KR101122591B1 - Apparatus and method for speech recognition by keyword recognition - Google Patents

Abstract

PURPOSE: A voice recognition apparatus and a method thereof are provided to analyze each section by detecting a keyword part, above-mentioned, and a later-mentioned part from voice data. CONSTITUTION: A keyword extraction unit(110) determines keyword recognition parts from voice data. A keyword recognition unit(120) outputs a keyword recognition result. A compensation section decision unit(130) determines an above-mentioned part and a later-mentioned part from the keyword recognition part. A recognition compensation unit(140) outputs an addition recognition result by comparison between stored data on the above-mentioned part and data on the later-mentioned part. A voice recognition unit(150) outputs the voice recognition result.

Description

Apparatus and method for speech recognition by keyword recognition

The present invention relates to an apparatus and method for recognizing speech by keyword recognition, and more particularly, to an apparatus and method for determining a speaker's intention by analyzing a keyword extracted from input voice data.

Automated Answering Service (ARS) using speech recognition technology is a service that responds to a user according to the presented announcement and analyzes and processes the user's response, and is currently widely used in various fields.

However, the conventional speech recognition technology has a problem that the speech recognition is successfully performed only when the correct term registered in the database is input, and the recognition rate decreases when the user inputs a speech signal that does not match the grammar. In addition, it is difficult to properly divide and analyze the long voice data input by the user, so that when the user inputs a long sentence, the user's intention cannot be accurately determined.

Korean Patent No. 0933895 describes a configuration in which the input speech data is divided into a plurality of pieces to improve the speech recognition rate, and the speech recognition is performed in parallel on the divided speech data. However, even in this case, when the length of the voice data is long, there is a problem that it is difficult to accurately distinguish the portion corresponding to the keyword and other portions, so that the user's intention cannot be easily understood.

In addition, Korean Patent No. 1021215 describes a parallel speech recognition apparatus and method by speech segmentation, but this method is also proposed to improve the speed of speech recognition and is a method for improving the accuracy of speech recognition. Can't. Therefore, there is still a problem in that it is necessary to accurately grasp the intention of the user and perform the corresponding process.

In order to solve the above problem, there is a need for a method capable of accurately extracting a keyword and processing it according to a user's intention even when a long voice signal that does not conform to a grammar is input from the user.

The technical problem to be achieved by the present invention, when performing the speech recognition processing the speech data input from the user to extract the keyword section and the additional speech section to complement the speech recognition from the long speech data to determine the user's intention The present invention provides a speech recognition apparatus and method using keyword recognition.

Another technical problem to be solved by the present invention is to extract the user's intention by extracting the keyword section and the additional speech section to complement the speech recognition from the long speech data when performing the speech recognition processing the speech data input from the user. The present invention provides a computer-readable recording medium having recorded thereon a program for executing a voice recognition method by recognizing a keyword.

In order to achieve the above technical problem, the apparatus for recognizing speech by keyword recognition according to the present invention includes a keyword extracting unit for determining a plurality of keyword recognition sections which are data of minimum length for extracting keywords from speech data input from a user. ; A keyword recognition unit for outputting a keyword recognition result by comparing the voice data corresponding to each keyword recognition section with keyword data previously stored in a database; A complementary section determination unit determining a keyword recognition section located before the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, and a keyword recognition section located after the keyword section, as described later; A recognition complementary unit configured to output additional recognition results by comparing the voice data corresponding to the tactic unit and the following unit with the tactic unit data and the following unit data previously stored in the database; And a speech recognition unit for combining the keyword recognition result and the additional recognition result to output a final speech recognition result indicating the user's intention.

In order to achieve the above technical problem, the method of speech recognition by keyword recognition according to the present invention comprises: (a) determining a plurality of keyword recognition sections which are data of minimum length for extracting keywords from speech data input from a user; step; (b) outputting a keyword recognition result by comparing the voice data corresponding to each keyword recognition section with keyword data previously stored in a database; (c) determining, by the tactic unit, a keyword recognition section located before the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, and determining a keyword recognition section located after the keyword section, as described later; (d) outputting additional recognition results by comparing the voice data corresponding to the tactic unit and the below-described unit with the tactic unit data and the below-described section data previously stored in the database; And (e) combining the keyword recognition result and the additional recognition result to output a final speech recognition result indicating the user's intention.

According to the apparatus and method for recognizing a speech by keyword recognition according to the present invention, a tactical section and a later section for assisting in a speech recognition process as well as a keyword section in which a user's intention is best expressed from the speech data for analyzing speech data. By detecting the and analyzing the respective sections it is possible to increase the accuracy of speech recognition to perform the operation in accordance with the user's intention.

1 is a block diagram showing the configuration of a preferred embodiment of a speech recognition apparatus using keyword recognition according to the present invention;
2 is a diagram showing the size of voice data over time for performing automatic gain adjustment;
3 is a diagram illustrating a start point, an end point, and a break point between syllable data;
4 is a diagram for explaining a process of dividing speech data including noise;
FIG. 5 is a diagram illustrating a form of voice data input by a user uttering a sentence, 'Um ... find a brunch restaurant near COEX', and,
6 is a flowchart illustrating a process of performing a preferred embodiment of a speech recognition method using keyword recognition according to the present invention.

Hereinafter, exemplary embodiments of a speech recognition apparatus and method using keyword recognition according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing a configuration of a preferred embodiment of a speech recognition apparatus using key word recognition according to the present invention.

Referring to FIG. 1, in the speech recognition apparatus according to the present invention, the keyword extractor 110, the keyword recognizer 120, the supplementary section determiner 130, the recognition complementer 140, and the speech recognizer 150 are provided. And a voice processing unit 160.

The speech recognition apparatus according to the present invention may be applied to the above-mentioned answering service and used to perform a corresponding process according to a voice input by a user, but is not limited thereto and is applicable to all fields requiring a speech recognition technology. It is possible.

First, when a user inputs a voice signal in the form of an analog signal through a microphone, the input voice signal is converted into a digital signal by an analog-to-digital converter, in which the sampling frequency is 8 kHz or 16 kHz. The resolution of the sample converted into the digital signal can be 16 bits. Next, voice data in the form of digital data is obtained through the encoding process for the digital signal. This process is performed on the voice signal inputted by the user by one utterance, and the 'voice data' described below means 'digital data' in which the above process is performed.

The keyword extracting unit 110 determines a plurality of keyword recognition sections that are data of minimum length for extracting keywords from the voice data input from the user.

In this case, the keyword recognition section may be determined based on a result of dividing the speech data into a plurality of word data corresponding to the spacing unit of the sentence and a result of dividing each word data into a plurality of syllable data corresponding to the minimum unit of the word. . The method of dividing the word data and the syllable data will be described later in detail.

Before dividing the voice data into the word data and the syllable data, a process of converting the size of the voice data in the form of digital data to be within a preset size range may be performed. Also, a process of removing noise included in voice data may be performed for accurate voice recognition. This process may be performed by a speech enhancement module (not shown) separately provided in the speech recognition apparatus according to the present invention.

The voice signal input through the microphone may vary in size depending on the system characteristics and the equipment and usage used for the voice signal input. As described above, if the voice signal has a different size every time the voice signal is input, accurate voice recognition may be difficult.

In order to solve this problem, the speech recognition apparatus according to the present invention first converts the gain of the speech data into digital data at the most suitable level for speech recognition. This process is called Automatic Gain Control (AGC).

2 is a diagram illustrating the size of voice data over time for performing automatic gain adjustment. In FIG. 2, a dotted line denoted by 3 is a line indicating a compression threshold value of the size range of the voice data which is preset for speech recognition. When the size of the voice data becomes larger than the upper limit value, it is attenuated so as to fall within the size range. The time taken for this is the attack time indicated by ①. In addition, the section denoted by (2) indicates a release time taken for the input audio data larger than the upper limit to be restored to its original size after being attenuated. Also, a compression slope that attenuates voice data larger than the upper limit may be set in advance.

In FIG. 2, only the case where the size of the voice data is larger than the upper limit of the size range is shown. However, when the size of the voice data becomes too small and smaller than the expander threshold of the size range, it is amplified. The expander slope of the signal may also be set in advance.

As a result, the speech data adjusted to a value within the size range by the speech recognition apparatus according to the present invention can prevent the recognition of the speech because the influence of the microphone, the system or the user environment is excluded.

Next, a process of removing noise, that is, noise components, from the voice data on which AGC is performed is performed. Since the noise component included in the voice data may cause the success rate of the speech recognition to be reduced, it is necessary to remove the noise component before the speech recognition process. Noise reduction may be performed using existing noise reduction algorithms, and the following method may be used as an example.

The speech recognition apparatus according to the present invention extracts a feature by dividing the speech data equally at a predetermined division interval and converting the speech data into a frequency domain. The division interval may be set to one of 5 to 20 msec. Next, the extracted feature is used to determine the region where the speech component is the main region and the noise component the region by the omnidirectional search technique. After the determination of each region is completed, the noise component to be removed is estimated by using the adaptive forgetting coefficient based on psychoacoustic.

The speech recognition apparatus according to the present invention then uses an improved differential spectral subtraction method to remove the estimated noise components, resulting in speech components in the frequency domain from which the noise is removed. Finally, speech data from which noise components are removed is obtained by converting speech components in the frequency domain into the time domain.

The improved speech data obtained by the gain adjustment and noise removal process described above is processed sequentially in the subsequent process and used to extract keywords, which are key words representing the user's intention, from the speech data.

The word division module (not shown) included in the speech recognition apparatus according to the present invention generates a plurality of word data by dividing the improved voice data into word units. The word "word" means each word that constitutes a sentence, which corresponds to the unit of spacing. For example, the sentence "How much is this month" consists of four words: "this time", "month", "price" and "how much". The word division module generates a plurality of word data by dividing the improved speech data into spaces.

The division of word units is started by first dividing the improved speech data into a plurality of frames of the same size. The division of the frame may be performed at the same interval as the above-described division interval, or the voice data may be divided by any one value selected from a range of 5 to 20 msec.

The word division module extracts a feature vector from each frame and generates word data using the extracted feature vector. As the feature vector, power or energy, formant information, or linear predictive coding (LPC) coefficient in the time or frequency domain of the voice data may be used.

The word segmentation module smoothes the feature vector corresponding to each frame by calculating an average of the feature vector corresponding to each frame and the feature vectors of the adjacent frame before applying the feature vector to the generation of the word data. A 'feature vector' used to generate word data described below means a 'smooth feature vector'.

The word segmentation module sequentially inspects a plurality of frames and determines whether a start point of each frame corresponds to a start point or end point of a word. To this end, a threshold value to be compared with a value of the feature vector of each frame is set in advance. A plurality of thresholds may be set according to the applied frame.

For example, with respect to the nth frame, the size of the feature vector of the n-1th frame is equal to or less than the first threshold, and the size of the feature vector of the nth, n + 1th, and n + 3th frames is the second threshold, respectively. If the value is greater than or equal to the third threshold value and the fourth threshold value, the word division module determines the start point of the nth frame as a word start point candidate. That is, the nth frame becomes the first frame of one word data among a plurality of word data obtained from the voice data. In this case, the magnitudes of the first to fourth threshold values may be empirically set.

When the word start point candidate is determined, the word segmentation module sequentially checks subsequent frames to determine the word end point candidate. For example, when the size of the feature vector of the n + 1th frame is less than or equal to the fifth threshold, the start point of the n + 1th frame is determined as a word end point candidate. Since the start point of the nth frame is determined as a word start point candidate, the nth frame and the n + 1th frame constitute one word data.

Alternatively, even when the n + 1th frame satisfies Equation 1 below, the start point of the n + 1th frame may be determined as a word end point candidate.

Here, E _smt () represents the magnitude of the feature vector of the frame, and i values are 0 and 1.

When the above-described process is performed for the entire frame of speech data, a word data candidate from a word start point candidate to a word end point candidate is obtained. This is called a primary word data candidate. A plurality of primary word data candidates can be obtained from one voice data. That is, if the first word start point candidate, the first word end point candidate, the second word start point candidate, and the second word end point candidate are sequentially determined for the frames of the voice data, two primary word data candidates are obtained.

When the first word data candidates are determined, the word segmentation module removes candidates that cannot be considered to be meaningful phrases because the length is too short among the first word data candidates. That is, the word segmentation module selects only the first word data candidate having a predetermined first reference length or more to determine the second word data candidate, and removes the other first word data candidate from the keyword extraction target. The first reference length may be set within the range of 150 to 170 msec.

After the second word data candidate is determined, the word segmentation module merges two adjacent second word data candidates if the length between adjacent second word data candidates among the second word data candidates is smaller than a preset second reference length. do. For example, if the length between the end point of the k th second word data candidate and the start point of the k + 1 th second word data candidate is less than the second reference length, the k th second word data candidate and the k + 1 th second One secondary word data candidate consisting of the word data candidates is determined. In this case, the second reference length may be set to a value within a range of 200 to 250 msec.

After the plurality of secondary word data candidates are sequentially scanned and the merge process described above is completed, secondary word data candidates having a predetermined third reference length or more among non-merged secondary word data candidates and merged secondary word data candidates are preset. Is determined as the third word data candidate. The third reference length may be set to 400 msec.

Finally, the word segmentation module considers the size of the feature vector of the second word data candidate to determine the final word data among the third word data candidates.

First, the word segmentation module calculates an average size of a feature vector of frames included in the second word data candidate for each second word data candidate, and calculates a maximum average size among the calculated average sizes. The word division module then scans each secondary word data candidate and determines whether it can be determined as the final word data.

Specifically, the maximum mean size is greater than or equal to a preset multiple of a preset multiple of a second word data candidate, and the corresponding second word data candidate corresponds to the third word data candidate or the third word data candidate. If included, the word division module determines the third word data candidate including the corresponding second word data candidate as the final word data. The size of the reference fold is set in advance by experiment, and may be set to a value between 2.5 and 4.

In summary, the word segmentation module divides the speech data into frames having the same size, determines the first word data candidates based on the size of the feature vector of each frame, and removes candidates having too short length from the first word data candidates. To determine the second word data candidate. Next, candidates having a short gap between the second word data candidates are merged with each other, and a second word data candidate having a predetermined length or more is determined as the third word data candidate. Finally, the final word data is determined among the third word data candidates by considering the size of the feature vector of the frames constituting the second word data candidate.

The speech recognition apparatus according to the present invention divides the plurality of final word data thus obtained into a plurality of syllable data, and this process may be performed by a syllable division module (not shown) separately provided. 'Syllable' is a basic phonetic unit of a word or word that is intuitively recognized. It can have only a vowel or one or more consonants before and after the vowel. For example, in the example described above, the word 'rate' consists of three syllables: 'yo', 'gold', and 'silver'. The syllable segmentation module may divide one word data such as 'rate' into three syllable data corresponding to 'yo', 'gold', and 'silver'.

In order to divide the syllable data, the information of the final word data, the smoothed feature vector of the frame constituting each final word data, and the information of the smoothed feature vector for each subband in which the frame is divided may be used. Also, when there are a plurality of final word data, the process of dividing each final word data into syllable data may be performed in parallel. Therefore, the syllable division module may be composed of a plurality of syllable division modules. Hereinafter, a process of dividing one final word data into a plurality of syllable data will be described in detail.

When splitting is performed on the last word data, the start point of the last word data is the start point of the first syllable data candidate of the last word data, and the end point of the last word data is set to the end point of the last syllable data candidate of the last word data. .

Next, the syllable segmentation module calculates power for each DFT-pointer in the frequency domain by performing an N-pointer Discrete Fourier Transform (DFT) on the frames included in the final word data. By grouping into L subbands, a feature vector for each subband is calculated.

Thereafter, the syllable segmentation module sequentially scans each frame of the final word data to determine a start point and an end point candidate of the syllable data. Furthermore, a candidate for a breakpoint, which is a point between syllable data corresponding to different syllable data but the distinction is not clear, may be determined.

3 is a diagram illustrating a start point, an end point, and a break point between syllable data. Referring to FIG. 3, the starting point and the ending point may be easily divided into a time point at which a syllable is started or finished. In the case where a plurality of syllables appear in succession, an end point of a previous syllable and a start point of a next syllable may overlap between syllables, and in this case, a start point candidate of the syllable data is determined. In the case of the division point, as shown in FIG. 3, the voiced sound is continuous when the syllable is uttered, and thus it is difficult to find the point compared with the beginning and end points of the syllable, and a discontinuity point occurs when the division is performed.

The syllable segmentation module compares the k-1th frame and the kth frame of the final word data with the feature vector of the lth subband of the k-1th frame not equal to 0 and the feature vector of the lth subband of the kth frame. Counting the number of times that the size of 0 is 0, and if the number is greater than the preset reference number of times, the start point of the k-th frame is determined as the end point candidate of the specific syllable data. In this case, the number of subbands obtained from each frame may be determined as 6 on a log scale, and the value of the reference frequency may be set to 2 or 3.

The syllable segmentation module also sorts the subbands in descending order by comparing the magnitudes of the smoothed feature vectors of each of the L subbands obtained from each frame of the final word data. As a result, when the order of subbands of the k-1th frame and the order of subbands of the kth frame are compared, the starting point of the kth frame is determined if the number of subbands whose order is larger than the first reference number that is set in advance. The syllable data is determined as a candidate for breaking point. Here, the first reference number may be set to 4 or 5.

As another method of segmenting syllable data, the syllable segmentation module extracts the rate of change (increase or decrease) of the smoothed feature vector of each frame of the final syllable data and processes it according to the following four different cases.

In the first case, if the rate of change of the k-th frame is greater than 0 and the size of the smoothed feature vector of the k-1th frame is greater than the size of the smoothed feature vector of the kth frame, the syllable segmentation module starts the k-th frame. Is determined as an end point candidate of the syllable data, and the rate of change of the size of the k-th frame is set to -1.

In the second case, the smoothed feature vector of the k-th frame is smaller than zero and the smoothed feature vector of the k-1th frame is smaller than the smoothed feature vector of the kth frame. If the size is larger than the sixth preset threshold, the start point of the k-th frame is determined as the start point candidate of the syllable data, and the rate of change of the k-th frame is set to one. The sixth threshold value is obtained by averaging the feature vectors of the entire frames, extracting frames having a smaller feature size than the average from the frames, calculating their average feature vectors, and then calculating the average feature vectors within the range of 1.3 to 1.7 times the calculated value. Can be set at

In the third case, if the rate of change of the k-th frame is 0 and the size of the smoothed feature vector of the k-th frame is larger than the sixth threshold, the start point of the k-th frame is determined as a start point candidate of the syllable data. Set the rate of change of size to 1.

In the fourth case, if the size change rate of the kth frame is not 0 and the size of the smoothed feature vector of the kth frame is 0, the start point of the kth frame is determined as a candidate for the end point of the syllable data, and the size change rate of the kth frame is determined. Set to zero.

The syllable segmentation module determines the final syllable data while sequentially scanning the syllable start point candidate, the end point candidate, and the break point candidate determined by the above-described process. Specifically, the breakpoint candidates determined within the first reference distance preset before and after each end point candidate are removed, and the breakpoint candidates determined within the second reference distance preset before and after each start point candidate are also removed. Here, the first reference distance and the second reference distance may be set within the range of 80 to 120 msec and 160 to 200 msec, respectively.

When unnecessary breakpoint candidates are removed by the above process, the syllable segmentation module determines the data between the syllable starting point candidates and the syllable ending point candidates as the final syllable data. A plurality of such final syllable data is obtained from one final word data. In addition, as described above, since the partitioning process is performed in parallel on the plurality of final word data, the processing time can be shortened.

4 is a diagram for describing a process of dividing speech data including noise. FIG. 4A is a form of voice data first input from a user, and includes a lot of noise. AGC and noise reduction are performed on the voice data to obtain voice data in the form of (b), and the data in the form of (c) is obtained by dividing the voice data. A detailed process of generating final word data and final syllable data is omitted in FIG. 4.

As described above, when a plurality of final word data and a plurality of final syllable data are determined from the voice data, the keyword extractor 110 may extract the voice data based on the information of the last word data and the final syllable data. It is divided into a plurality of keyword recognition sections, which are data lengths.

In this case, the voice data to be divided into keyword recognition sections may be voice data performed only with AGC for the first voice data, and the voice data performed until the noise removal process is supplemented to compensate for the degradation of the keyword recognition rate due to noise. Can be used as

First, the keyword extractor 110 determines that the final word data as one keyword recognition section when the number of the final syllable data divided from each last word data is equal to or less than a second preset reference number. In addition, when the number of the final syllable data divided from each last word data is larger than the second reference number, the adjacent syllable data is merged while the final syllable data is scanned to determine one keyword recognition section.

For example, if the final word data is data corresponding to 'Olympic Park' and the length of the keyword recognition section is three syllables, the keyword extractor 110 scans the final word data from the first syllable data by three syllables. Determine the recognition section. In other words, three keyword recognition sections of 'Olympic', 'Lympic Ball' and 'Pick Park' are determined from the final word data of 'Olympic Park'. Here, the scan unit of the syllable data and the length of the keyword recognition section may be determined as an appropriate value in consideration of characteristics of a system to which the speech recognition apparatus according to the present invention is applied and length information of frequently input keywords.

Also, if the number of the final syllable data included in one final word data is greater than the preset third reference number, the keyword recognition section determined in the same order as the third reference number from the final word data is not included in the other keyword recognition section. Include all remaining final syllable data.

When the keyword recognition sections are determined as described above, in order to improve the recognition rate of the keywords, the silent section data having a predetermined length is added before and after each keyword recognition section so that the keywords can be extracted from the middle point of the section. The length of the silence section may be set in advance to a value between 100 and 500 msec. In addition, as the data used as the silent section, data of the silent section included in the actual voice data may be used.

On the other hand, when the start point and the end point of the keyword recognition section are the break points of the syllable data rather than the start point and the end point of the syllable data, the discontinuity point of the signal may cause a problem that the feature vector is distorted. In the section corresponding to the division point of, a discontinuous point of the signal may be removed by applying a hanning window.

When the number of keyword recognition sections determined by the above-described process is equal to or greater than a preset maximum number, the keyword extractor 110 or a separately provided control module (not shown) outputs an error message and inputs a new voice message from the user. I can receive it.

The case where the keyword recognition section of the maximum number or more is determined from the voice data corresponds to a case in which the user fails to input a sentence clearly indicating his or her intention. For example, if a lot of meaningless phrases such as interjection are included in a sentence, the number of keyword recognition sections is increased, resulting in a large amount of computation and a long processing time for speech recognition.

Therefore, the keyword extractor 110 outputs an error message to the user when the keyword recognition interval of the maximum number or more is determined from the voice data so that the user inputs a simple sentence including the correct keyword portion.

Next, the keyword recognition unit 120 compares the voice data corresponding to each keyword recognition section determined by the keyword extraction unit 110 with keyword information previously stored in the keyword database 170 and outputs a keyword recognition result. In this case, the voice data used for keyword recognition is voice data performed only by AGC, and the voice data from which the noise is removed may be used as an auxiliary means in case the recognition result is not good.

The keyword database 170 used for keyword recognition stores various keywords determined according to characteristics of a system to which the speech recognition apparatus according to the present invention is applied and types of keywords mainly input by a user. In addition, in order to improve the success rate of keyword recognition, keywords written in a grammar of a Backus-Naur Form (BNF) may be stored in the keyword database 170.

As in the case of dividing each final word data into a plurality of final syllable data, a keyword recognition process for each keyword recognition section may be performed in parallel. When the keyword recognition unit 120 is configured with a plurality of speech recognition modules to perform keyword recognition for a plurality of keyword recognition sections in parallel, each speech recognition module may refer to the same keyword database 170. Hereinafter, a case of recognizing a keyword included in one keyword recognition section will be described.

The keyword recognition unit 120 compares the voice data corresponding to the keyword recognition section with the keyword data stored in the keyword database 170 to determine a plurality of keyword candidates, and determines one of the keyword candidates as the final recognition result.

In detail, if the reliability value of the keyword candidate having the highest reliability among the plurality of keyword candidates is equal to or greater than a preset seventh threshold value, the keyword recognition unit 120 determines the keyword candidate as the final recognition result. In addition, if the confidence value of the keyword candidate having the highest reliability is smaller than the seventh threshold value, the keyword candidates having the confidence value greater than the eighth threshold value set in advance are presented to the user so that the user can directly select the final recognition result.

If the reliability value of the keyword candidate having the highest reliability is smaller than the eighth threshold, the keyword database 170 is configured to display data corresponding to the keyword recognition interval in the speech data from which the noise is removed, assuming that the feature vector is affected by the noise component. The recognition result is output against the keyword data stored in the.

If the reliability value of the keyword candidate having the highest reliability among the keyword candidates obtained as a result of using the noise-free speech data is smaller than the eighth threshold, it is determined that keyword recognition has failed.

The seventh and eighth threshold values may be determined through experiments. Preferably, the seventh threshold value may be determined to be 47 to 60%, and the eighth threshold value may be determined to be 40 to 47%. In addition, the number of keyword candidates that may be determined for one keyword recognition interval may be determined as 3 or 4.

FIG. 5 is a diagram illustrating a form of voice data input by a user uttering a sentence, 'Um ... find a brunch restaurant near COEX'.

In the voice data of FIG. 5, since the “um ...” portion is short, it is removed in the process of dividing word data. Accordingly, the final word data obtained from the voice data of FIG. 5 is five of 'coex', 'nearby', 'brunch', 'restaurant', and 'find'. The number of final syllable data obtained from each final word data is three or two.

When the value of the second reference number set to determine the keyword recognition section is 4 and the number of the final syllable data included in one final word data is less than or equal to the second reference number, the final syllable data of the final syllable data may be determined. When the scan unit is assumed to be one syllable and the size of the keyword recognition section is set to three syllables, the final word data divided from the speech data of FIG. 5 satisfies the condition for determining the keyword recognition section.

The keyword recognition unit 120 performs keyword recognition by comparing the five keyword recognition sections determined above with keyword data stored in the keyword database 170, respectively. When the keyword recognition unit 120 includes a plurality of speech recognition modules, each of the speech recognition modules may simultaneously perform keyword recognition for each of the keyword recognition sections.

If 'brunch' and 'delicious' are included in the keyword data stored in the keyword database 170, 'brunch' and 'delicious' are selected from the keyword database 170 among the five keyword recognition sections determined from the voice data of FIG. Since it matches the stored keyword data, it is output as a keyword recognition result.

In the above description, the determination of the keyword recognition section and the keyword recognition are performed using speech data on which AGC is performed, and the speech data on which noise removal is performed is used as an auxiliary means, but the recognition rate and recognition speed are improved. In order to achieve this, all of the keyword recognition sections determined from the speech data on which AGC has been performed and the speech data on which noise has been removed may be input to a plurality of speech recognition modules and processed in parallel.

When a portion corresponding to a keyword in the voice data is determined by the keyword recognizer 120, the supplemental section determiner 130 uses the keyword recognition section previously located in the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, as the tactic unit. The keyword recognition section positioned after the keyword section is determined later.

In the conventional speech recognition method, only keywords, which are key words representing the user's intention, are recognized and processed accordingly. Therefore, the voice recognition result may be obtained differently from the user's intention, and thus the accuracy of the voice recognition is required.

The speech recognition apparatus according to the present invention includes a supplementary section determiner 130 to supplement the keyword recognition result obtained by the keyword recognizer 120 to improve the accuracy of speech recognition. Decide In other words, the voice recognition is performed by further selecting a keyword recognition section corresponding to the keyword recognition result from the voice data, that is, a keyword recognition section located before the keyword section and a keyword recognition section located after the keyword section, thereby accurately determining the user's intention. To make it possible.

For example, since the recognition result of 'brunch' and 'delicious' is output as the keyword recognition result among the five keyword recognition sections determined from the voice data of FIG. 5, 'brunch' and 'delicious' are not included in the voice data of FIG. 5. Corresponds to the keyword section. The complementary section determination unit 130 determines the 'coex' and 'near', which are the keyword recognition sections previously located in the keyword section, as the tactic unit, and determine the 'find', which is a keyword recognition section located after the keyword section, as described later. .

In this case, the number of the tactical part and the hereinafter part determined from the voice data may be set in advance, and only the keyword recognition section adjacent to the front and back of the keyword part may be determined as the tactical part and the hereinafter. In addition, when the keyword portion is located at the beginning or the end of the sentence, the above-described portion or the later portion may not be determined.

The recognition compensator 140 compares the voice data corresponding to the tactical part and the hereinafter part with the tactical part data and the later part data previously stored in the tactical part database 180 and the later part database 190 to obtain additional recognition results. Output If only one of the tactical part or the hereinafter part is determined in the voice data, only the speech recognition of the tactical part or the hereinafter part is performed by the recognition complementer 140.

The tactical database 180 and the hereinafter described database 190 are provided separately from the keyword database 170, and the method of constructing the database is the same as the keyword database 170. That is, words that can be frequently used at the front of the sentence are stored in the tactics database 180, and words that can be frequently used at the rear of the sentence are stored in the database 190 below.

Thus, by separating the tactics section database 180 and the following section database 190 can shorten the time for the speech recognition performed in addition to the keyword recognition, the accuracy by using additional tactics section and the later section for speech recognition The speech recognition time can be prevented from being excessively increased while obtaining the effect of improving.

The tactical part recognition process and the later part recognition process by the recognition complementer 140 may be performed in parallel by a plurality of voice recognition modules, respectively. In this case, the voice recognition module for performing the tactical part recognition and the later part recognition may be the same as the plurality of voice recognition modules constituting the keyword recognizer 120. That is, the speech recognition apparatus according to the present invention may be provided with a plurality of speech recognition modules to adaptively perform the functions of the keyword recognition unit 120 or the recognition supplement unit 140.

Accordingly, each voice recognition module selects an appropriate database according to the characteristics of the keyword recognition section, which is a recognition target, to perform voice recognition. In case of performing tactical recognition, refer to the tactical database 180 and refer to the later recognition. In the case of doing so, refer to the database 190 described later. Specific tactic part recognition and a later part recognition process are the same as the keyword recognition method by the keyword recognition unit 120, and thus a detailed description thereof will be omitted.

Meanwhile, the tactical part recognition module that performs the tactical part recognition process among the plurality of speech recognition modules operating as the recognition complementer 140 may be assisted if the tactical part data corresponding to the tactical part does not exist in the tactical part database 180. As a means, data corresponding to the tactical part can be retrieved from the hereinafter referred to database 190. Similarly, the below-described part recognition module that performs the below-described part recognition process may also search for the data corresponding to the later-described part from the above-described part database 180 if there is no later-described part data corresponding to the later-described part in the later-described part database 190. .

The operation of the speech recognition module is because the structure of the sentences input by all users is not always constant, and the phrase used in the front of the sentence may generally be included in the back and vice versa. For example, in the voice data of FIG. 5, when the tactical part data corresponding to 'near' among 'coex' and 'near' of the tactical part is not retrieved from the tactical part database 180, the recognition complementer 140 will be described later. The sub-data 190 may retrieve later-part data corresponding to "near".

Even if the search results are not obtained in both the tactics database 180 and the later database 190 according to the above-described process, additional recognition results for the corresponding tactic or later section are not output.

The speech recognition unit 150 outputs a final speech recognition result indicating the user's intention based on the keyword recognition result and the additional recognition result.

For example, when only the recognition result of 'Brunch', which is a keyword part, 'COEX', and 'Find', which is described later, is obtained from the voice data of FIG. 5, the voice recognition unit 150 recognizes a keyword about 'Brunch'. The result is combined with the additional recognition result of 'COEX' and 'Find', and outputs a result such as 'Find Coex Brunch' as the final voice recognition result of the voice data input by the user.

The speech recognition apparatus according to the present invention may not only output the final speech recognition result for the voice data input by the user, but may also grasp the user's intention from the recognition result and perform an operation corresponding thereto.

The voice processor 160 performs a pre-stored operation in response to the keyword recognition result and the additional recognition result included in the final speech recognition result.

For example, since the final voice recognition result output from the voice recognition unit 150 with respect to the voice data of FIG. 5 is 'Find a COEX brunch,' the voice processing unit 160 may store an operation according to a user's command in advance. The database 200 searches for an operation corresponding to the final speech recognition result. When the operation corresponding to the final speech recognition result is searched as 'output of the brunch restaurant list near COEX' from the search result scenario database 200, the voice processing unit 160 lists the brunch specialty restaurants located near the COEX among the stored restaurant lists. To the user.

In this case, in order to quickly perform an operation according to a user's intention, the scenario database 200 may separately store operations corresponding to a keyword recognition result, a tactical part recognition result, and a later part recognition result among the final speech recognition results. . Therefore, when the final speech recognition result is obtained, the speech processing unit 160 may operate according to the scenario obtained by searching for the scenario corresponding to the keyword recognition result and the scenario corresponding to the tactic part and the later part recognition results, respectively.

6 is a flowchart illustrating a process of performing a preferred embodiment of a speech recognition method using keyword recognition according to the present invention.

Referring to FIG. 6, the keyword extractor 110 determines a plurality of keyword recognition sections that are data of minimum length for extracting a keyword from voice data input from a user (S1010), and the keyword recognizer 120 determines that The keyword recognition result is output by comparing the voice data corresponding to each keyword recognition section with the keyword data previously stored in the database, that is, the keyword database 170 (S1020).

Before the keyword recognition section is determined, automatic gain adjustment and noise removal of the speech data may be performed as a preprocessing process for the speech data, and the speech data may be divided into a plurality of word data and syllable data. . Since the specific processing for this has been described above, a detailed description thereof will be omitted.

The complementary section determination unit 130 determines the keyword recognition section located before the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, as the tactic unit, and determines a keyword recognition section located after the keyword section to be described later (S1030). . The recognition complementer 140 further recognizes the voice data corresponding to the tactical part and the hereinafter part by comparing the tactical part data and the hereinafter part data respectively stored in the tactical part database 180 and the later part database 190 respectively. The result is output (S1040).

The speech recognizer 150 combines the keyword recognition result and the additional recognition result and outputs a final speech recognition result indicating the user's intention (S1050).

Finally, the voice processor 160 performs an operation previously stored in the scenario database 200 in response to the keyword recognition result and the additional recognition result included in the final speech recognition result (S1060).

The present invention can also be embodied as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium includes all kinds of recording devices in which data that can be read by a computer system is stored. Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, and the like, and may be implemented in the form of a carrier wave (for example, transmission via the Internet) . The computer readable recording medium can also be distributed over network coupled computer systems so that the computer readable code is stored and executed in a distributed fashion.

Although the preferred embodiments of the present invention have been shown and described above, the present invention is not limited to the specific preferred embodiments described above, and the present invention belongs to the present invention without departing from the gist of the present invention as claimed in the claims. Various modifications can be made by those skilled in the art, and such changes are within the scope of the claims.

110-Keyword Extractor
120-Keyword Recognition
130-complementary interval determination unit
140-Awareness Complement
150-Speech Recognition
160-voice processing unit
170-Keyword database
180-Tactical Database
190-Database below
200-scenario database

Claims (15)

A keyword extraction unit that determines a plurality of keyword recognition sections that are data of minimum length for extracting keywords from voice data received from a user;
A keyword recognition unit for outputting a keyword recognition result by comparing the voice data corresponding to each keyword recognition section with keyword data previously stored in a database;
A complementary section determination unit determining a keyword recognition section located before the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, and a keyword recognition section located after the keyword section, as described later;
A recognition complementary unit configured to output additional recognition results by comparing the voice data corresponding to the tactic unit and the following unit with the tactic unit data and the following unit data previously stored in the database; And
And a speech recognition unit configured to combine the keyword recognition result and the additional recognition result to output a final speech recognition result indicating the user's intention. The method of claim 1,
And a speech processor configured to perform a pre-stored operation in response to the keyword recognition result and the additional recognition result included in the final speech recognition result. 3. The method according to claim 1 or 2,
The keyword extractor is configured to divide the speech data into a plurality of word data corresponding to a spacing unit of a sentence and to divide the word data into a plurality of syllable data corresponding to a minimum unit of a word. And a keyword recognition section. 3. The method according to claim 1 or 2,
And the keyword extracting unit outputs an error message to the user and receives new voice data from the user when the number of the keyword recognition sections determined from the voice data is greater than or equal to a preset maximum number. 3. The method according to claim 1 or 2,
The keyword recognizer determines a plurality of keyword candidates by comparing the voice data corresponding to each keyword recognition section with the keyword data, and determines a keyword candidate having the highest reliability as the keyword recognition result. Recognition device. 3. The method according to claim 1 or 2,
The recognition compensator outputs the additional recognition result in comparison with the voice data corresponding to the tactic part and the later data if the additional recognition result for the voice data corresponding to the tactic part is not retrieved from the tactic part database. And outputting the additional recognition result in contrast to the voice data corresponding to the later-described part and the tactical-part data if the additional recognition result of the voice data corresponding to the later-described part is not retrieved from the later-described part database. Speech recognition device characterized in that. 3. The method according to claim 1 or 2,
And the recognition complementary unit performs speech recognition on the tactic unit and the following unit in parallel by a plurality of voice recognition modules corresponding to the tactic unit and the following unit, respectively. (a) determining a plurality of keyword recognition sections, the minimum length of data for extracting keywords from voice data input from a user;
(b) outputting a keyword recognition result by comparing the voice data corresponding to each keyword recognition section with keyword data previously stored in a database;
(c) determining, by the tactic unit, a keyword recognition section located before the keyword section, which is a keyword recognition section corresponding to the keyword recognition result, and determining a keyword recognition section located after the keyword section, as described later;
(d) outputting additional recognition results by comparing the voice data corresponding to the tactic unit and the below-described unit with the tactic unit data and the below-described section data previously stored in the database; And
and (e) combining the keyword recognition result and the additional recognition result and outputting a final speech recognition result indicating the user's intention. The method of claim 8,
and performing a pre-stored operation in response to the keyword recognition result and the additional recognition result included in the final speech recognition result. The method according to claim 8 or 9,
In the step (a), based on a result of dividing the speech data into a plurality of word data corresponding to the spacing unit of the sentence and a result of dividing each word data into a plurality of syllable data corresponding to the minimum unit of a word And a plurality of keyword recognition sections. The method according to claim 8 or 9,
In the step (a), if the number of the keyword recognition interval determined from the voice data is greater than or equal to a preset maximum number of voice recognition method characterized in that for outputting an error message to the user to receive new voice data from the user . The method according to claim 8 or 9,
In the step (b), a plurality of keyword candidates are determined by comparing the voice data corresponding to each keyword recognition section with the keyword data, and a keyword candidate having the highest reliability is determined as the keyword recognition result. Voice recognition method. The method according to claim 8 or 9,
In the step (d), if the additional recognition result for the voice data corresponding to the tactic part is not retrieved from the tactic part database, the additional recognition is performed in comparison with the voice data corresponding to the tactic part and the data to be described later. Outputs a result, and if the additional recognition result of the voice data corresponding to the below-mentioned part is not retrieved from the below-described part database, the additional recognition result is compared with the voice data corresponding to the below-mentioned part and the tactical part data. Outputting the voice recognition method. The method according to claim 8 or 9,
In the step (d), the voice recognition method for performing the voice recognition for the tactical section and the hereinafter described section in parallel by a plurality of speech recognition modules respectively corresponding to the tactical section and the hereinafter described section. A computer-readable recording medium having recorded thereon a program for executing the speech recognition method according to claim 8 or 9.

Images