TWI242729B - Speech database establishment and recognition method and system thereof - Google Patents

Abstract

The invention provides a speech database establishment and recognition method and system thereof. The method includes enabling a word segmentation module to divide a speech signal inputted by a user through an input unit into at least a vowel speech module in accordance with a rule pre-defined by the user and to store the vowel speech module in a database through a storage module; enabling the storage module to store a vowel speech module arrangement sequence corresponding to the speech signal inputted by the user; enabling a speech recognition module to divide the speech signal into at least a vowel speech module to be recognized in accordance with the rule pre-defined by the user while the user inputs the speech signal through the input unit; enabling the speech recognition module to search if there is a match with arrangement sequence data of the vowel speech module in the database, retrieving the arrangement sequence data if yes, and listing a possible combination matched with the vowel speech module arrangement sequence if no. With the voice database establishment and recognition method and the word segmentation mechanism of the system, it is able to provide a concise speech recognition database structure and a speech recognition method and system varied according to the feature of user.

Description

1242729 V. Description of the invention (1) [Technical field to which the invention belongs] A method and system for establishing and identifying a speech database, more specifically, a method and system for improving speech training and identification efficiency through word segmentation technology . [Previous technology] With the rapid development of the electronic information industry, various powerful and low-priced consumer electronic information products have come out. As for the most common computers, various software and hardware have continued to strengthen their functions. In contrast, the work that the computer can handle is no longer limited to program operation or data processing as before, but plays a role of a media of video and audio transmission. In other words, computers have moved from companies or laboratories into the field of home appliances. This is not only the case with computers. On the other hand, the electrification products around life are increasingly emphasizing computerization. Through various embedded systems, such as televisions, refrigerators, or washing machines, they have gradually become small computers. In other words, the user can set and operate different function options through a simple human-machine interface. Furthermore, in addition to the one-way setting operation, the user can still communicate with the electrified product, and even communicate with the outside world via email. Yes, in the past, simple home electrical products were also computerized 'and developed in the direction of Yudai Electric. According to the previous description, whether it is a computerized home appliance or a computerized home appliance, the user must communicate with the machine through a human-machine interface. Taking the input unit as an example, the most commonly used is a keyboard button, mouse, or other similar input unit. . Although these input units can provide users with

17458 MSI.ptd Page 5 1242729 V. Description of the Invention (2) The volume of required instructions or resources is important; secondly, the traditional input side computer appliances or the input method to solve this choice can be greatly reduced. Need to communicate with the people, for users who want to use the voice database to test the input of the Chinese Mingguo Voice Learning System, but it still requires light, short and small users may not be familiar with the computer with a home appliance computer Conversion is an inconvenience. For example, in terms of the input unit, it is usually difficult for the computer person to dim the computer, or it is difficult for them to pass through. For the implementation of the above problems, using voice, only one that is not as dark as the communication of Mai Shao ’s products is used as a loser, and there is a bulletin and its learning, which is used to identify the input. Voice to training training device. When the model is almost covered, it can effectively input the signal. The above-mentioned language system has used a set of all the media that are based on the volume and the oral brain operation, to identify the method 3 0 8 6 6 6 ", the identification and use of example sentences The user ’s learning example of the body ’s voice and voice model is input in a gram-like manner. The user must first recognize the system ’s patented technology, and the special feature of the voice message. With the training characteristics of the voice model, the internal language is used for traditional text or images. Selected sound input unit, display space. Furthermore, it is convenient to use the instruction to enter the machine. But there must be a wealth of information to reveal the characteristics of the number and the calculation of a borrowing type and later, which should result in the "characteristics of wisdom" identified by the mechanical parameters, and the use of the new among which the user officially recognizes the type of Mandarin. After investigating first, the sound learning and recognition systems and methods of the user using the voice line of the economic and academic materials are the current speech recognition technology. However, it has considerable shortcomings.

Guoyuan

17458 MSI.ptd Page 6 1242729 V. Description of the invention (3) The user must first establish a user with clear features to establish and recognize, not only lacking in humanity, repeated trial and error, if there are changes, the rate will decline. Also, learn about the number and content of Model (HMM) models before you enter. And another DTW) is used to complete the speech of the language. In other words, the amount of recognized speech is bound to establish the above-mentioned hidden horse comprehensive establishment and identification party. [Summary of the invention] To solve the problem of providing a speech data close to predetermined standards Speed and volume read example sentences, borrowed features of the voice 俾 reduce the chance of the system to identify errors, and the same habit of entering speech with stable reading. This other way of speech requires the user to adapt to the recognition habits of the machine. For users with less agility, it is necessary to use a formula to obtain better recognition results. In addition, the user must re-adjust the user's characteristics, otherwise Hidden Markov is used as the judgment criterion for speech recognition. The disadvantage is that the module is preset. When the user sets the number of models and Eight. The speech data that conforms to these models to complete the model. Dynamic Time Correction (Dynamic T ime W arping; the technology of sound identification, based on the user's input in advance as a comparison benchmark, there is no so-called module The number of data and content of the person's rotation determines its ability to distinguish and its content. Once a certain degree of recognition is required, a very large database, the same situation will also occur in the Kuff model speech recognition technology "How to provide a more efficient speech database method and system has become a problem to be solved urgently. Describe the shortcomings of the conventional technology" The main purpose of the present invention is to establish a database and identification method and system, through words

17458 MSI.ptd 7th! 1242729 V. Description of the invention (4) The segmentation mechanism can increase the number of samples in the database and increase the probability of successful speech training and recognition. Another object of the present invention is to provide a method and system for establishing and recognizing a speech database. Through the word segmentation mechanism, users do not need to repeatedly learn the pronunciation speed, frequency, and / or intonation of example sentences from the beginning, so they can save before using speech recognition. Time to establish personal voice characteristics. Another object of the present invention is to provide a method and system for establishing and identifying a voice database. Through a word combination mechanism, a certain amount of voice data can be arranged and combined into a complex word combination, so a large amount of database data can be saved. . Yet another object of the present invention is to provide a method and system for establishing and recognizing a voice database. Through the word segmentation mechanism, even if the user's pronunciation does not meet the standard, a fairly close recognition result can still be obtained. In order to achieve the above and other objectives, the speech database establishment and identification system of the present invention includes: a word segmentation module, which is used to convert the voice signal input by the user through an input unit according to the user's preset Based on the reference, the voice signal is divided into at least one mother voice module, and the mother voice module is stored in a database; a storage module is used to divide the at least the word division module into the at least one A mother voice module and the arrangement order of the mother voice module corresponding to the input signal are stored in the database; a speech recognition module is used to input a voice signal through the input unit according to the user The preset standard divides the voice signal into at least one mother voice module to be identified, and searches the database for data that allows the arrangement order of the mother voice module to be identified, and if so, retrieves the arrangement order.

17458 MSI.ptd Page 8 1242729 V. Description of the invention (5) Sequence information; if not, list the possible combinations that allow the arrangement order of the parent voice module. The method for establishing and recognizing a system through the voice database and performing voice training and recognition is as follows: First, the word segmentation module is configured to convert the voice signal input by the user through an input unit and the voice according to a preset reference of the user. The signal is divided into at least one mother voice module, and the mother voice module is stored in a database through a storage module; secondly, the storage module will correspond to the mother voice module of the voice signal input by the user The arrangement order of the groups is stored in the database; then, when the user inputs a voice signal through the input unit, the voice recognition module divides the voice signal into at least one mother voice mode to be recognized according to a preset reference of the user. Furthermore, the voice recognition module module is caused to search the database for the arrangement order data of the mother voice module to be identified, and if so, retrieve the arrangement order data; if not, list them This permits a possible combination of the arrangement order of the mother voice module. Compared with the conventional voice training and recognition technology, the method and system for establishing and recognizing a voice database of the present invention can save the number of samples in the database and increase the probability of successful voice training and recognition. Feature time. In addition, even if the user's pronunciation is not up to standard, a fairly close recognition result can still be obtained to increase the chance of successful recognition. [Embodiment] The following describes the embodiment of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be modified by other

17458 微星 .ptd Page 9 1242729 -----__ V. Description of the invention (6) ^ ~~ -—____ 白 hotel # 骨 杳 # ^ 可 基 ': Two examples are implemented or applied, each item in this specification The details are the same, and the same viewpoints and applications do not deviate from the modifications of the present invention. S 5 < Each one is created under the accumulation of the first figure 'in this embodiment' The voice database of the present invention is applied to a personal computer 1 through the system to provide use, η The voice database of this meal is established Communicate with the identification method and the control panel, Book 1 such as operation and / or setting. It should be specially explained that "'Jie Xi Acupoint', the voice database establishment and identification system of the present invention, and the software and hardware architecture of the personal electricity / transmission / sacrifice system are more complicated, in order to highlight the technology of the present invention. Only the parts related to the technical features of the present invention will be shown t °, and the method of establishing and recognizing the voice database of the present invention and the system retransmission are applied to workstations' notebook computers, LCD computers, tablet computers, palmtop computers, personal digital Assistants and mobile phones. The speech database establishment and recognition system of the present invention includes at least: an input unit 10, a word segmentation module 12, a database 14, a storage module 16, and a speech recognition module 18. The input unit 丨 〇 is used to provide a user to input a voice signal into the voice database establishment and identification system with a sound collection function unit, in this example, it is a microphone (in icr 〇ph ο ne). The word segmentation module 12 is used to divide the voice signal input by the user through the input list 10, and divide the voice signal into at least one mother voice module according to the preset reference of the user. In this embodiment, the word division group 1 2 includes an analog digital conversion unit (not shown) for converting an analog voice signal input by a user into a digital signal. Therefore, when using

17458 MSI.ptd

Page 〖〇1242729 V. Description of the invention (7) When the user sets up a voice database, he must input the analog voice signal of a group of words "the weather is fine today" through the input unit 10, the word segmentation module 1 2 This is then converted into a digital signal for processing. After the conversion of the digital signal format is completed, the word segmentation module 12 then divides the group of words "f at" and performs segmentation of the group and word according to the speech segmentation criterion set by the user. In this embodiment, the word segmentation module of the present invention analyzes the distribution relationship of the speech signal on the frequency spectrum. In other words, when the user inputs the voice from the mouth through the input unit 10, he must go through the operation of frequency conversion (Fourier transform) to obtain the data of the voice signal on the frequency spectrum. The original data includes at least the frequency. , Energy and time relationship, at a time point near a certain time point t (... t-2, t-1, t + 1, t + 2 · · ·) to get the energy data at each frequency, by calculating The average number and correlation coefficient are used to obtain the differences between them. In addition, in the two-dimensional data of "frequency" and "time", the edge detection principle of the two-dimensional image is used to obtain the boundary between two dissimilar speech segments, and then a variable threshold is used. The data and environment vary, so as to identify that the energy change in frequency at one time point and another time point has a significant performance that exceeds the threshold, and is used as the basis for segmenting words. In addition, a similar mother voice module is obtained between the division lines and the division lines. In other words, after inputting a certain set of word data, after the calculation and processing of the aforementioned word segmentation technology, at least one mother speech module can be obtained. According to the foregoing description, in this embodiment, the set of words input by the user may be divided into three parts such as "f", "a", and "t". Yu Benshi

] 7458 微星. (1 Page 11 1242729 V. Description of the invention (8) In the embodiment, let the three parts "f", "a" and "t" be the mother voice modules "A", "A", " B "and" C ". That is, the module" ABC "composed of the mother voice module stands for" fat ". The storage module 1 4 'is used to divide the word into modules 12 The at least one mother voice module and the arrangement order of the mother voice module corresponding to the input signal are stored in the database 14. According to the foregoing description, in this embodiment, the user uses the input unit 10 The input group of words may be divided into three parts, such as "f", "a", and "t". Therefore, the storage module 14 immediately follows the three mother voices such as "A", "B", and "C" The module and the module "ABC" are stored in the database 14. In addition, during the establishment of the database 14, the user recovers the sequence relationship between the input sound and the sound through the input unit 10 (Sequentia 1 cue) longer "fat" (fat between fa) and short sequence relationship (fa-1 between fa-1) "Fat". Among them, it is assumed that the module corresponding to the "fat" of the prolonged sound between fa is "DC", and the module corresponding to the "fat" that has a shorter sequence between fa-1 is "fat". E ". The user will regard the phrase corresponding to the" ABC "," DC "and" E "modules as" fat ". The speech recognition module 18 is used by the user to pass When the input unit 10 inputs a voice signal, the voice signal is divided into at least one mother voice module to be identified according to a preset reference by the user, and a search is performed in the database 14 for whether the mother voice module to be identified is allowed. Group arrangement order data, if available, retrieve the arrangement order data; if not, list possible combinations that allow the arrangement order of the mother voice module. As mentioned earlier, in this embodiment, the phrase

17458 微星 .ptd Page 12 1242729 V. Description of the invention (9) The voice signal segmentation method of the voice recognition module 18 is the same as that of the word segmentation module 12 described above. Through the aforementioned segmentation technology, the user must pass The voice signal to be recognized input by the input unit 10 is divided into at least one mother voice module to be recognized. At this time, if the user inputs a set of words "f at", the speech recognition module 18 will be divided into three to-be-recognized mother speech modules such as "f", "a", and "t", that is, three A to-be-recognized mother voice module "A", "B" and "C" are composed of the to-be-recognized module "ABC". Then, by using the technology of dynamic time correction, it is searched in the database 14 whether there is stored word data that allows the module "ABC" to be identified, and if so, it is identified that the user inputs through the input unit 10 The word system is "fat"; if there is no matching mother voice module arrangement order, the possible combinations that match the mother voice modules are retrieved from the database 14 for further confirmation by the user What words are entered? Based on this, the user can create the sequence data according to the listed possibilities. It should be noted that if the “fat” inputted by the user through the input unit 10 is “fat” with a long tone between f-a or “f a t” with a shorter sequence between f-a-1. Then the modules identified by the speech recognition module 18 will be "DC" or "E", respectively. According to the previous description, during the establishment of the database 10, the user has set the corresponding phrase of the "fat" module "DC" or "E" of the aforementioned long or short note to "fat" . Therefore, even if the user inputs non-standard "f at" voice data through the input unit 10, the voice recognition module 18 still has to recognize the phrase "fat".

17458 微星 .ptd Page 13 1242729 V. Description of the invention (10) On the other hand, if the user creates another module "ABFC" for the corresponding phrase "fact". Then, when the user inputs "fact" through the input unit 10, but the mother voice module of the "c" sound is not recognized because the user's pronunciation is not standard, the voice recognition module 18 will have With a weightedva 1 ue mechanism such as low recognition probability, the module corresponding to the non-standard speech is determined to be "ABC" or "ABFC". If the recognition probability of "ABC" is high, then the The speech recognition module 18 will recognize the speech input by the user into the phrase "fat" corresponding to the "ABC" module. Please refer to FIG. 2, which shows the steps of the method for establishing and recognizing the voice database of the present invention: In step S2 01, the word segmentation module 12 is configured to direct the voice signal input by the user through the input unit 10 , The voice signal is divided into at least one mother voice module according to a preset reference set by the user. According to the foregoing description, in this embodiment, when the user is creating a voice database, the analog voice signal of a group of words "f at" can be input through the input unit 10, and the word segmentation module 12 will then This is converted into a digital signal for processing, and it is divided into three parts: "f", "a", and "t". After the segmentation is completed, the different mother voice modules of δ are stored in the database 14. Then, step S202 is performed. In all directions, in step S202, the storage module 丨 4 is configured to store the at least one mother voice module which is separated from the word segmentation module 1 2 and the mother corresponding to the input signal =: the board group is stored Go to the database 14. According to the previous description, in # K ^ 例 ’the storage module 1 4 has to pass the user through the input unit

Page 14 1242729 V. Description of the invention (11) The order of the three mother voice modules "f", "a" and "t" entered in "fat" is stored in the database 14 Then, proceed to step S203. In step S203, the voice recognition module 18 is caused to divide the voice signal into at least one mother voice mode to be recognized when the user inputs a voice signal through the input unit 10 according to a preset reference of the user. group. According to the foregoing, in this embodiment, the voice signal segmentation method of the speech recognition module 18 is the same as that of the word segmentation module 12 described above. Through the aforementioned segmentation technology, the user can pass the input unit 1 0 The input voice signal to be recognized is divided into at least one mother voice module to be recognized. According to the previous description, if the user enters a set of words "fat", the speech recognition module 18 will be divided into three to-be-recognized mother speech modules such as "f", "a", and "t", and then proceed Step S204: In step S204, the speech recognition module 18 is caused to search whether there is data in the database 14 that permits the arrangement order data of the mother speech module to be recognized. According to the foregoing description, in this embodiment, the speech recognition module 18 uses dynamic time correction technology to search whether there is word data in the database 14 that allows the "fat" arrangement order, and if so, enter Go to step S 2 0 5; if not, go to step S 2 0 6. In step S205, the speech recognition module 18 is made to recognize that the word entered by the user through the input unit 10 is "f a t". In step S206, the speech recognition module 18 is made to retrieve possible combinations that match the mother speech modules from the database 14 for further confirmation by the user of the entered word data. Why.

17458 微星 .ptd Page 15 1242729 V. Description of the invention (12) In summary, the method and system for establishing and identifying a voice database of the present invention, in addition to increasing the number of samples in the database and not extending the voice of the database infinitely The number of samples was raised earlier to increase the efficiency of speech training and recognition success, so as to save time in establishing personal speech features. In addition, the method and system for establishing and recognizing a speech database of the present invention combine text to speech (TTS) to become an interactive dialogue system. The above-mentioned embodiments are merely illustrative for explaining the principle of the present invention and its effects, and are not intended to limit the present invention. Anyone skilled in the art can modify and change the above embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the rights of the present invention should be listed in the scope of patent application mentioned later.

17458 MSI. Mouth Seven () Page 16 1242729 Simple illustration of the diagram [Simplified illustration of the diagram] FIG. 1 is a block diagram showing the system architecture of the speech database creation and identification system of the present invention; and 2 The figure is a flowchart showing the steps of the method for establishing and identifying the speech database of the present invention. 1 Personal computer 10 Input unit 12 Word segmentation module 14 Database 1 6 Storage module 18 Speech recognition module

] 7458 MSI.pt-d Page 17

Claims (1)

1242729 VI. Scope of patent application 1. A method for establishing and recognizing a speech database, which can be applied to a data processing device to provide the function of speech recognition of the data processing device, including: (1) ordering a word segmentation module The group divides a voice signal input by a user through an input unit, divides the voice signal into at least one mother voice module according to a preset reference of the user, and stores the mother voice module in a data through a storage module. In the database; (2) order a storage module to store the mother voice module arrangement order corresponding to the voice signal input by the user in the database; (3) order a voice recognition module in the user through the input When the unit inputs a voice signal, the voice signal is divided into at least one to-be-recognized parent voice module according to a preset reference of the user; (4) The voice recognition module module is caused to search whether there is a permit in the database Recognize the arrangement order data of the mother speech module. If there is, go to step (5); if not, go to step (6); (5) Make the speech recognition module extract the data that allows the arrangement order. Data; and (6) order the speech recognition module to list possible combinations that allow the arrangement order of the mother speech module. 2. If the method of item 1 of the patent application scope is included, before the word segmentation module divides the voice signal, make the word segmentation module convert the received analog voice signal into a digital signal format. 3. The method of item 1 in the scope of patent application, wherein the word segmentation module analyzes the distribution relationship of the speech signal on the frequency spectrum. 17458 MSI. 卩 士 (1 Page 18 1242729 VI. Patent Application Scope 4. For the method of patent application No. 3, where the distribution relationship of the frequency spectrum includes two-dimensional data of "frequency" and "time", using two-dimensional data The edge detection principle of the image is used to obtain the boundary between two dissimilar speech segments. 5. For the method in the scope of patent application No. 4, wherein the boundary of the speech segment is a variable threshold, the threshold will be The voice data and the environment vary, so as to identify that the energy change in frequency between one time point and another time point has a significant performance that exceeds the threshold value, and is used as the basis for segmenting words. 6. If applying for a patent The method of the first item in the scope, wherein the word segmentation module is based on one of the speed, energy, and frequency of the voice data. 7. If the method of the first item in the scope of the patent application is applied, the method includes the different The arrangement order of the mother speech module is set to correspond to the same language phrase. 8. For the method of the first scope of the patent application, wherein the speech recognition module is corrected by dynamic time Dynamic T i in e W arping; DTW) technology to compare the mother voice module and the specific mother voice module arrangement order in the database, in order to obtain the closest to the user's input voice content Result. 9. The method of item 1 of the patent application scope, wherein the voice recognition module module searches for a predetermined weighting value in the database as to whether or not the arrangement order of the mother voice module to be identified is allowed. The criterion for judging the data. 10. If the method of the first item of the scope of patent application, the data processing device may be a data processing system compatible with a personal computer and embedded 17458 MSI.ptd Page 19 1242729 6. Scope of Patent Application. 1 1. The method according to item 10 of the patent application scope, wherein the personal computer compatible data processing system may be a workstation, a personal computer, a notebook computer, an LCD computer, a tablet computer, a palmtop computer, a personal digital assistant, and a mobile Call one of them. 1 2. A speech database establishment and recognition system, which can be applied to a data processing device to provide the speech recognition function of the data processing device, includes: a word segmentation module, which is used to use The voice signal input by an input unit is divided into at least one mother voice module according to a preset reference by the user, and the mother voice module is stored in a database; a storage module, It is used to store the at least one mother speech module divided by the word segmentation module and the arrangement order of the mother speech modules corresponding to the input signal to the database; and a speech recognition module, which is When the user inputs a voice signal through the input unit, the voice signal is divided into at least one to-be-recognized mother voice module according to a preset reference of the user, and a search is performed in the database to see if the to-be-recognized mother is allowed. Voice module arrangement order data, if available, retrieve the arrangement order data; if not, list possible combinations that allow the arrangement order of the mother voice module. 13. The system of item 12 in the scope of patent application, wherein the word segmentation module includes an analog digital conversion unit for converting an analog voice signal input by a user into a digital signal. 17458 MSI.ptd Page 20 1242729 VI. Scope of Patent Application 1 4. For the system of No. 12 scope of patent application, the word segmentation module analyzes the distribution relationship of speech signals on the frequency spectrum. 15. The system according to item 14 of the scope of patent application, wherein the distribution relationship of the frequency spectrum includes two-dimensional data of "frequency" and "time", and the edge detection principle of the two-dimensional image is used to obtain two dissimilar speech segments. Demarcation. 16. If the system of item 15 of the scope of patent application, the delimitation of the voice segment is a variable threshold value, and the threshold value may be changed due to different voice data and environment, so as to identify a certain The energy change in frequency between one time point and another time point has a significant performance that exceeds the threshold value, and 俾 is used as the basis for segmenting words. 17. The system according to item 12 of the scope of patent application, wherein the word segmentation module is based on one of the speed, energy and frequency of the speech data. 18. The system of item 12 in the scope of patent application, wherein the voice recognition module is based on the technology of Dynamic Time Correction (DTW) and the mother voice modes in the database. The group and the specific mother voice module arrangement order are compared to obtain the result closest to the user's input voice content. 19. If the system of item 12 in the scope of patent application, wherein the voice recognition module module searches for a predetermined weighted value in the database as to whether the order of the mother voice module to be recognized is allowed Judgment basis of information. 2 0. The system of item 12 in the scope of patent application, wherein the data processing device 17458 MSI.ptd Page 21 1242729 VI. Scope of Patent Application The device can be a personal computer compatible and embedded data processing system. 2 1. If the system of the scope of patent application No. 20, wherein the personal computer compatible data processing system can be a workstation, personal computer, notebook computer, LCD computer, tablet computer, palmtop computer, personal digital assistant and mobile Call one of them. 2 2. The system according to item 12 of the scope of patent application, wherein the database is a relational database. 17458 MSI 邛 1: (1 page 22