KR20000033276A - Frame compression method using representative characteristic column and audio recognition method using the same - Google Patents

Abstract

PURPOSE: A frame compression method using representative characteristic column and an audio recognition method using the same are provided to reduce the number of frame without degradation of performance by obtaining a representative frame of similar frames after extracting a characteristic column from frames of constant time interval. CONSTITUTION: A frame compression method using representative characteristic column comprises steps of: dividing compression signal by prescribed time interval into a frame; extracting a characteristic column about the divided frame; obtaining the similarity between the extracted characteristic columns; obtaining a representative characteristic column of similar characteristic columns using the obtained similarity.

Description

Frame Compression Method Using Representative Feature Strings and Speech Recognition Method Using the Same

The present invention relates to a frame compression method using a representative feature sequence and a computer readable recording medium having recorded thereon a program for realizing the method.

The present invention also relates to a voice recognition method for processing an input voice in units of frames and to recognize the voice based on the same, and a computer-readable recording medium having recorded thereon a program for realizing the method. The present invention relates to a speech recognition method that reduces the number of frames by using a compression method to improve recognition processing speed, and a computer-readable recording medium that records a program for realizing the method.

In general, a speech recognition system is a device that inputs a human voice and outputs speech content.

Conventional speech recognition systems perform speech recognition by processing features by extracting features at a fixed time interval regardless of the content of speech using a fixed frame rate.

Looking at US Patent No. 4,979,212 (Speech recognition system in which voiced intervals are broken into segments that may have unequal durations) of OKI Electric Industry Co., Ltd., Japan, for use in speech recognition. In extracting the feature string, the length of the speech section to be extracted is determined by changing the power of the speech.

In general, a vowel portion having a large power forms one frame, and a consonant between vowels forms another frame, and a feature string proportional to the number of syllables is extracted.

Accordingly, one syllable is normalized to a feature string having a constant length regardless of the utterance length of the syllable.

As described above, the conventional speech recognition system that extracts feature strings from frames at predetermined time intervals and processes them to perform speech recognition has a problem in that the speed of speech recognition is reduced because a large number of frames are generated every predetermined time. .

In addition, the widely used Hidden Markov Model type speech recognition system has a problem of duration modeling.

The present invention devised to solve the above problems can reduce the number of frames without loss of performance by extracting feature strings from frames at regular time intervals and then obtaining one representative frame (representative feature string) for similar frames. It is an object of the present invention to provide a frame compression method and a computer-readable recording medium having recorded thereon a program for realizing the method.

In addition, the present invention provides a frame compression method in which the length information of each frame is added as a feature by adding the number of actual frames represented by the extracted representative feature string as a feature, and a computer program for recording the program for realizing the method. The purpose is to provide a recording medium that can be used.

In addition, the present invention provides a computer readable recording medium recording a speech recognition method and a program for realizing the speech recognition method using the frame compression method to improve the speech recognition speed and the speech recognition performance (recognition rate). There is another purpose to provide.

1 is an exemplary configuration diagram of a speech recognition system to which the present invention is applied.

2 is a conceptual diagram of a frame compression method using a representative feature sequence according to the present invention;

3 is a flowchart illustrating a frame compression method using a representative feature sequence and a speech recognition method using the same according to the present invention.

Figure 4 is an example of a feature sequence similarity measurement process using a multi-layer perceptron artificial neural network in accordance with the present invention.

* Explanation of symbols for the main parts of the drawings

102: microphone 103: A / D conversion unit

104: feature extraction unit 105: optimal model search unit

106: voice data 107: voice model selection unit

108: voice model training department

In accordance with another aspect of the present invention, there is provided a frame compression method, comprising: a first step of dividing an input signal into frames at predetermined time intervals; Extracting a feature sequence for each of the divided frames; Obtaining a similarity between the extracted feature strings; And a fourth step of reducing the number of frames by obtaining a representative feature sequence representing similar feature sequences using the obtained similarity.

The frame compression method may further include a fifth step of adding to the representative feature string a number of original (actual) feature strings represented by the representative feature string.

On the other hand, the voice recognition method of the present invention for achieving the above another object, the voice recognition method applied to the voice recognition system, the first step of dividing the voice input signal into a predetermined time interval divided into frames; Extracting a feature sequence for each of the divided frames; Obtaining a similarity between the extracted feature strings; A fourth step of reducing the number of frames by obtaining a representative feature sequence representing similar feature sequences using the obtained similarity degree; And a fifth step of performing voice recognition on the voice signal having the reduced number of frames and outputting a result thereof.

In addition, the speech recognition method may recognize the speech using the duration of the speech by adding the number of original (real) feature strings represented by the representative feature string to the representative feature string after performing the fourth step. It further comprises a sixth step to make.

On the other hand, the present invention provides a frame compression apparatus having a processor, comprising: a first function of dividing an input signal into frames at predetermined time intervals; A second function of extracting a feature sequence for each of the divided frames; A third function of obtaining similarity between the extracted feature strings; And a computer readable recording medium having recorded thereon a program for realizing a fourth function of reducing the number of frames by obtaining representative feature strings representing similar feature strings using the obtained similarity.

The present invention also provides a speech recognition system including a processor, comprising: a first function of dividing a voice input signal into predetermined frames at predetermined time intervals; A second function of extracting a feature sequence for each of the divided frames; A third function of obtaining similarity between the extracted feature strings; A fourth function of reducing the number of frames by obtaining representative feature strings representing similar feature strings using the obtained similarity degree; And a computer readable recording medium having recorded thereon a program for realizing a fifth function of performing speech recognition on a speech signal having a reduced number of frames and outputting a result thereof.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is an exemplary configuration diagram of a speech recognition system to which the present invention is applied.

First, when a person speaks, the voice 101 is input to the voice recognition system through the microphone 102. Then, the analog / digital (A / D) converter 103 converts the voice signal input through the microphone 102 into digital data, and the feature extractor 104 converts the voice signal into digital data. The feature necessary for speech recognition is extracted from the digital data converted at 103.

The optimum model searcher 105 compares the feature vectors extracted by the feature extractor 104 with the speech model 109 and outputs a word represented by the most similar speech model as a recognition result. In this case, the voice model 109 for comparison selects a voice model from the voice data 106 recorded from a plurality of people (107), and trains the voice model to obtain the voice model (108).

2 is a conceptual diagram of a frame compression method using a representative feature sequence according to the present invention, and FIG. 3 is a flowchart of a frame compression method using a representative feature sequence and a speech recognition method using the same according to the present invention.

First, the concept of the present invention is outlined, while the conventional speech recognition system processes speech recognition using frames at regular time intervals, whereas in the present invention, one representative frame (representative feature string) is obtained for similar frames. In addition, by performing speech recognition on the representative frame, the number of frames to be processed by the speech recognition system is reduced to improve the processing speed.

In addition, the present invention adds the number of actual frames represented by the frame to the feature for speech recognition obtained from the frame, and, unlike the conventional speech recognition system, uses speech length information to recognize speech performance (recognition rate). Improve).

2 and 3, a frame compression method using a representative feature sequence and a speech recognition method using the same according to the present invention will be described in detail.

First, when a voice is input from the user through the microphone 102 (301), the analog voice signal is converted into digital (302). In FIG. 2, 201 shows a waveform of an input voice.

Thereafter, the input speech waveform is divided into frames at regular time intervals (303). As shown in 201 of FIG. 2, one section divided by vertical lines is referred to as one frame. These frames may overlap some.

Then, the feature string is extracted 304 for each divided frame. 202 of FIG. 2 shows the extracted feature string. In this case, the general speech recognition system performs speech recognition using this feature sequence.

However, in the present invention, if a similar feature sequence is contiguous by looking at the similarity level between the feature sequence and the neighboring feature sequence, one representative feature sequence 203 that can represent them is obtained (305), and speech recognition is performed thereto (307). In operation 308, the voice recognition result is output.

Further, in the present invention, after further performing step 306 of adding the feature 204 of the feature string represented by the feature string to the representative feature string 203 as a feature, speech recognition using the duration of speech is performed. In operation 307, the voice recognition result is output (308).

Meanwhile, when the similarity between the feature strings is measured in more detail, when the similarity between two or more consecutive feature strings is larger than a predetermined value, the consecutive feature strings are represented as one representative feature string. . At this time, the representative feature string is determined as an average of each feature string.

As an example of a method for measuring the similarity of feature strings, the following three methods are proposed.

As a first example, we propose a method that uses the inverse of the Euclidean distance.

Similarity between feature columns A and B = , Where A _i Feature of column A i Value of the second dimension, B _i Feature of column B i Is the value of the second dimension, and N represents the number of dimensions of each feature column.

As a second example, a method using the output of multilayer perceptron artificial neural networks is proposed.

As shown in Fig. 4, two feature strings 401 are input to the neural network (402), and one output node 403 is placed to make the value of this output node a similarity. The target pattern for the training of neural networks is the state of each hidden Markov model of the speech recognition system using the Viterbi algorithm on the training speech data, and the state of one state. The feature strings belonging to are given a similarity of 1, and the feature strings belonging to other states are trained by giving a target value of similarity 0.

And the third example is as follows.

After the feature string is obtained for the training voice data, the difference between two consecutive feature strings is obtained. Vector quantization is performed on the feature sequence differences, and a Gaussian distribution function is obtained for each group. For each Gaussian function obtained, compare the majority of the feature strings in the group with or without the same state and label them equal or different. In the recognition situation, the difference of feature strings is input to each distribution function to determine the probability density, and according to the tag of the distribution function having the highest probability density, the similarity is set to 1 for 'equal' and 0 for 'different'. .

The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes can be made in the art without departing from the technical spirit of the present invention. It will be apparent to those of ordinary knowledge.

The present invention as described above has the effect of greatly improving the performance and recognition performance of the conventional frame-based speech recognition system.

That is, in the present invention, while the existing speech recognition system processes speech recognition using frames at regular time intervals, one representative frame (representative feature string) is obtained for similar frames, and the speech recognition is performed for this representative frame. By reducing the number of frames to be processed in the speech recognition system has an effect that can improve the processing speed.

In addition, the present invention adds the number of actual frames represented by the frame to the feature for speech recognition obtained from the frame, and, unlike the conventional speech recognition system, uses speech length information to recognize speech performance (recognition rate). ) Can improve the effect.

Claims (24)

In the frame compression method applied to the frame compression apparatus, A first step of dividing the input signal into frames at predetermined time intervals; Extracting a feature sequence for each of the divided frames; Obtaining a similarity between the extracted feature strings; And A fourth step of reducing the number of frames by obtaining a representative feature sequence representing similar feature sequences using the obtained similarity; Frame compression method comprising a. The method of claim 1, A fifth step of adding to the representative feature string the number of original (actual) feature columns represented by the representative feature string Frame compression method further comprising. The method according to claim 1 or 2, The process of measuring the similarity between the feature strings of the third step, And if both similarities between two or more consecutive feature strings are larger than a predetermined value, expressing the consecutive feature strings as one representative feature string. The method of claim 3, wherein The representative feature string is, And determining the average of each successive feature string having the similarity greater than a predetermined value. The method according to claim 1 or 2, The third step, Frame compression method characterized by using the inverse of the Euclidean distance. The method of claim 5, The process of obtaining the similarity of the third step is The similarity between the feature strings A and B is calculated by the following equation. ( A _i Feature of column A i Value of the second dimension, B _i Feature of column B i Value of the first dimension, N is the number of dimensions of the feature column) The method according to claim 1 or 2, The third step, Frame compression method characterized by using the output of the Multilayer Perceptron Artificial Neural networks. The method of claim 7, wherein The multilayer perceptron artificial neural network, A frame compression method using a state boundary of a Hidden Markov model as a target value. The method according to claim 1 or 2, The third step, A frame compression method characterized by using the magnitude of the probability density of a Gaussian distribution function. The method of claim 9, The third step, A sixth step of obtaining a difference between two successive feature strings after obtaining the feature strings for the training voice data; A seventh step of performing vector quantization on the obtained feature string differences and obtaining a Gaussian distribution function for each group; An eighth step of comparing each of the obtained Gaussian functions with each other whether or not the difference in the feature strings belonging to the group belongs to the same state; And A ninth step of inputting a difference of feature strings into each distribution function to obtain a probability density and specifying similarity according to the comparison result of the eighth step with respect to the distribution function having the highest probability density; Frame compression method comprising a. In the speech recognition method applied to the speech recognition system, A first step of dividing the voice input signal into frames at predetermined time intervals; Extracting a feature sequence for each of the divided frames; Obtaining a similarity between the extracted feature strings; A fourth step of reducing the number of frames by obtaining a representative feature sequence representing similar feature sequences using the obtained similarity degree; And A fifth step of performing speech recognition on the speech signal having the reduced number of frames and outputting the result; Speech recognition method comprising a. The method of claim 11, After performing the fourth step, adding the number of original (actual) feature strings represented by the representative feature strings to the representative feature strings so that the speech can be recognized using the duration of the voices. Speech recognition method further comprising. The method according to claim 11 or 12, The process of measuring the similarity between the feature strings of the third step, And if the similarity between two or more consecutive feature strings is greater than a predetermined value, expressing the consecutive feature strings as one representative feature string. The method of claim 13, The representative feature string is, And determining the average of each successive feature string whose similarity is greater than a predetermined value. The method according to claim 11 or 12, The third step, A speech recognition method using an inverse of the Euclidean distance. The method of claim 15, The process of obtaining the similarity of the third step is The similarity between the feature strings A and B is calculated by the following equation. ( A _i Feature of column A i Value of the second dimension, B _i Feature of column B i Value of the first dimension, N is the number of dimensions of the feature column) The method according to claim 11 or 12, The third step, A speech recognition method using the output of multilayer perceptron artificial neural networks. The method of claim 17, The multilayer perceptron artificial neural network, A speech recognition method using a state boundary of a Hidden Markov model as a target value. The method according to claim 11 or 12, The third step, A speech recognition method using the magnitude of the probability density of a Gaussian distribution function. The method of claim 19, The third step, A seventh step of obtaining a difference between two consecutive feature strings after obtaining a feature string for the training voice data; An eighth step of performing vector quantization on the obtained feature string differences and obtaining a Gaussian distribution function for each group; A ninth step of comparing each of the obtained Gaussian functions with each other whether or not the difference in the feature strings belonging to the group belongs to the same state; And A tenth step of inputting a difference of feature strings into each distribution function to obtain a probability density and specifying similarity according to the comparison result of the eighth step with respect to the distribution function having the highest probability density; Speech recognition method comprising a. In a frame compression device having a processor, A first function of dividing an input signal into frames at predetermined time intervals and dividing the input signal into frames; A second function of extracting a feature sequence for each of the divided frames; A third function of obtaining similarity between the extracted feature strings; And A fourth function of reducing the number of frames by obtaining a representative feature string representing similar feature sequences using the obtained similarity A computer-readable recording medium having recorded thereon a program for realizing this. The method of claim 21, A fifth function of adding to the representative feature string a number of original (actual) feature columns represented by the representative feature string A computer-readable recording medium that records a program for further realization. In a speech recognition system having a processor, A first function of dividing a voice input signal into frames at predetermined time intervals and dividing the voice input signal into frames; A second function of extracting a feature sequence for each of the divided frames; A third function of obtaining similarity between the extracted feature strings; A fourth function of reducing the number of frames by obtaining representative feature strings representing similar feature strings using the obtained similarity degree; And A fifth function of performing speech recognition on the speech signal having the reduced number of frames and outputting the result; A computer-readable recording medium having recorded thereon a program for realizing this. The method of claim 23, A sixth function that adds the number of original (actual) feature strings represented by the representative feature string to the representative feature string after performing the fourth function to recognize the voice using the duration of the voice; A computer-readable recording medium that records a program for further realization.