KR100293465B1 - Speech recognition method - Google Patents

Abstract

PURPOSE: A voice recognition method is provided to draw slant lines having gradients 1 from start lattice points and end lattice points, and to move the slant lines to both sides as many as determined values to set a window, then to use lattice points only between the two slant lines for matching, so as to easily set the window. CONSTITUTION: A voice recognition system generates a 2-dimensional vertical coordinates system having M times N lattice points, to match two sequences. The system detects start lattice points(1, 1) and end lattice points(M, N) of the 2-dimensional vertical coordinates system, and draws slant lines having gradients 1 from the start lattice points(1, 1) and the end lattice points(M, N), then moves the slant lines as many as determined values to define a search section for matching. The system calculates distance between two features of each lattice point of each string within the search section, and selects a path where the distance is the minimum as an optimum path. The system repeats the step of selecting the optimum path for all strings within the search section. The system divides a minimum accumulation distance of the end lattice points(M, N) by a length sum of the two sequences, and calculates a final matching score.

Description

Speech Recognition Method {SPEECH RECOGNITION METHOD}

The present invention relates to a speech recognition method for recognizing a speech signal by comparing a reference pattern of a databaseized speech signal with a test pattern of an input speech signal.

The simplest of the speech recognition techniques is speaker dependent isolated word recognition. It can recognize only the voice of the trained person, and can only recognize the voice spoken in units of words (or short sentences). Many speech recognition algorithms for this purpose are already known and can be divided into speech section detection, feature extraction, and matching.

In this speaker dependent isolated word recognition, the normalization of the test pattern and the reference pattern on the time axis is greatly influenced by the speech recognition rate.

In general, since the time intervals do not coincide between the test pattern and the reference pattern, a time scale is determined through a nonlinear time warping method. The determination of this time scale is realized through the Dynamic Time Warping (DTW) technique, which uses three processes: word boundary detection, nonlinear time alignment, and recognition. Therefore, a recognition error is not caused by an error generated in word boundary and time alignment. In other words, even if the same person utters the same word, it is stretched nonlinearly on the time axis due to the difference in pronunciation speed. Therefore, the similarity between the two patterns is determined by comparing the test pattern with the reference pattern to remove the variation on the time axis. The time base normalization calculation method is DTW.

Various DTW methods have been proposed, one of which measures a spectral distance between a test pattern and a database-based reference pattern and uses a reference pattern having a spectral distance closest to the test pattern as a recognition pattern. There is a way to choose. That is, after dividing the digital signal of each speech pattern into frames, the LPC (Linear Predictor Coefficient) is calculated for each frame, and then converted into a Capstrum coefficient, which is a feature vector of the test pattern, and compared with each of the previously-referenced reference patterns. To measure the spectral distance. However, this DTW method is slow in response and has a problem of requiring a large storage capacity because it is a floating point format.

In order to solve this problem, there is a method of setting a window to limit an optimal path search range and converting the analyzed speech pattern into a feature set of an integer format. At this time, the spectral distance between the test pattern and each reference pattern is measured by comparing the feature set of the test pattern in the integer format with the database-based reference pattern, and the reference pattern having the spectral distance closest to the test pattern is recognized. Select as. However, this DTW method has less storage capacity and faster response time than the floating point format, but by connecting the start point and the end point and setting the window interval whose width is odd by the connected line. In addition, when the number of frames of the test pattern and the reference pattern are different, it is difficult to calculate the slope, and it is difficult to check which grid points are included in the window. There is a problem that the amount of calculation increases.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to draw an oblique line having a slope of 1 from a starting grid point and an end grid point, and divide a predetermined value, that is, the number of frames by an integer multiple of 2. The present invention provides a speech recognition method that does not need a slope calculation and a divider by setting a window section by moving diagonal lines to both sides by a value.

1 is a configuration block diagram for performing a voice recognition method according to the present invention

2 is a coordinate diagram showing a search interval setting state according to the present invention;

3 is a DTW coordinate diagram according to the present invention

Explanation of symbols for main parts of the drawings

11: microphone 12: A / D converter

13 voice section detection unit 14 feature extraction unit

15: reference data memory 16: DTW unit

In order to achieve the above object, a voice recognition method includes comparing a test pattern of an input voice with a feature set of each reference database pre-database and outputting a reference pattern having a feature most similar to the test pattern as a recognized voice. In the speech recognition method, in order to match two sequences of a feature of a test pattern and a feature set of a reference pattern, a two-dimensional two-dimensional grid having MxN points (M and N are the number of frames of the test pattern and the reference pattern). A first step of generating a vertical coordinate system, and after detecting a starting grid point and an ending grid point of the two-dimensional vertical coordinate system, an oblique line having an inclination of 1 is drawn from the detected starting grid point and the ending grid point, and from there, N A second step of determining a search section for matching by moving diagonal lines on both sides by dividing the (number of frames) by an integer multiple of 2; A third step of calculating a distance between two features at each grid point and selecting a path having a minimum distance between the two features as an optimal path, and a fourth step of repeating the third step for all columns of the search interval; And a fifth step of calculating the final matching score by dividing the minimum cumulative distance at the end grid point by the sum of the two sequence lengths when the fourth step is performed.

When the minimum cumulative distance value at each lattice point of each column in the search interval exceeds the range of integers, the maximum integer value is replaced.

Each lattice point of each column in the search interval has a minimum cumulative distance value up to the m th and n th features of the two sequences of the test pattern and the reference pattern.

When the DTW method is used for speech recognition, it is easy to set a window without requiring a tilt calculation, and the calculation speed is small, and the response speed is increased.

Other objects, features and advantages of the present invention will become apparent from the detailed description of the embodiments with reference to the accompanying drawings.

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

FIG. 1 is a block diagram illustrating a speech recognition system for performing a speech recognition method, including a microphone 11, an analog / digital converter (A / D) converter 12, a speech section detector 13, and feature extraction. The unit 14 is composed of a reference data memory 15 and a DTW unit 16.

According to the present invention configured as described above, in order to store the reference pattern of the voice signal, when the voice signal is input through the microphone 11, the A / D converter 12 converts it to a digital signal and then the voice section detector 13 Will output The voice section detector 13 divides the digital voice signal into a short section of a signal (ie, a frame), and then actually uses the energy, zero crossing rate, and time length information of each frame. The starting point and the end point of the spoken speech section are detected. Then, the feature extractor 14 extracts the spectral coefficients of the frame corresponding to the speech section to form a reference pattern, and stores the reference pattern in the reference data memory 15. The reference pattern is repeated in the reference data memory 15 by repeatedly performing the voice signal.

On the other hand, in the case of performing the process of recognizing the spoken voice signal, the voice input through the microphone 11 is the A / D converter 12, the voice section detection unit 13, and the feature extraction unit 14 The start point and the end point of the actually spoken speech section are detected from the inputted signal, and the spectral coefficient of the frame corresponding to the detected speech section is extracted to make a test pattern of the input speech. At this time, the DTW unit 16 compares the test pattern with each reference pattern stored in the reference data memory 15 at least one or more, and outputs a reference pattern having the most similar characteristics to the test pattern as a recognized voice.

That is, in the DTW unit 16, first, the lengths of two sequences to be matched are defined as M and N, respectively, as shown in FIG. At this time, if the larger one of the sequence lengths M and N to be compared is more than two times smaller than the smaller one, there is almost no likelihood. Therefore, a matching distortion is given as a maximum value without performing a recognition process. This is because the input word and the registered word are more than two times different, so there is almost no likelihood, so that the similarity of the recognized word is excluded.

In order to match the two sequences, a two-dimensional vertical coordinate system having M × N grid points is created. At this time, by comparing the frame length of the database-based reference pattern and the frame length of the test pattern, if the pattern having a longer frame is located on the M axis, the calculation can be smoothly performed.

Here, the speech patterns R and T may be expressed as follows as a sequence of feature vectors for feature extraction.

R = [R ₁ , R ₂ , ..., R _m , ..., R _M ]

T = [T ₁ , T ₂ , ..., T _n , ..., T _N ]

The patterns R and T change along the m-axis and the n-axis, respectively, and the difference in the feature vector between the voice patterns is expressed as a series of points C (k) as shown in FIG.

F = C (1), C (2), ..., C (k), ..., C (K)

Here, C (k) = (m (k), n (k)), and F is called a warping function, which is a function of projecting from the time axis of the test pattern to the time axis of the reference pattern.

The DTW unit 16 finds the optimal path m = W (n) that becomes the minimum distance using the warping function.

In order to reduce unnecessary calculation, a window for limiting the optimum path search range is set. In other words, if the same person talks, there is no big change, so the window is set to limit the optimal path search section.

In the present invention, it is easy to set the window without the need for the slope calculation, and the window is determined as follows in order to reduce the amount of calculation and to increase the response speed.

First, an oblique line having a slope 1 is drawn from the starting grid points 1 and 1 and the ending grid points M and N, respectively. From this , Where N is the number of frames, n is a natural number, and n is 2, which is most suitable.) If the diagonal lines are moved to both sides, the grid points between the two diagonal lines are searched for matching. At this time, the width of the window When N is divided by an integer multiple of 2, the integer multiple of 2 is simple and efficient because a shifter can be used without a complicated divider. N may be the number of frames of the test pattern or may be the number of frames of the reference pattern.

In the search window, the grid points m and n have minimum cumulative distance values between the m th and n th features of the two sequences. Here, the feature value is scaled to have an integer value between 0 and 5000.

At this time, in order to avoid excessive compression and elongation in a specific time interval, the limit of the small section pass is limited, and the method of determining the small section path is shown in FIG. 3 as an embodiment. In Figure 3 the warping function moves one step simultaneously from one of the three possible directions. For example, three possible ways to reach grid points (m, n) are to move directly from grid points (m-1, n-1) to grid points (m, n), or grid points (m) There is a method of indirectly moving from -1, n to grid point (m, n) or from grid point (m, n-1) to grid point (m, n).

At this time, each arrival lattice point (m, n) is associated with a weight W _mn such as the Euclidean or captral distance between the m th frame of the pattern R and the n th frame of the pattern T. The weight W _mn is applied to each of the indirect paths and the weight 2W _mn is applied to the direct path, so that the distance between two features at each grid point (m, n) is defined as d _{m, n in} Equation 1 below. .

That is, the distance between each feature is obtained by adding all the difference between the values corresponding to the orders of the two features. At this time, the minimum cumulative distance at the grid point (m, n) is calculated as in Equation 1, and when the value is out of the range of integers, it is replaced by the maximum integer value.

Then, starting from the bottom row, the minimum cumulative distance value is obtained for the grid points within the search range while sequentially moving upward. In order to obtain the minimum accumulated distance value of the current column, the minimum accumulated distance value of the next row is needed.

The final matching score is a value obtained by dividing the minimum cumulative distance at the grid points M and N by the sum of two sequence lengths (M + N).

As described above, according to the speech recognition method of the present invention, an oblique line having a slope 1 is drawn from the starting grid point and the ending grid point, respectively, , N is the number of frames, n is the natural number, and sets the window to be the search section by moving the diagonal lines on both sides, so that only the grid points between the two diagonal lines are used for matching. The grid point check is also easy because the window boundary line is connected to the grid point and the grid point. In addition, the complexity of the divider is eliminated, resulting in less computation and faster response.

Claims (7)

A voice recognition method for comparing a test pattern of an input voice with a feature set of each reference pattern previously databased and outputting a reference pattern having a feature most similar to a test pattern as a recognized voice. To match the two sequences of the feature of the test pattern and the feature set of the reference pattern, generate a two-dimensional vertical coordinate system with M × N grid points (M and N are the number of frames of the test pattern and the number of frames of the reference pattern). With the first step, After detecting the starting grid points (1,1) and the ending grid points (M, N) of the two-dimensional vertical coordinate system, the slopes are respectively determined from the detected starting grid points (1,1) and the ending grid points (M, N). A second step of drawing an oblique line having 1, and moving the oblique line to both sides by a predetermined value therefrom to determine a search section for matching; A third step of calculating a distance between two features at each grid point of each column in the search section and selecting a path having a minimum distance between the two features as an optimal path; A fourth step of repeating the third step with respect to all columns of the search period; When the fourth step is performed, the speech recognition method comprises a fifth step of calculating a final matching score by dividing the minimum cumulative distance at the end grid points (M, N) by the sum (M + N) of the two sequence lengths. Way. The speech recognition method of claim 1, wherein the distance between two features at each grid point is obtained by adding up the difference between the values corresponding to each order of the two features, and the minimum cumulative distance is defined by the following equation. Where D _{m, n} is the minimum accumulation distance at the lattice point (m, n), d _{m, n is} the distance between two features at the grid point (m, n) , a ⁱ _{1, m} : i-th value of the m th characteristic of the first sequence, a ⁱ _{2, n} : i-th value of the n th feature of the second sequence, P: degree of feature. The method of claim 2, Speech recognition method characterized in that the minimum cumulative distance value at each grid point (m, n) is replaced by the maximum integer value when it exceeds the range of integers. 3. The voice of claim 2, wherein each grid point (m, n) of each column in the search interval has a minimum cumulative distance value up to the mth and nth features of two sequences of a test pattern and a reference pattern. Recognition method. 5. The grid point (m, n) of claim 4, wherein each grid point (m, n) of each column in the search interval has a distance value moving directly from the previous grid point (m-1, n-1) and two neighbors of the grid point (m, n). And generating a new pass value repeatedly as a function of at least one of the distance values indirectly moving from the grid points (m-1, n) and (m, n-1). 6. The speech recognition method of claim 5, wherein the minimum cumulative distance value of the immediately below column is stored to obtain a minimum cumulative distance value of the current column. The predetermined value for moving the oblique line to both sides in the second step is (Where N is the number of frames and n is a natural number).