JP2885399B2 - Pattern recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Object of the Invention] (Industrial application field) The present invention relates to a pattern recognition device that effectively learns a recognition dictionary used for pattern recognition. (Prior Art) In pattern recognition in a voice recognition device such as a voice word processor or in recognition processing of handwritten characters, in order to improve the recognition performance (recognition rate), a large number of recognition target patterns are collected in advance, and these patterns are collected. Learning of recognition dictionaries is performed by using them to enhance the recognition dictionaries. In the learning of the recognition dictionary, for example, a word to be input is displayed on a display device when a voice recognition device for a specific speaker is used as a word unit, and the speaker speaks the word specified on the display device. In this way, the procedure is repeated by collecting one learning pattern for the word, displaying the next character when the registration of the recognition dictionary is completed, and collecting the next learning pattern. However, according to such an input method, an idle time corresponding to the thinking time of the speaker after the word to be input is displayed on the display device until the speaker actually speaks the displayed word. Occurs. This vacant time is a time that cannot be ignored when registering an extremely large number of words, and is a major obstacle in improving the efficiency of the learning process of the recognition dictionary. In addition, such an obstacle not only increases the burden on the speaker who makes an effort to reduce the learning processing time, but also leads to the possibility of inducing erroneous input. In the character recognition device, when the writer's character is registered online, the same problem occurs because the input instruction is performed in units of characters and words. In speech recognition, research is being conducted on continuously uttered speech recognition in addition to word-based input. In this continuous speech recognition, information on a so-called crossover portion between words is used. Is important information in performing the recognition process. However, in the conventional method of inputting a learning pattern as described above, the continuity between words is impaired due to the presence of an idle time between the designation of an input word by the display device and the actual input of the word, There was a problem that information on the part between words could not be collected. (Problems to be Solved by the Invention) As described above, each time the registration of the input pattern is completed, the next input word is displayed, thereby sequentially collecting the input patterns and learning the recognition dictionary. In such a case, there is a problem that a large amount of time is required for the learning process, and a burden on an input person and an erroneous input increase. Further, particularly in continuous speech recognition, there is a problem that information on a change between words cannot be obtained. The present invention provides a pattern recognition device that can improve the efficiency of the learning process and reduce the burden on the input person and reduce erroneous input. In particular, in continuous speech recognition, a pattern recognition device capable of collecting information on changes between words and phrases is provided. The purpose is to provide. [Structure of the Invention] (Means for Solving the Problems) A pattern recognition apparatus according to the present invention includes a recognition unit that recognizes an input pattern using a recognition dictionary, and an input unit that inputs an input pattern when learning the recognition dictionary. While the characters to be displayed are displayed on the display device in a lump based on the order in which the characters to be input are collectively input, a group of the characters to be input this time should be input and a group of the input should be input later. Characters to be input this time and characters to be input later are displayed in a first display method that is distinguished from a group of characters, and characters to be input this time are displayed in the group of characters displayed in the first display method. Display control means for repeatedly performing an operation of displaying in a second display method distinguished from each other for each time,
Learning means for learning the recognition dictionary based on a recognition result by the recognition means for an input pattern input in association with the character to be input currently displayed on the display device by the second display method. It is characterized by having done. Further, the pattern recognition device according to the present invention includes a recognition unit that recognizes an input pattern using a recognition dictionary, and a group of characters to be input as an input pattern in a predetermined number of times when learning the recognition dictionary. In a state of being displayed on the display device in a manner based on the input order,
Display control means for repeatedly performing, for each input, an operation of displaying a character to be input this time in the group of characters being displayed in a predetermined display method that is distinguished from a character that has already been input and a character that is to be input later Learning means for learning the recognition dictionary based on a recognition result by the recognition means for an input pattern input in association with the character to be input currently displayed on the display device by the predetermined display method. The display control means is characterized in that characters to be input are displayed in a pair of a kanji notation and a kana notation. Further, the pattern recognition device according to the present invention includes a recognition unit that recognizes an input pattern using a recognition dictionary, and a group of characters to be input as an input pattern in a predetermined number of times when learning the recognition dictionary. In a state of being displayed on the display device in a manner based on the input order,
Display control means for repeating, for each input, an operation of displaying a character to be inputted this time in the group of characters being displayed in a color distinguishable from a character which has already been inputted and a character to be inputted later; Learning means for learning the recognition dictionary based on a recognition result by the recognition means for an input pattern input in association with the character to be input currently displayed on the display device in the predetermined display method. It is characterized by having done. (Operation) According to the present invention, not only the characters to be currently input to the display control means but also the already input characters and the characters to be input thereafter are displayed. Therefore, those who want to enter the input pattern, while entering the characters to be entered now,
Since the characters that have already been input and the characters to be input can be confirmed, it is possible to input the next character immediately after inputting one character accurately, promptly, and without burdening the input user. Can be achieved. Therefore, according to the present invention, the learning time can be reduced, the burden on the input person can be reduced, and erroneous input can be prevented. In the case of continuous speech recognition, information on a changed portion such as between words or between phrases can be collected. Will be possible. Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic configuration diagram of a speech recognition device according to the present embodiment. Although this apparatus performs pattern recognition of an input voice, it can be similarly applied to pattern recognition such as handwritten character recognition. The voice input unit 1 is a part that includes a microphone and an amplifier and converts input voice into an electric signal. The audio signal input via the audio input unit 1 is supplied to a feature extraction unit 2 composed of, for example, an eight-channel filter bank, where it is frequency-decomposed and each frequency component is divided into, for example, eight points in the time axis direction. The input speech pattern is converted into a sampled 64 (8 × 8) -dimensional feature vector. Note that feature extraction can also be performed by fast Fourier transform analysis, cepstrum analysis, or the like, instead of the above-described filter analysis. The recognizing unit 3 calculates the similarity S between the feature vector f of the input speech and the vector component of each n-plane of the dictionary pattern {φn ^(l) } of the category 1 registered in the recognition dictionary memory 4.
^(l) [f] is The recognition result is determined by calculating the composite similarity. Here, λn ^(l) and φn ^(l) indicate the nth eigenvalue and the eigenvector of the covariance matrix K obtained in advance from many sample patterns for each recognition target category, and these are the recognition dictionary memory. 4 is stored. Instead of using the similarity S to determine the recognition result, it is also possible to obtain the recognition result using the Mahalanobis general distance, Euclidean distance, probability, likelihood, and the like. The learning training pattern storage unit 5 associates the learning training pattern obtained by frequency decomposition of the voice data input at the time of the learning process with the feature extraction unit 2 with the recognition result of the recognition unit 3. It is something to memorize. The learning unit 6 inputs the input voice pattern obtained through the feature extracting unit 2 and the learning training patterns collected and stored in the learning training pattern storage unit 5 and stores them in the recognition dictionary memory 4. The learned dictionary pattern is learned for each recognition target category. The learning of the dictionary pattern in the learning unit 6 is performed by, for example, iterative updating of the covariance matrix K and KL expansion of the covariance matrix K. That is, the covariance matrix K of the recognition dictionary of the category to be learned is input, and the covariance matrix is basically updated as follows. That is, Is performed to update the covariance matrix K. Here, V is a learning pattern, for example, given as a 64-dimensional vector. K ′ is a covariance matrix after learning. W is a weighting coefficient (a value of a learning parameter) that takes a positive or negative value obtained from the recognition result of the recognition unit 3, and when positive, increases the similarity of the characteristic nucleus of the covariance matrix K to the input pattern. An effect is exhibited, and if negative, an effect of reducing the similarity is exhibited. The display unit 7 displays the recognition result recognized by the recognition unit 3 during normal recognition processing, and displays characters to be input for learning when learning the recognition dictionary. Characters to be displayed on the display unit 7 are stored in the generated character string storage unit 8. The underline / reverse display control unit 9 also underlines a unit indicating one unit of the character string displayed on the display unit 7,
The display control for reversely displaying the character to be input at present is performed. The category of the character to be currently input specified by the underline / reverse display control unit 9 is associated with the obtained learning training pattern and the learning training pattern storage unit 5.
Is stored in In FIG. 1, the control unit 10 controls the above-described pattern recognition processing and recognition dictionary learning processing. When learning the recognition dictionary in the present apparatus configured as described above, first, the control unit 10 reads a character string to be learned and input from the utterance character string storage unit 8 and causes the display unit 7 to display the character string. The underline / reverse display control unit 9 underlines one unit, and performs a reversal process on the character to be input. FIG. 2 shows a display form of the display unit 7 obtained by such processing. Here, “up to this way to similarity S” is a unit, and this character string is underlined to indicate that “in this way” is the character (phrase) to be currently input. Is shown. Then, when the speaker speaks “this way” according to this display, the voice input unit 1 inputs this voice,
The feature extraction unit 2 extracts the feature pattern. This feature pattern is stored in the training training pattern storage unit 5 via the recognition unit 3 as a training training pattern. Recognition unit 3
Is stored in the learning training pattern storage unit 5 in association with the learning training pattern. Next, the learning unit 6 performs the learning process as described above, expands the covariance matrix K ′ obtained as a result of the learning by KL, and specifies its eigenvalues and eigenvectors in the specified recognition target category “in this manner. Is registered in the recognition dictionary memory 4. When the registration of the recognition dictionary is completed, the underline / reverse display control unit 9 reversely displays the phrase "to be determined" to be input next. Then, similar learning is performed. FIG. 3 shows the procedure of such a display process. As shown in this figure, the display unit 7 firstly outputs the uttered character string storage unit 8
, A plurality of characters to be displayed are read out and displayed (11), and a unit indicating one unit of the displayed characters is underlined (12) and an input instruction is awaited (13). Then, if the input is possible (13), the character to be input among the underlined display characters is highlighted (14) and the input is awaited (15).
When the specified character is input, the highlighted display is moved to the next character to be input (16). If there is no character to be input next (17), the underline display is moved to the next group (18). If there is no next unity (19), the uttered character string storage unit 8
To see if there are any characters left to display (2
0), if any, update the display contents (2
1). According to this device, when the speaker utters the first phrase “in this way”, it is known in advance by the display on the display unit 7 that the next phrase to be uttered is “determined”. When the input is requested, the requested character can be uttered and input immediately. Therefore, the input time can be reduced, the burden on the speaker can be reduced, and erroneous input can be prevented. In addition, it is possible to collect information of a so-called crossover between phrases in continuous speech. As shown in FIG. 4, if the input units are separated by a slash or the like, the length of a character string to be input next is also known in advance, so that quicker and more reliable input is possible. In the above embodiment, a group of characters to be input is underlined. This does not require that the entire content be uttered, but has the effect that the user can utter in a more natural environment by showing the user a meaningful group. FIG. 5 shows an example of input in single syllable units, and FIG. 6 shows an example of input in word units. In the example of FIG. 5, the unit of the word “ぽ ぱ” is underlined, and the syllable “ぽ” to be inputted in the underlined word is highlighted. The example in FIG. 6 is an example in which the hiragana and kanji notation of a word are displayed as a pair, thereby preventing the utterance error of the word. Here, the pair of the hiragana notation and the kanji notation is underlined, and the hiragana notation is highlighted as a word to be currently input. In FIG. 5, the words “address”, “turn on” and “kiori” and “Tokyo Tokyo” and “Yurakucho Yurakucho” in FIG. These are distinguished from other words by being displayed in high brightness. Also,
Such a high-brightness display, underline display, reverse display, or the like can be replaced with color classification when a display portion capable of color display is used. In addition, the present invention may add a function of displaying whether the input is valid or invalid. The validity of the input can be indicated by moving the reverse display portion or by erasing the input character. When the total number of input units and the number of input units input so far are displayed on the display unit 7, the speaker can predict the remaining character input amount, so that the burden on the speaker is further reduced. Although the present invention has been described above with reference to a voice recognition device as an example, the present invention can be applied to pattern recognition processing devices conforming to various recognition algorithms, such as handwritten character recognition. [Effects of the Invention] As described above, according to the present invention, when learning the recognition dictionary, not only characters to be currently input to the display control means,
Since characters to be input are also displayed thereafter, the process of registering the learning pattern can be performed efficiently, and the burden on the user can be reduced and erroneous input can be prevented. In particular, when the present invention is applied to continuous speech recognition, it is possible to easily collect information on a changed portion such as between words or between phrases, and to produce a recognition dictionary with excellent performance.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a schematic configuration diagram of a voice recognition device according to an embodiment of the present invention, FIG. 2 is a diagram showing a display example of a display unit in the device, and FIG. 4 to 6 are views showing other display examples of the apparatus. 1 ... voice input unit, 2 ... feature extraction unit, 3 ... recognition unit,
4 ... Recognition dictionary memory, 5 ... Learning training pattern storage unit, 6 ... Learning unit, 7 ... Display unit, 8 ... Speech character string storage unit, 9 ... Underlined / reversed display control unit, 10 ... ... Control unit.

Continuation of the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00 521 G10L 3/00 561

Claims (1)

(57) [Claims] A recognition unit for recognizing an input pattern using a recognition dictionary; and, in a process of learning the recognition dictionary, a state in which characters to be input as an input pattern are collectively and displayed on a display device in an arrangement manner based on an input order. A first set of characters that is distinguished from a set of parts to be input this time from a set of parts already input and a part of a set to be input later.
Is displayed in the second display method in which characters to be input this time are distinguished from characters that have already been input and characters that are to be input later in the group of parts displayed in the first display method. Display control means for repeatedly performing an operation to be displayed each time, and the recognition means for an input pattern input in association with the character to be input this time displayed on the display device by the second display method. A pattern recognition device comprising: learning means for learning the recognition dictionary based on a recognition result. 2. The first display method is to display the group to be input this time underlined, and the second display method is to reversely display the characters to be input this time. The pattern recognition device according to claim 1, wherein 3. The pattern recognition device according to claim 1, wherein both the first display method and the second display method are based on color-coded display. 4. 2. The pattern recognition apparatus according to claim 1, wherein said display control means inserts a separation mark such as a space or a slash for each input unit to specify the input unit. 5. 2. The pattern recognition device according to claim 1, wherein the characters to be input are displayed in pairs of kanji and kana. 6. A recognition unit for recognizing an input pattern using a recognition dictionary; and a learning method for the recognition dictionary, wherein a set of characters to be input as an input pattern in a predetermined number of times is arranged collectively and based on an input order. In a state where the characters are displayed on the display device, an operation of displaying the characters to be input this time in the group of characters being displayed in a predetermined display method that is distinguished from the already input characters and the characters to be input later, A display control unit that repeats every input, and the recognition based on a recognition result by the recognition unit for an input pattern input in association with the character to be input this time displayed on the display device in the predetermined display method. Learning means for learning a dictionary, wherein the display control means displays a character to be inputted in a pair of a kanji notation and a kana notation. That pattern recognition device. 7. 7. The pattern recognition apparatus according to claim 6, wherein the predetermined display method is to reversely display the character to be input this time. 8. 7. The pattern recognition apparatus according to claim 6, wherein the display control means inserts a separation mark such as a space or a slash for each input unit to specify the input unit. 9. A recognition unit for recognizing an input pattern using a recognition dictionary; and a learning method for the recognition dictionary, wherein a set of characters to be input as an input pattern in a predetermined number of times is arranged collectively and based on an input order. In the state of being displayed on the display device, the operation of displaying the characters to be input this time in the group of characters being displayed in a color that can be distinguished from the already input characters and the characters to be input later is performed by each input. Display control means that is repeated every time, and the recognition dictionary based on a recognition result by the recognition means for an input pattern input in association with the character to be input this time, which is displayed on the display device in the predetermined display method. And a learning means for learning the pattern. 10. 10. The display control means according to claim 9, wherein the display control means inserts a separation mark such as a space or a slash for each input unit to specify the input unit.
The pattern recognition device according to the paragraph. 11. 9. The ninth aspect of the present invention, wherein characters to be input are displayed in pairs of a kanji notation and a kana notation.
The pattern recognition device according to the paragraph.