognition on the noise-reduced voice obtained in step S2, and obtain the entire sentence pinyin string corresponding to the voice.

[0015] Further, the permission identification includes the following steps:

[0016] S11, obtaining a wake-up voice;

[0017] S12, extracting the Fbank energy features of each frame in the speech obtained in step S11;

[0018] S13, input the Fbank energy features into a pre-trained DNN model and obtain a high-level feature expression based on deep learning;

[0019] S14. Perform cosine similarity matching between the high-level feature expression obtained in step S13 and the user voice feature bound to the user login account; if the cosine similarity is lower than or equal to the set value, it is determined that the user does not match the user bound to the login account and has no use permission; if it is higher than the set value, it is determined that the user has use permission.

[0020] Further, the construction of the difference frequency special word library includes the following steps:

[0021] S1. Collect text data related to local work, including work plans, work files, meeting records, equipment records and call texts on the local machine and the local area network;

[0022] S2. Clean and segment the text data collected in the previous step, and obtain a vocabulary list. Then, perform vocabulary frequency statistics on the vocabulary list, and obtain a local special word frequency dictionary, where word frequency = the number of times the word is repeated × the length of the word / the total number of words in all the data. The local special word frequency dictionary contains all the words and their word frequencies that appear in the local text data;

[0023] S3, searching the academic articles of the profession on the Internet, performing word frequency statistics on the obtained vocabulary list, and obtaining a professional term frequency dictionary, which includes all the words and their word frequencies that appear in the academic articles of the profession;

[0024] S4. Perform word frequency statistics on general data, including the People's Daily corpus and news from three major websites, Sina, Sohu, and NetEase, to obtain a general word frequency dictionary, which includes all the words and their frequencies that appear in the general data;

[0025] S5. Perform a frequency difference operation on each word in the professional term frequency dictionary, namely:

[0026] The difference frequency value = the frequency of a word's professional terminology - k 1 × its common word frequency, here k 1 is the set coefficient. The professional term frequency and general term frequency are obtained from the professional term frequency dictionary and general term frequency dictionary respectively. The frequency of words that cannot be found in the dictionary is counted as 0.

[0027] The words with primary difference frequency value less than or equal to 0 are discarded, and those with difference frequency value greater than 0 are stored in the secondary difference frequency sub-bank. The higher the difference frequency value, the higher the word is in the queue of the difference frequency sub-bank.

[0028] S6. Perform a secondary frequency difference operation on each word in the local special word frequency dictionary, namely:

[0029] Second difference frequency value = local specific frequency of a word - k 2 × Frequency of its professional terms - k 1 × its common word frequency, here k 2 is the set coefficient, and the local specific word frequency is obtained from the local specific word frequency dictionary;

[0030] Words with secondary difference frequency values ​​greater than 0 are stored in the primary difference frequency sub-bank, otherwise they are discarded. Words with higher difference frequency values ​​are queued at the front of the difference frequency sub-bank.

[0031] S7. Associating each word in the difference frequency special word library with its corresponding pinyin, and finally forming a difference frequency special word library for professional fields, wherein the difference frequency special word library is composed of the first and second level difference frequency sub-libraries.

[0032] Furthermore, the level priority matching step includes:

[0033] Reverse word selection: select the pinyin of the word with the highest difference frequency value from the words that have not been matched in the first-level difference frequency sub-library, and set its name as B. If the pinyin of the words in the first-level difference frequency sub-library have been matched, then move on to the second-level difference frequency sub-library;

[0034] Convert pinyin at any position: Search for a substring C similar to B in A. If B matches C successfully, convert C into the corresponding Chinese phrase. If there are multiple substrings similar to B in A, repeat the above operation. The substring C can be located at any position in A.

[0035] Further, the matching comprises the following steps:

[0036] Calculate the phoneme edit distance: refers to the minimum number of phoneme edit operations required to convert one pinyin string into another; the phoneme refers to the initial consonant or final of the pinyin, and the permitted edit operations include: inserting an initial consonant / final, deleting an initial consonant / final, replacing one initial consonant / final with another, and a replacement between ambiguous sounds is only counted as 0.5 times, and the above operations do not include tones;

[0037] Calculate the pitch distance: if |xy|≥2, add 1 to the pitch distance, otherwise do not add. Here, x and y are the tones after the corresponding finals in the two pinyin strings. The tones after all the finals are calculated and accumulated to get the pitch distance.

[0038] Calculate the matching value = phoneme edit distance + pitch distance;

[0039] If the match is a special word, then when the match value is less than a given threshold, the match value and the match success signal are output, otherwise a match failure signal is given;

[0040] If the match is a common term, the matching value is output.

[0041] Furthermore, the matching value is further modified according to the length of the vocabulary. Assuming that the pinyin of the vocabulary to be matched corresponds to j characters, according to the calculated matching value D, the modified matching value is

[0042] Furthermore, the threshold is calculated according to the following formula:

[0043] Threshold = k 3 × the average matching value of the current user, where k 3 The coefficient is set. The identity of the current user is obtained by authority recognition. During the user's use, each time a matched text sentence is output, the average matching value of all the words in the sentence is calculated, and then the total average value of all the sentences used in the user's history is accumulated and recorded under the user name.

[0044] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0045] 1. According to a method for improving speech recognition rate in professional fields provided by the present invention, the speech recognition system can automatically distinguish between common common words and special words in professional or industry fields, especially those special words of regional departments, so that each regional department does not need to manually build a database. The system will store these special words in a hierarchical difference frequency special word library and continuously refresh, update and replace them, thereby saving a lot of time and energy.

[0046] 2. The level priority matching of the present invention focuses on special vocabulary and gives priority to matching, which reduces the errors caused by the prior popular general vocabulary of the existing speech recognition method, thereby improving the accuracy of speech recognition in professions or industries, reducing errors in dispatch instruction recognition, reducing accidents, improving efficiency, and better realizing the intelligence of regional departments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method of the present invention.

[0048] Figure 2 It is a flow chart of the voiceprint recognition process.

[0049] Figure 3 It is a flow chart of speech noise reduction process. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0051] See also Figure 1 As shown, this embodiment discloses a method for improving the speech recognition rate in a professional field, and the specific implementation steps are as follows:

[0052] S1. Permission recognition: Perform voiceprint recognition on the wake-up voice input by the user. If the user has the usage permission, go to step S2; otherwise, exit.

[0053] S2. Voice noise reduction: Perform voice noise reduction on the user's voice to reduce the interference of noise on recognition.

[0054] S3. Voice to pinyin conversion: Use a well-known deep learning voice recognition algorithm to perform voice analysis and recognition on the noise-reduced voice obtained in S2, and obtain the whole sentence pinyin corresponding to the voice.

[0055] For example, when the command voice of the power dispatcher is:

[0056] "Insert the grounding switch No. 1234 of the Tie Nitrogen Tanghua Stone Line at Tielin Station", through the conversion in step S3, the following pinyin string A will be obtained: [tou2ru4tie3lin2zhan4tie3dan4tang2hua2shi2xian4yao1liang3san1si4jie1di4dao1zha2].

[0057] S4. Match pinyin and special vocabulary with priority based on the differential frequency special vocabulary library, and perform Chinese text hierarchical matching on A.

[0058] For example, Tielin Station, Tie Nitrogen Tanghua Stone Line, and grounding switch are all special vocabulary. The differential frequency value ranking is: Tielin Station (Level 1) > Tie Nitrogen Tanghua Stone Line (Level 1) > grounding switch (Level 2).

[0059] a. Reverse word extraction: Extract words one by one from the first-level differential frequency sub-library in descending order of the differential frequency value. For each word extracted, check whether there is a matching substring in A. The current matching method is to extract pinyin from A and search in the vocabulary library. The method of this patent is the opposite, so it is called reverse word extraction.

[0060] b. Arbitrary position conversion: Different from the current method that starts converting words from the first letter, the substring in this method can be converted at any position in A. If the matching gap is greater than the given threshold, give up and take the next word until the pinyin of Tielin Station "tie3lin2zhan4" can match a part of A, so A becomes [tou2ru4 Tielin Station tie3dan4tang2hua2shi2xian4yao1liang3san1si4jie1di4dao1zha2]. Here, reverse word extraction and arbitrary position conversion are specifically designed for differential frequency special vocabulary and are different from the currently well-known methods. Similarly, the remaining special vocabulary in A will be converted later: [tou2ru4 Tielin Station Tie Nitrogen Tanghua Stone Line yao1liang3san1si4 grounding switch].

[0061] S5, frequency priority matching pinyin and common vocabulary; when all the special words in string A are converted, the common vocabulary is matched according to the well-known frequency priority method: in order from front to back, take tou2 ru4, check the general dictionary, get "investment", string A becomes: [investment Tielin Station iron nitrogen Tang Huashi line yao1liang3san1si4 grounding knife switch];

[0062] S6, the remaining pinyin matches a single Chinese character to obtain the entire sentence text [put into Tielin Station iron nitrogen Tanghuashi line 1234 grounding knife switch];

[0063] S7, output the whole sentence;

[0064] S8, refresh the vocabulary frequency and difference frequency value, refresh the frequency of special and general vocabulary, and recalculate the difference frequency value, that is, refresh the frequency of the difference frequency vocabulary "Tielin Station", "Iron Nitrogen Tanghua Stone Line", "Earth Switch", recalculate the difference frequency value of the above vocabulary and update its order, refresh the frequency of the vocabulary "investment" in the general library, and update its order, and the vocabulary that has not appeared does not need to be frequently refreshed;

[0065] S9. End.

[0066] Figure 1 The details of the difference frequency vocabulary, matching algorithm, authority identification, and speech noise reduction involved in the process are described below.

[0067] In the above process, the level priority matching must be based on the difference frequency special word library. The difference frequency special word library is automatically constructed according to the following principles of this patent:

[0068] For professional fields, the importance of special words is higher than that of common words, so its recognition rate should be guaranteed first, so a special word library should be established. Furthermore, a noisy environment affects the similarity of matching, so when the matching similarity is not much different, high-level special words should be matched first. In addition, the system should automatically build a library through a program. To automatically distinguish special words from common words, their differences must be utilized. Special words, especially local special words, such as the special words mentioned above, such as "iron nitrogen Tanghua stone line", generally do not appear in ordinary news or articles, but appear in local power dispatch logs, work tickets, equipment records and local call text records. On the contrary, the general words such as "investment" mentioned above will appear frequently in ordinary articles or online articles. In addition, the word "grounding knife switch" may appear in both local texts and news reports. Therefore, this patent proposes that the level of a word is determined by the difference between its two frequencies, that is, the higher the frequency of appearance in special materials, the higher the level of the word, and the higher the frequency of appearance in general materials, the lower the level of the word.

[0069] Taking a local power dispatching department as an example, the automatic construction method of the difference frequency special word library is as follows:

[0070] S1. The voice recognition system collects text data such as power dispatch logs, work tickets, and call texts on the local area network and the local machine;

[0071] S2. Clean and segment the text data collected in the previous step, and obtain a vocabulary list, then count the word frequency of the vocabulary list to obtain a local special word frequency and a local special word frequency dictionary. The word frequency = the number of times the word is repeated × the length of the word / the total number of words in all the data. The local special word frequency dictionary contains all the words of the local data (local special words, professional terms and general words) and their word frequencies;

[0072] S3, the speech recognition system searches the professional academic articles on the Internet, counts the word frequency of the vocabulary list, and obtains the professional term frequency and professional term frequency dictionary, which contains all the words (professional terms, general words) and their word frequencies that appear in the professional academic articles;

[0073] S4. Perform word frequency statistics on general data (People's Daily corpus and news from three major websites, Sina, Sohu, and Netease) to obtain general word frequency and general word frequency dictionary, which includes all words (general words, some professional terms) and their word frequencies that appear in general data;

[0074] S5. Perform a frequency difference operation on each word in the professional term frequency dictionary, namely:

[0075] The difference frequency value = the frequency of a word's professional terminology - k 1 × its common word frequency, here k 1 The coefficient is set and can be obtained through statistics. The professional term frequency and general term frequency are obtained from the professional term frequency dictionary and the general term frequency dictionary respectively. The frequency of words that cannot be found in the dictionary is counted as 0;

[0076] The words with primary difference frequency value less than or equal to 0 are discarded, and those with difference frequency value greater than 0 are stored in the secondary difference frequency sub-bank. The higher the difference frequency value, the higher the word is in the queue of the difference frequency sub-bank.

[0077] S6. Perform a secondary frequency difference operation on each word in the local special word frequency dictionary, namely:

[0078] Second difference frequency value = local specific frequency of a word - k 2 × Frequency of its professional terms - k 1 × its common word frequency, here k 2 The coefficient for setting can be obtained through statistics, and the local special word frequency is obtained from the local special word frequency dictionary;

[0079] Words with secondary difference frequency values ​​greater than 0 are stored in the primary difference frequency sub-bank, otherwise they are discarded. Words with higher difference frequency values ​​are queued at the front of the difference frequency sub-bank.

[0080] S7, associate each word in the frequency difference special word library with its corresponding pinyin, and finally form a frequency difference special word library for power dispatching. The frequency difference special word library is composed of primary and secondary frequency difference sub-libraries.

[0081] The above process is still explained using the above-mentioned power sentence "1234 grounding switch of the iron-nitrogen-Tanghua-Shi line at Tielin Station is put into use" as an example: "put into use" is a general term and will appear in local dispatching logs, power academic articles and news reports. Its primary difference frequency value is less than 0 and will be stored in the general vocabulary library. "grounding switch" is an electric power professional term and will appear in local dispatching logs and electric power academic articles, but rarely appears in news reports. Its primary difference frequency value is greater than 0, but its secondary difference frequency value is less than 0 and will be stored in the secondary difference frequency sub-library. "iron-nitrogen-Tanghua-Shi line" will appear in local dispatching logs, but generally will not appear in electric power academic articles and news reports. Its primary and secondary difference frequency values ​​are both greater than 0 and will be stored in the primary difference frequency sub-library.

[0082] exist Figure 1 In the process, the matching of pinyin with words and characters can be carried out according to known methods. The present invention provides a preferred matching scheme comprising the following steps:

[0083] S1. Calculate the phoneme edit distance: refers to the minimum number of phoneme edit operations required to convert one phoneme into another between two pinyin strings. The phoneme refers to the initial consonant or final of pinyin. The permitted edit operations include: inserting an initial consonant / final, deleting an initial consonant / final, replacing one initial consonant / final with another. The replacement between ambiguous sounds is only counted as 0.5 times. The above operations do not include tones.

[0084] S2. Calculate the pitch distance: if |xy|≥2, the pitch distance is increased by 1, otherwise it is not increased. Here, x and y are the tones after the corresponding finals in the two pinyin strings. The tones after all the finals are calculated and accumulated to obtain the pitch distance.

[0085] Calculate the matching value = phoneme edit distance + pitch distance;

[0086] S3, according to the length of the vocabulary, the matching value is further modified. Assuming that the vocabulary to be matched contains the pinyin of j characters, according to the matching value D calculated in step S2, the modified matching value is Example: Assume that due to noise interference and the speaker's Mandarin is not standard, the pinyin string A is different from the pinyin string A. 0 :

[0087] [tou2ru4tie3ling1zhang1tie3dan4tang2hua2shi2xian4yao1liang3san1si4jie1di4dao1zha2]

[0088] Reverse word selection: select words one by one from the first-level difference frequency sub-library according to the difference frequency value from large to small, and each word selected will be in the pinyin string A 0 Search for a matching substring in the string until the pinyin of Tielin Station is "tie3lin2zhan4", which matches the pinyin string A. 0 The part "tie3ling1zhang1" in the text is calculated by S1: the wrong ing and ang can be replaced by the correct in and an. Although the number of operations is 2, this is a fuzzy sound replacement, and the phoneme edit distance = 0.5 + 0.5 = 1; the pitch distance is calculated by S2: the pitch difference between lin2 and ling1 is 1, that is, |2-1| = 1 < 2, and the pitch distance is added by 0, but the pitch difference between zhan4 and zhang1 is 3, that is, |4-1| = 3 ≥ 2, and the pitch distance is added by 1, the pitch distance = 0 + 1 = 1, and the matching value D = 1 + 1 = 2;

[0089] S4. According to the length of the vocabulary, the matching value is further modified. In the sub-library, "tie3lin2zhan4" contains k=3 pinyin characters. The modified matching value is The corrected matching value is less than the given threshold value of 1.7, a matching success signal is given, and a matching value of 1.59 is output. This is because the longer the vocabulary, the greater the probability of error.

[0090] In the above steps, the given threshold of 1.7 varies from person to person. For some people who speak Mandarin well, the threshold can be set to 1.2 for higher accuracy. For some people who speak Mandarin poorly, the threshold can be set to 2 to improve fault tolerance. Therefore, the given threshold can be bound to the user. The method is:

[0091] Threshold = k 3 × the average matching value of the current user, where k 3 The coefficient is set. The identity of the current user is obtained by authority recognition. During the user's use, each time a matched text sentence is output, the average matching value of all the words in the sentence is calculated, and then the total average value of all the sentences used in the user's history is accumulated and recorded under the user name.

[0092] exist Figure 1 The process involves permission identification. For details, see Figure 2 As shown, the process is as follows:

[0093] 1) Get the wake-up voice. The voice is acquired through the microphone device and stored on the hard disk in the format of a .wav file. The computer reads the voice by calling the wave module in Python.

[0094] 2) Extract the Fbank energy features of each frame of the speech obtained in step 1). In this part, the python_speech_features module of python is used to obtain the Fbank energy features of the speech.

[0095] 3) Input the Fbank energy features into a pre-trained DNN model and obtain a high-level feature expression based on deep learning. In this section, the DNN model architecture includes four hidden layers, which is implemented using the keras module of python.

[0096] 4) Perform cosine similarity matching on the high-level feature expression obtained in step 3) and the user voice feature bound to the user login account; if the cosine similarity is lower than or equal to the threshold, it is judged that the user does not match the user bound to the login account and has no use permission; if it is higher than the threshold, it is judged that the user has use permission.

[0097] exist Figure 1 The process involves speech noise reduction. For details, see Figure 3 As shown, the process is as follows:

[0098] 1) Get the voice read by the user.

[0099] 2) Use short-time Fourier transform (STFT) to extract the amplitude spectrum and phase spectrum of the speech obtained in 1). In this step, the speech is divided into several small segments according to a window of 1 second. Then, the short-time Fourier transform (STFT) provided by the librosa module of Python is used to obtain the amplitude spectrum and phase spectrum of each small segment.

[0100] 3) Input the amplitude spectrum obtained in step 2) into a pre-trained Unet model to obtain a denoised amplitude spectrum based on deep learning. In this step, the amplitude spectrum of each small segment is converted into a denoised amplitude spectrum through the Unet model. The Unet model is implemented using the keras module of python.

[0101] 4) The denoised amplitude spectrum obtained in step 3) and the phase spectrum obtained in step 2) are converted into denoised speech through inverse short-time Fourier transform (ISTFT). Through this step, the denoised speech corresponding to each small segment of speech in 2) is obtained, and then the denoised speech corresponding to the speech in step 1) is obtained through speech combination. The specific function of this step is implemented by the librosa module of python.

[0102] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for improving speech recognition rate in a professional field, characterized in that: include: Level priority matching: after converting speech to pinyin, a pinyin string consisting of letters and tones is obtained, which is named A. In the process of converting A into text, it is preferentially matched with the pinyin of the first-level difference frequency sub-library vocabulary of the difference frequency special word library. If the match is successful, part of the pinyin of A is converted into text. If the match is unsuccessful, the second-level difference frequency sub-library is considered. The matching is to find the similarity between a part of the pinyin of A and the correct pinyin of a Chinese phrase or character, which is also referred to as the matching of pinyin and text or vocabulary; the difference frequency special word library is composed of local special words and their pinyin to form a first-level difference frequency sub-library, and professional terms and their pinyin to form a second-level difference frequency sub-library. The sound is composed of a secondary difference frequency sub-library; the queuing order of a word in the difference frequency sub-library is determined by the difference of its frequency, that is, the higher the frequency of the word in the special data, the higher the queuing order of the word, and the higher the frequency of the word in the general data, the lower the queuing order of the word; the word refers to a Chinese phrase, a Chinese phrase includes at least 2 Chinese characters, the local special vocabulary refers to the vocabulary used only in the local machine, local area network, a specific region, group or department, local special vocabulary and professional terms are collectively referred to as special vocabulary, and the others are called general vocabulary, and the difference in frequency is referred to as difference frequency; Frequency priority matching: After the level priority matching is completed, the remaining pinyin of A is matched with the pinyin of common words. Non-special words with high frequency appearing in common data are matched first, and finally the remaining pinyin is matched with the pinyin of a single Chinese character; The construction of the difference frequency special word library comprises the following steps: S1. Collect text data related to local work, including work plans, work files, meeting records, equipment records and call texts on the local machine and the local area network; S2. Clean and segment the text data collected in the previous step, and obtain a vocabulary list. Then, perform vocabulary frequency statistics on the vocabulary list, and obtain a local special word frequency dictionary, where word frequency = the number of times the word is repeated × the length of the word / the total number of words in all the data. The local special word frequency dictionary contains all the words and their word frequencies that appear in the local text data; S3, searching the academic articles of the profession on the Internet, performing word frequency statistics on the obtained vocabulary list, and obtaining a professional term frequency dictionary, which includes all the words and their word frequencies that appear in the academic articles of the profession; S4. Perform word frequency statistics on general data, including the People's Daily corpus and news from three major websites, Sina, Sohu, and NetEase, to obtain a general word frequency dictionary, which includes all the words and their frequencies that appear in the general data; S5. Perform a frequency difference operation on each word in the professional term frequency dictionary, namely: The first difference frequency value = the professional term frequency of a word - k1 × its general term frequency, where k1 is a set coefficient. The professional term frequency and general term frequency are obtained from the professional term frequency dictionary and the general term frequency dictionary respectively. The word frequency that cannot be found in the dictionary is counted as 0; The words with primary difference frequency value less than or equal to 0 are discarded, and those with difference frequency value greater than 0 are stored in the secondary difference frequency sub-bank. The higher the difference frequency value, the higher the word is in the queue of the difference frequency sub-bank. S6. Perform a secondary frequency difference operation on each word in the local special word frequency dictionary, namely: Second difference frequency value = local specific word frequency of a word - k2 × its professional term frequency - k1 × its general word frequency, where k2 is a set coefficient, and the local specific word frequency is obtained from the local specific word frequency dictionary; Words with secondary difference frequency values ​​greater than 0 are stored in the primary difference frequency sub-bank, otherwise they are discarded. Words with higher difference frequency values ​​are queued at the front of the difference frequency sub-bank. S7. Associating each word in the difference frequency special word library with its corresponding pinyin, and finally forming a difference frequency special word library for professional fields, wherein the difference frequency special word library is composed of the first and second level difference frequency sub-libraries.

2. A method for improving speech recognition rate in a professional field according to claim 1, characterized in that: The steps of obtaining a pinyin string composed of letters and tones include: S1. Authority identification: Perform voiceprint recognition on the wake-up voice input by the user. If the user has the authority to use, proceed to step S2, otherwise, exit; S2, speech noise reduction: perform speech noise reduction on user speech to reduce the interference of noise on recognition; S3, voice to pinyin conversion: perform voice analysis and recognition on the noise-reduced voice obtained in step S2, and obtain the entire sentence pinyin string corresponding to the voice.

3. A method for improving speech recognition rate in a professional field according to claim 2, characterized in that: The authority identification comprises the following steps: S11, obtaining a wake-up voice; S12, extracting the Fbank energy features of each frame in the speech obtained in step S11; S13, input the Fbank energy features into a pre-trained DNN model and obtain a high-level feature expression based on deep learning; S14. Perform cosine similarity matching between the high-level feature expression obtained in step S13 and the user voice feature bound to the user login account; if the cosine similarity is lower than or equal to the set value, it is determined that the user does not match the user bound to the login account and has no use permission; if it is higher than the set value, it is determined that the user has use permission.

4. A method for improving speech recognition rate in a professional field as claimed in claim 1, characterized in that: The level priority matching step includes: Reverse word selection: select the pinyin of the word with the highest difference frequency value from the words that have not been matched in the first-level difference frequency sub-library, and set its name as B. If the pinyin of the words in the first-level difference frequency sub-library have been matched, then move on to the second-level difference frequency sub-library; Convert pinyin at any position: Search for a substring C similar to B in A. If B matches C successfully, convert C into the corresponding Chinese phrase. If there are multiple substrings C similar to B in A, repeat the above operation. The substring C can be located at any position in A.

5. A method for improving speech recognition rate in a professional field as claimed in claim 1 or 4, characterized in that: The matching comprises the following steps: Calculate the phoneme edit distance: refers to the minimum number of phoneme edit operations required to convert one pinyin string into another; the phoneme refers to the initial consonant or final of the pinyin, and the permitted edit operations include: inserting an initial consonant / final, deleting an initial consonant / final, replacing one initial consonant / final with another, and a replacement between ambiguous sounds is only counted as 0.5 times, and the above operations do not include tones; Calculate pitch distance: Then the pitch distance is increased by 1, otherwise it is not increased. Here, x and y are the tones after the corresponding finals in the two pinyin strings. The tones after all the finals are calculated and accumulated to get the pitch distance; Calculate the matching value = phoneme edit distance + pitch distance; If the match is a special word, then when the match value is less than a given threshold, the match value and the match success signal are output, otherwise a match failure signal is given; If the match is a common term, the matching value is output.

6. A method for improving speech recognition rate in a professional field as claimed in claim 5, characterized in that: According to the length of the word, the matching value is further modified. Assuming that the pinyin of the word to be matched corresponds to j characters, according to the calculated matching value D, the modified matching value is .

7. A method for improving speech recognition rate in a professional field as claimed in claim 5, characterized in that: The threshold is calculated according to the following formula: Threshold = k3 × average matching value of the current user, where k3 is a set coefficient. The identity of the current user is obtained by authority identification. During the user's use, each time a matched text sentence is output, the average matching value of all the words in the sentence is calculated, and then the total average value of all the sentences used in the user's history is accumulated and recorded under the user name.