KR100339525B1 - Device for guiding living information using two step keyword recognizing method - Google Patents

Abstract

PURPOSE: A device for guiding living information using a two step keyword recognizing method is provided to recognize a keyword through two stages by dividing keywords according to classes, dividing keywords which represent each class into a class keyword and a sub class keyword. CONSTITUTION: A voice feature extracting unit(100) extracts a feature from a filtered and quantized voice signal with respect to a voice signal inputted through a microphone(9). The first stage keyword recognition unit(200) extracts the first stage class keyword from the feature. A second stage keyword recognition unit(300) extracts a sub class keyword possessed in the class keyword. A recognition result combining unit(21) combines recognition languages of the first stage keyword recognition unit(200) and the second stage keyword recognition unit(300) and outputs the combined recognition languages. The first stage class keyword extracting recognition unit(14) progresses a viterbi process using a class keyword model(13) and filler models(15). If an accumulated probability value of the class keyword model(13) and a probability value in the filler model(15) with respect to a searched section are larger than a critical value, the first stage class keyword extracting recognition unit(14) extracts a class keyword. A recognition unit(16) judges whether the class keyword extracted in the first stage class keyword extracting recognition unit(14) is recognized or rejected, and recognizes or rejects the class keyword.

Description

Life information guide device using the 1st and 2nd stage central word recognition method

The present invention relates to speech recognition, and in particular, solves the limitations and unnaturalness of vocabulary, which is a disadvantage in the recognition system using the existing isolated language, and uses "news", "weather", "securities" and "professional baseball" life as a natural dialogue The present invention relates to a living information guide device using the 1,2-step central word recognition method that inquires related information and informs the result.

The conventional center word recognition apparatus has a low pass filter 2 and an analog / digital converter 3 for filtering and quantizing voice information input through the microphone 1, as shown in FIG. The feature extractor 4 extracts the features of the quantized speech through the digital converter 3 and the core words and the filler model 7 trained in advance by inputting the features through the feature extractor 4. The central word extraction recognizer 6 outputting a cumulative probability value by performing the Viterbi process, and the cumulative activity rate value output from the central word extract recognizer 6 and the threshold are compared to reject the corresponding central word. It consists of the recognition part 8 which decides whether to do it.

Looking at the conventional technology configured as described above are as follows.

When the voice path s (t) is input through the microphone 1, the low pass filter 2 performs low pass filtering and converts the low pass voice signal into a digital value by the analog / digital converter 3. It is delivered to the feature extraction unit (4).

Then, when the feature extractor 4 extracts and outputs a feature of the voice, the center word extraction recognizer 6 that receives the output is the center word and the filler model 7 of the data storage unit 5 that are pre-learned and stored. When the Viterbi process is performed and output to the recognizer 8, the recognizer 8 compares the cumulative probability value obtained through the central word extraction recognizer 6 with a threshold value and becomes larger than a threshold value. It is said that it is recognized as the key word for the voice section, and it is not recognized for the other sections.

However, in the related art, a common error is frequently generated and recognized to be applied by applying a device that recognizes a corresponding center word for a voice interval exceeding a specific threshold by comparing the cumulative probability values equally with respect to all the central object models to be recognized. When the number of central words increases, there is a problem that it is difficult to recognize in real time in limited memory and calculation speed.

Accordingly, an object of the present invention is to divide the central words used in the living information guide service by each class, and to classify the central words representing each class as the class key words. The present invention provides a living information guide device using a 1,2-step central word recognition method, which divides a class key word into two levels.

Another object of the present invention is that it is impossible to recognize a large number of central words at the same time using only limited resources such as memory or computational speed, but it is recognized at once by dividing them into a central word set representing a class of a target service and a set of central words belonging to it. In order to reduce the number of core words to be processed in real time to provide a sage information guide device using the first and second steps of the central word recognition method.

The present invention for achieving the above object is a voice feature extraction unit 100 for extracting features from the voice signal filtered and quantized for the voice signal input through the microphone 9, as shown in FIG. The sub-class center word belonging to the classification of the classification-center word extracted by the first-stage central word recognition unit 200 and the first-stage central word recognition unit 200 from the feature extracted by the voice feature extraction unit 100. And a recognition result combining unit 21 for combining and outputting the recognized words of the first-stage central word recognition unit 200 and the second-stage central word recognition unit 300 to extract the two-stage central word recognition unit 300.

In addition, the first-stage central word recognition unit 200 performs the Viterbi process using a pre-modeled classification core word model 13 and a filler model 15 corresponding thereto, and the classification core word for the section found. A first-stage classification core word extraction recognizer 14 that obtains a cumulative probability value for the model 13 and a corresponding probability value in the filler model 15 corresponding thereto and extracts it as a classification center if it is larger than a threshold value; And a recognition unit 16 for determining whether to recognize or reject the classification center word extracted at (14).

In addition, the two-stage central word recognition unit 300 uses the partial configuration of the central word model 17 and the filler model 19 as the model used in performing the same configuration and the same operation as the first-stage central word recognition unit 200. do.

When described in detail with respect to the operation and effect of the present invention configured as described above.

When the voice signal S (t) is input through the microphone 9, it is first filtered through the low pass filter 10 and quantized through the analog / digital converter 11 (S (N)), and the quantized signal. The feature extractor 12 extracts feature vectors representing voice signals at predetermined time intervals, and the extracted feature vectors are input to the first-stage central word recognition unit 200 to extract one of the classification core word models. do.

Then, the first-stage classification core word recognition recognition unit 14 of the first-stage central word recognition unit 200 uses a pre-modeled classification core word model 13 and a filler model 15 corresponding thereto to perform a Viterbi process. The recognition unit which finds a speech section, finds a cumulative probability value for the classification core word model 13 and a probability value in the filler model 15 corresponding to the found section, and compares them using these values. Decide whether to recognize or reject the central word compared in (16).

In the case of the rejection, the feedback is inputted to the first stage classification core word extraction recognition unit 14, and when it is recognized, the second stage central language recognition unit 300 is output.

Then, the second stage central word recognition unit 300, the second stage subclass center word extraction recognition unit 18, the subclass key word model 17 belonging to the classification center word classification and the filler model 19 corresponding thereto. The central word is extracted from the two-stage subclass center word extraction recognition unit 18 using the same process.

That is, the second stage subclass center word extraction recognition unit 18 performs a Viterbi process using a pre-modeled subclass center word model 17 and a corresponding filler model 19 to find and recognize a speech segment. When outputting to the unit 20, the recognition unit 20 obtains the cumulative probability values for the subclass center word model 17 and the probability values in the filler model 19 corresponding to the found voice intervals. Are compared to each other to decide whether to recognize subtypes as a central word or not.

In the first and second core words and filler models used in the first and second steps, as shown in FIG. 3 and FIG. do.

For example, "How's the weather tomorrow?" In the first step of the sentence, the classification core word set Ⅱ = {extracts the classification key word "weather" from "news", "weather", "movie" ......} and sets the subclass set of weather Θ = { "today". The words "Tomorrow", "Weekend", "South", "Seoul" ......} are extracted.

Therefore, the recognition results of "weather" and "tomorrow" are combined in the recognition result combining unit 21 to grasp the meaning of "weather tomorrow" and output the corresponding weather forecast.

The core word set used in the present invention is the words shown in FIG. 5 and the words located in the second line are the words corresponding to the class center words, and the words connected below the sub-words belong to the classification.

Instead of recalculating the feature vectors obtained for the input sentences, only the Viterbi process is repeated by using networks of the same structures as in FIGS.

As mentioned above, in order to perform real life information service, it is necessary to recognize the central words of dozens of words, but it is impossible to recognize a large number of central words at the same time by using only limited resources such as memory and computation speed, Real-time processing is possible by reducing the number of central words recognized at once by dividing them into a set of central words and a set of central words belonging to them.Because only one classification core word representing a classification is extracted from a sound signal pronounced in one step, errors that are incorrectly recognized can be reduced. In addition, the subclass center words belonging to the classification may be recognized in two stages.

1 is a block diagram of a conventional center word recognition apparatus.

2 is a block diagram of the life information guide device using the method of recognizing the central and first stage of the present invention.

3 is a network explanatory diagram of a core word and a filler used for extracting a first-stage classification key word in FIG.

4 is a network explanatory diagram of a core word and a filler used for extracting a two-stage classification key word in FIG.

5 is a central word set used in FIG.

***** Explanation of symbols for main parts of drawing *****

1,9: Microphone 2,10: Low pass filter

3,11: Analog (A) / Digital (D) Converter

4,12: Feature extraction part: 6: Central word extraction recognition part

7,15,19: Filler model 8,16,20: Recognition unit

13: classification core word model 14: first stage core word extraction recognition unit

17: Sub-class core word model 18: Two-level core word extraction recognition unit

21: Recognition result combining unit 100: speech feature extraction unit

200: step 1 core word recognition unit 300: step 2 central word recognition unit

Claims (2)

Speech feature extraction means for extracting a feature vector from the speech signal filtered and quantized with respect to the speech signal input through the microphone, a classification model model pre-modeled from the feature vector extracted by the speech feature extraction means, and a corresponding filler model A first-stage central word recognition means for extracting a first-stage classification central word using the first-stage classification core word, a sub-class core word model pre-modeled from the feature vector extracted by the speech feature extraction means, and a filler model corresponding thereto 1 to 2, characterized in that it comprises a two-stage central word recognition means for extracting the subclass center words belonging to the classification of the classification center word extracted in the step; Life information guide device using step-centered word recognition method. The method of claim 1, wherein the first-stage central word recognition unit comprises: a first-stage classification center word extraction recognizing unit that performs a Viterbi process using a pre-modeled classification core model and a filler model corresponding thereto to find a speech section; Recognition means for determining whether or not to recognize the classification center word as classification center by comparing the threshold with the cumulative probability value for the classification center word model and the probability value in the filler model corresponding to the speech interval of Life information guide device using a 1,2-step central word recognition method characterized in that consisting of.