KR100285502B1 - Method for building phonetic database - Google Patents

Abstract

PURPOSE: A method for building a phonetic database is provided to promptly and accurately build a phonetic database by editing and labeling phonetic data simultaneous with the building of a phonetic file system proper to phonetic recognition training. CONSTITUTION: The method for building a phonetic database comprises steps of collecting phonetic database(500), designing a phonetic file system structure for database building(510), classifying the collected database with respect to the phonetic file system structure by converting the collected phonetic data to the phonetic data(520), labeling the respective classified files(530), confirming the appropriateness of labeling(540), and confirming if all phonetic files are appropriate to the phonetic file system structure(550). By building a phonetic database employing a labeling method, the accurate and prompt building is possible. Further, a cost for adding a phonetic database is reduced.

Description

How to build a voice database

The present invention relates to a speech database construction method, and more particularly, to a speech database construction method configured to efficiently construct a speech database required for training a speech recognizer.

In order to manufacture a speech recognition system, data must be converted from the collected speech data into a form that is easy to use for learning in the speech recognition system. In the beginning, when collecting voice data, a strategy can be collected and collected as it is, but this method has a detrimental effect on the natural voice collection, which degrades the performance of the entire voice recognition system. Therefore, it is desirable to collect natural voice data under real environment.

FIG. 1 is a diagram conceptually showing a general speech recognition system. When a speaker's voice is input to the feature extractor 10, the feature extractor 10 extracts a feature of the input voice and undergoes the feature extraction. The voice data is inputted to the training department 12, and then standardized through the training process, and then inputted to the reference pattern unit 14. The recognition department 16 shows the recognition result through comparison with the reference pattern generated by the training department 12.

Therefore, in the speech recognition system, it is important to know how to create a good reference pattern in the training process, and to do this, it is necessary to know exactly which words or sentences each voice file produces during the training process.

In general, the following two methods are employed to know what words or sentences the collected voice data are uttered.

First, the first method is to set a certain order before collecting voice data and collect the data in that order. This method has the advantage that it is easy to manage the voice data later because the order is set in advance, but there are many restrictions because the recording must be performed only in this order.

For example, if the speaker speaking has missed a given word or sentence incorrectly, and if you skip one or two or speak the same word twice, all subsequent data will be wrong and you will not be able to train properly. do. Even if the speaker is speaking correctly, the voice is very different from the natural voice of the general user because the speaker is speaking in a very tense state and thus acts as a factor that degrades the performance of the voice recognizer.

The second method is to have the narrator make a natural recording. This method allows the narrator to speak freely, while requiring additional tools to organize and manage such disordered data.

This second method is very similar to the voice to be recognized in commercial services using the voice recognition because the speaker speaks naturally. Therefore, the voice recognition results will be very good when using this method. In addition, the data collected during the service can be used in addition to training again, which has the advantage of enabling the performance of the repetitive speech recognizer.

According to the second method, the speaker has to go through the task of converting the voice data spoken out of order into the index defined in the index file. In this case, the operation of converting the index into the corresponding index is called labeling. Conventionally, in order to label voice files, only a person listens and checks by hand.

Therefore, a problem arises that the processing speed is slow and the labeling result is inaccurate.

The present invention has been made to solve the above-described problems, and by building a voice file system suitable for a training process for voice recognition, by listening to, editing and labeling voice data, it is possible to quickly and accurately construct a voice database. The purpose is to provide a voice database construction method.

According to a preferred embodiment of the present invention to achieve the above object, a process of designing a hierarchical voice file system structure for collecting voice data input by a speaker and constructing a database, and voice file the collected voice data There is provided a voice database construction method comprising the step of classifying according to the structure of the voice file system and the step of labeling all the classified voice files.

1 is a conceptual diagram of a general speech recognition system.

2 is a structural diagram of a voice file system for constructing a voice database according to an embodiment of the present invention.

3 is a screen configuration diagram of a voice labeling system according to an embodiment of the present invention.

4 is a flowchart of a voice labeling process according to an embodiment of the present invention.

5 is a flowchart for explaining a voice database construction method of the present invention.

* Explanation of symbols for main parts of the drawings

10: feature extraction unit 12: training department

14: reference pattern part 16: recognition and government

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

2 is a structural diagram of a voice file system for constructing a voice database according to an embodiment of the present invention. When the speaker naturally speaks and collects voice data, the voice data is stored in each call unit. Since the service under consideration in the embodiment of the present invention is a voice recognition service using a telephone, voice data is recorded through a telephone line, and when a call is over, a new directory is created and the voice is stored in the directory.

That is, data 0, data 1, ..., and data N are generated one at the end of each call, and voice files are stored in each directory. Each voice file starts with fO and f1, f2 ‥‥‥ fn, etc., and the voice file is automatically cut into words or sentences by the end point detector whenever the speaker speaks a word or sentence. Are made.

In the figure, a list file is a file that lists and stores information on which voice files are included in each voice directory.

The list file contains information on what files exist in each directory, which makes it easy to automatically load and edit the next or previous file of a voice file.

This list file is also created one for each voice directory, so it is cumbersome to input one by one when there are many voice directories. So, again, you can save this list of list files as table files so that you can freely load them with a single button for each of the list files.

These hierarchical stored voice files are labeled with appropriate names according to their contents.

3 is a screen configuration diagram of a voice labeling system according to an embodiment of the present invention, in which a screen of the voice labeling system automatically loads the next voice file so that a previous file may be loaded for display and confirmation on the screen. Function, to create a waveform window on the screen to remove the noise components present in the voice file as needed, to show the waveform and to cut out and listen to a specific part of the waveform, and to edit the minute details. A number of buttons (ie Zoom In, Zoom Out, Correction, Play, Prev.Play, Subsequent Play) to perform functions that allow you to zoom in and out. button to specify next, store, load, load list file, qult, save, +/-, and so on.

In order to implement the screen configuration of the voice labeling system shown in FIG. 3, a user must listen to a voice file and designate a corresponding index on the screen. According to an embodiment of the present invention, an index is specified using a mouse or a keyboard. It was.

In other words, if the size of the word to be recognized, that is, the index is small, the index is displayed in a window scrolling on the screen, and it is designated by clicking with a mouse. Since it is cumbersome to find, the keyboard is faster and more efficient than the mouse.

In addition, when the keyboard is input as you hear it, you can quickly find the necessary voice file name and wrong pronunciation, and it is more accurate than using the mouse.

This is because if you specify it with the mouse, you can specify a different index by mistake.

Meanwhile, according to the same drawing, the labeled results are displayed on the screen so that the user can check them visually, and even in the process of verifying the labeled results, the labels and the sound displayed on the screen are different from each other. You can only modify it using the mouse or keyboard.

According to the voice labeling system shown in FIG. 3, numbers are added in a certain order according to the words to be recognized. This is because the target word to recognize (a sentence is a sequence of words) is limited.

Namely, if the first word to be recognized is "Father", the number is 0, and if "Mother", it is numbered in order, which is called an index file. Then, files such as f0 and f1, which are automatically generated in the voice file system structure, are given a corresponding index again according to the contents of the voice file.

For example, if the user's third voiced f2 was a "father," the file would be changed to fO using the mouse or keyboard. If you change all the files in that way, you can see only the file name later to see what the voice is.

If the same voice is pronounced twice, the same file name is given. Therefore, numbers such as .0, .1 ... are added after the file name to distinguish them. Thus, (f0.0 indicates a voice for the first spoken ″ father ″ and f0.1 indicates a voice for the second spoken ″ father ″.

4 is a flowchart of a voice labeling process according to an embodiment of the present invention, and the voice labeling process in the present invention is performed by the following procedure.

When the voice editing rule starts, the hierarchically constructed voice file system structure first loads the table file (step 400), and then loads the list file from the loaded table file (step 405).

Thereafter, a plurality of voice directories and also a plurality of voice files existing in each directory are sequentially selected (step 410), and then the selected voice files are loaded (step 415). Then, the loaded voice file is reproduced by using a button (step 420). At this time, the voice waveform and the labeling result of the reproduced voice file are displayed on the screen.

If there is a label for the currently playing voice file (″ Yes ″ in step 425), then the label is checked for correctness. If the label is not correct (″ no ″ in step 430), use the mouse or keyboard to select the correct label. (Step 435). If there is no label for the currently played voice file, the operation of step 435 is immediately performed.

On the other hand, if there is a label for the currently playing voice file and the label is correct, it is determined whether the voice file is the last one. If the result is not the last one (″ NO ″ in step 440), the step 410 is performed. In step 440, if the voice file is the last one (" Yes " in step 440), it is determined whether the list file is the end.

As a result of the determination, if it is not the end of the list file (″ NO ″ in step 445), the process returns to step 405 and repeats the operation from that step. If it is the end of the list file (″ Yes ″ in step 445), the table file If it is not the end of the table file (″ NO ″ in step 450), the process returns to step 400 and repeats the operation from that step, whereas if it is not the end of the table file (step 450). ″ Yes ″) voice editing rule ends.

Referring to the flowchart of FIG. 5, the method for constructing the voice database of the present invention employing the voice file system structure and the voice labeling scheme described above is as follows.

When the voice data collection procedure is determined, the voice data is collected (step 500), and the voice file system structure for the database construction is designed (step 510). Subsequently, the collected voice data is voiced and classified according to the voice file system structure (step 520).

Then, labeling is performed for each classified voice file (step 530). If the labeling is not correct (NO in step 540), the process returns to step 530 to re-label the voice file. On the other hand, if the labeling is correct (" Yes " at step 540) then it is determined whether all voice files have been classified according to the voice file system structure.

As a result, if all voice files are not classified according to the voice file system structure ("NO" in step 550), the process returns to step 520 and repeats the operation from that step, while all voice files If it is classified according to the structure of the voice file system (YES in step 550), the voice database construction operation is terminated.

According to the present invention as described above, by constructing a voice database by adopting a voice file system structure suitable for the training process for voice recognition and a labeling method that can listen to, edit / label the voice data, it is faster than the method of the Jongle In addition, accurate construction work is possible, as well as data collected during service provision can be used for training, and the cost of building by adding a voice database can be reduced.

The present invention is not limited only to the above-described embodiments, but may be modified and modified without departing from the scope of the present invention. Techniques by such modifications and variations should be regarded as belonging to the following claims.

Claims (6)

Creating a new voice directory every call unit, registering the generated directory name in a table file, creating a list file in the directory, and registering a path of the list file in the table file; Collecting voice data in every call unit, storing the voice file in the generated directory, and registering the voice file name in a list file; And loading a table file to sequentially read the list file names stored in the table file, and labeling while selectively reproducing the voice file registered in the list file. The method of claim 1, wherein the labeling of the voice file while selectively playing the voice file comprises the user directly listening to the voice file and selecting the voice file from the displayed index. The method of claim 2, wherein the selecting from the index comprises selecting using a mouse. The method of claim 1, wherein the labeling of the voice file while selectively playing the voice file comprises inputting directly through a keyboard. The method of claim 4, wherein the step of directly inputting through the keyboard comprises notifying the voice file name to be stored if it does not match the name of the voice file to be stored, so as to detect a wrong pronunciation. The voice database of claim 1, wherein the reproducing and labeling of the voice file comprises removing the noise component if a noise component is present in the voice file and deleting the wrong recording data. How to build.

Images