KR100282048B1 - Wave file automatic classification method of car voice recognizer - Google Patents

Abstract

The present invention relates to a method for automatically classifying a wave file of a vehicle voice recognizer to automatically classify a wave file according to a desired condition and sound quality in a conditional voice database constructed for a vehicle voice recognition. Condition of sound quality, condition of judging gender of speaker who spoke each voice, condition of age of speaker who spoke each voice, condition of recording place, condition of driving condition, amplitude of voice waveform Wave condition determination process to determine the condition according to the condition and dialect area, directory initialization process to input information about SPL directory and wave directory, SPL file checking process to list SPL and confirm it to the user, and Create a list file that creates a list of wave files that match the conditions based on the conditions and directory structure. It achieved by performing the information one by one.

Description

Wave file automatic classification method of car voice recognizer

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vehicle, and in particular, a wave file of a vehicle speech recognizer that can automatically classify a wave file (WAVEFILE) according to a desired condition and sound quality in a conditional speech database constructed for vehicle speech recognition. Automatic classification method.

In general, speech recognition refers to a general operation of extracting a feature pattern from a human voice and comparing and recognizing it so that a machine recognizes a speech produced by a human being.

In order to perform recognition, first select a word (or command) for recognition, speak these words to many people, and then synthesize and analyze the data to make the words common in all words (all talkers), and other words. You must first find some pattern that is distinct from.

When this pattern is found, it is stored in the system, and when a user enters a certain word into the microphone, the voice is extracted in the same way, and then compared to the pattern of each word previously stored to find the closest word. I think it was recognized.

When the recognition is performed, a variety of vehicle electronics outputs a digital protocol that can be recognized, and the vehicle electronics (eg, air conditioners, car audio, etc.) are automatically operated according to this command protocol.

As an example of the voice and control commands input to the microphone, "volume up" is a command to increase the overall volume of the car audio system by one step, and "FM" is a command to set the mode of the car audio to FM.

In more detail, the internal operation sequence of the voice recognition system, first, the voice input unit uses an external microphone to convert the vibration of air from the outside into an electrical signal, and the noise input from the LPF module requires other than noise or other voices. It filters out missing signals and converts these filtered analog analog signals from A / D converters to digital signals.

Then, before performing the speech recognition, the preprocessing unit performs preprocessing to perform proper signal processing that minimizes the parts that are difficult to be machined in the voice signal and leaves only the parts that are easy to be identified by the machine.

In addition, the feature extractor extracts the time-frequency features of the voice that are easy for the machine to recognize from the speech, and compares the extracted speech features with the reference patterns that have been prepared for each word. The voice recognition system determines that the input voice is a word model.

And when the word is determined, the recognition system delivers the protocol to the electronic device to which the word is to be performed, and the electronic device operates accordingly.

However, it is very difficult to classify a wave file of a desired condition in a speech database of such a conventional speech recognition system, and it is often used without classification.

In addition, even to classify wave files had to go through each person's hand, there was a problem that leads to unnecessary waste of time and manpower.

The present invention has been made to solve such a conventional problem, and an object of the present invention is to enable a software to easily classify a large amount of database recorded according to the environment of a vehicle without going through a human hand. The present invention provides a method for automatically classifying a wave file of a vehicle voice recognizer to effectively use a voice database constructed by region, speed, sequence, and condition.

1 is a block diagram showing the configuration of a voice recognition system

2 is an overall operation flow diagram of a wave file automatic classification method of a vehicle voice recognizer of the present invention

FIG. 3 is a flowchart illustrating a wave condition determining process of FIG. 2.

4 is a flowchart illustrating a directory initialization process of FIG. 2.

FIG. 5 is a flowchart illustrating a process of checking an SPL file in FIG.

6 is a flowchart illustrating an operation of creating a list file of FIG. 2.

<Explanation of symbols for main parts of drawing>

1: Voice section detection section 2: Feature coefficient extraction section

3: Voice recognition section 4: Database section

5: post-processing unit

In order to achieve the above object, the present invention reads the information files in which the overall information of the database, the speaker information, and the sound quality are recorded in a directory unit, so that a list of wave files that satisfies a user's desired condition is set according to a predetermined standard. It is written and saved as a file in ASCII text format.

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

As shown in FIG. 1, the present invention provides a speech section detecting unit 1 for detecting a speech section from an input signal, and extracts a feature from the speech section detected by the speech section detecting section 1 using an MFCC coefficient. A feature coefficient extraction section 2, a speech recognition section 3 for recognizing a speech signal using an algorithm such as an HMM, a database section 4 storing a word model learned by the speech signal, and It is composed of a post-processing section 5 for outputting the recognized words by determining the effectiveness of the speech signal recognized by the speech recognition section (3).

The present invention having such a configuration is a DOS-based program, which must be run in English mode on a DOS-based basis, and can operate anywhere in the IBM-AT / 386 or higher model, and only 8.3 format is applied to directories and file names. It is possible.

As shown in FIG. 2, the present invention is largely composed of four steps: a wave condition determination process (S10), a directory initialization process (S20), a SPL (Speech Logging) file check process (S30), and a list file creation process (S40). It executes tasks by executing them sequentially.

First, the wave condition determination process (S10) is a part for determining a condition for selecting a wave file, as shown in FIG. 3, initializes the condition value (S11), corrects the condition value (S12), and then corrects it. The measured value is viewed through the monitor (S13).

If it is determined that there are no modifications, it returns if there is no modification, and if there is a modification, corrects again and returns to the step S12 (S14, S15).

Here, seven conditions are required to perform the recognition.

The first condition is sound quality, which is a classification condition according to sound quality of a voice file.

For each voice waveform stored in the voice database, it is classified into five levels from A to E according to whether noise is included and difference in sound quality.

Level A is an ideal waveform with little noise, while level B is relatively good in voice, but low noise, low cough, indefinite echoes, breaths, and slow pronunciation are acceptable, and level C is a moderate sound quality. The louder noise is when the front and rear silence interval is shorter than 100ms and longer than 10ms, and the D level is lower than the C level, when the instantaneous noise exceeds the voice, the sound is correct and understandable. When the silence period is shorter than 10ms, the recording state is bad, the E level is very low sound quality, when the silence period is less than 0ms, when the word is different from the recognition word, when the pronunciation is not understood.

Here, a plurality of sound quality classification conditions may be selected for each voice. For example, if the user inputs A, B as the condition, a voice having a sound quality condition of A or B is selected among all voice databases, and the rest are selected. It doesn't work.

The second condition is a classification condition for determining the gender of the speaker who spoke each voice.

The user can select one of [man and woman], [man only], and [woman only].

[Both male and female] means that the speaker who spoke the voice does not care whether it is male or female, and this condition is skipped in the search condition.

[Man only] means that the voice speaker selects a voice file that is a male, and [Woman only] allows a voice speaker to select only a voice file that is a female.

The third condition is a classification condition representing the age of the speaker who spoke each voice.

Each age group is divided into 20s, 30s, 40s and 50s.

The user is required to enter one of these numbers. For example, if you enter 20, only 20 are selected from the voice database. If you select 0, this condition is skipped.

The fourth condition is a classification condition according to the recording place.

There are two kinds of voice recordings recorded in the vehicle and in the office environment, and [wherever] means that it does not matter where you recorded. If you select this condition, this classification condition is skipped.

[In-car] is selected only when recording in a car, and [In-office] is selected only when recording in an office environment.

The fifth condition is a classification condition according to the operating state.

This condition is required when the recording place is a vehicle among the classification conditions. This classification condition is classified according to the speed and driving state of the vehicle, as shown in Table 1 below.

TABLE 1

division speed engine Pan road 001 0 off off downtown 002 0 off off parking 003 0 off On downtown 004 0 off On parking 005 0 On off downtown 006 0 On off parking 007 0 On On downtown 008 0 On On parking 009 40 On off downtown 010 40 On off downtown 011 60 On On Suburb 012 60 On On Suburb 013 80 On off General road 014 100 On off General road 015 100 On On highway 016 100 On On highway

The user can select one of these.

[State 1-8] selects states 1 to 8, [State 9-12] selects states 9 to 12, and [State 13-16] selects states 13 to 16.

The sixth condition is a condition for classifying the amplitude of the speech waveform.

In the voice database, the amplitude level is divided into two cases of low level and other cases in the inspection process, because it is necessary to distinguish these two cases in the training process of recognition of voice data. In both cases, it means to accept. [Low level only] is selected only when the amplitude of the speech waveform is low, and [Relative level only] is selected only when the amplitude of the speech waveform is normal.

The seventh condition is the classification condition according to the dialect region.

This condition is to distinguish the provinces using the dialects, and the voices were recorded from speakers who speak various dialects in seven areas of Seoul, Gyeonggi, Busan, Daejeon, Gwangju and Gangneung.

This condition is added because speech recognition training needs to be performed by dividing this into dialects.

Next, a process of initializing the directory, which is an S20 process, will be described with reference to FIG. 4.

This process is to input information about SPL directory and wave directory. Basically, the initial file is read first, and the directory information is read from the initialization file (S21).

At this time, it is determined whether the directory is read as a file or input by hand, and in the case of a file, directory information is read as a file (S22, S23), or the directory is input by hand (S24).

Next, it is determined whether the directory information is correct, and if the directory information is correct, the monitor shows that the directory information is correct (S25, S26).

Here, the SPL directory is a directory in which the SPL file exists, and the wave directory is a directory one level up from the directory in which the wave file exists.

In other words, the wave directory and the SPL directory must be at the same level in the database, and the SPL and wave must always be matched one-to-one.

The information in these two directories requires accuracy, and if confused, the program is very likely to malfunction, so the user must be careful when entering this information, because the program only checks to see if the directory exists. .

Next, the SPL file check process S30 will be described with reference to FIG. 5.

This process is to process SPL list and confirm it to the user. The user can check the SPL list by using the arrow keys and confirm that the SPL is entered correctly.

That is, the information is shown in the SPL file (S31), and it is determined whether the SPL file is correct.

Next, the list file creation process S40 will be described with reference to FIG.

This process finally creates a list of wave files that meet the conditions according to the conditions and directory structure entered by the user, and briefly reports the result to the user. The screen shows only the number of wave files according to the list file name and words. Is displayed.

That is, the number of SPL files (i) is set to i = 0 (S41), the information is read, the condition of the voice file of each wave is checked (S42, S43), and the list is made with all the SPL files created. (S44, S45)

If it is smaller than i by determining whether the SPL file is smaller than the number of SPL files (i), the information is displayed on the directory through the monitor.

As described above, the present invention can easily classify a large amount of database recorded according to the environment of a vehicle by software without human hands when classifying a vehicle voice database, and by region, speed, age, and condition. It can effectively classify the constructed voice data in a short time to prevent unnecessary waste of time and manpower, and it is possible to build many sub databases for one main database by varying the classification conditions. There is.

Claims (5)

Condition of sound quality of voice file, condition of judging gender of speaker who spoke each voice, condition of age of speaker who spoke each voice, condition of recording place, condition of driving condition, amplitude of voice waveform A wave condition determination process for determining a condition according to a condition and a dialect region according to the present invention; A directory initialization process for inputting information about an SPL directory and a wave directory; Checking the SPL file by listing the SPL to the user; A process of creating a list file which makes a list of wave files matching the conditions according to the conditions and directory structure entered by the user; Wave file automatic classification method of the vehicle voice recognizer, characterized in that to perform sequentially. The method of claim 1, wherein the wave condition determination process, Initializing the condition value, Correcting the condition value and viewing the corrected value through a monitor; Determining whether there is a correction, and if there is no correction, returning the correction and performing correction again. The method of claim 1, wherein the directory initialization process, Reading directory information from the initialization file, Determining whether to read the directory as a file or input by hand, in the case of a file, read the directory information as a file, or enter the directory by hand, Determining whether the directory information is correct, and if the directory information is correct, the step of showing through the monitor characterized in that the wave file automatic classification method of the vehicle voice recognizer. According to claim 1, SPL file check process, Displaying the information in the SPL file, Determining whether the SPL file is correct, if it is not correct, it is terminated as it is, and if it is correct, the wave file automatic classification method of the vehicle voice recognizer, characterized in that the step. The method of claim 1, wherein the list file creation process Setting the number of SPL files (i) to 0, Reading the information and checking the condition of the voice file of each wave; Create a list and keep a list of all the SPL files, Determining whether the SPL file is smaller than the number of SPL files (i), and when the size of the SPL file is smaller than i, displaying directory information through a monitor.