JP2000099082A - Voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make an input code split string to be obtained by storing words of foreign origin split character strings which are to be obtained by dividing words of foreign origin into a constat length by relating them to their words of foreign origin split voice code strings to estimatingly generate a word of foreign origin which is not registered in a word dictionary. SOLUTION: A word dictionary 3 storing words by relating them to registered word character strings and registered word phoneme strings and also an n-gram dictionary (words of foreign split character storage means) 4 storing words of foreign original split character strings which are to be obtained by diving words of foreign original into a constant length by relating them to their word of foreign original split phomene strings are provided. Then, a word estimating means 5 and a katakana unknown word correcting means 6 compare an input phoneme string with registered word phoneme string- a word of foreign split phoneme strings and they generate an input code split string in which the input phoneme string is divided at every coincident part. Thus, the input code split string can be generated by estimatingly generating a word of foreign original which is not registered in the word dictionary 3.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device for generating a recognition character string based on a speech input,
The present invention relates to an improvement capable of generating a recognized character string by recognizing speech including an undefined foreign word.

[0002]

2. Description of the Related Art FIG. 11 is a block diagram showing the configuration of a conventional speech recognition apparatus disclosed in Japanese Patent Application Laid-Open No. 4-73694. In the figure, reference numeral 1 denotes a microphone that outputs an analog signal in accordance with an audio input, and 2 denotes an analog signal
A phonological probability calculating means for generating a network-structured phonological lattice composed of a plurality of input phonological sequences separated into phonological units and an input appearance probability indicating the likelihood of each phonological unit by performing D conversion, quantization, and spectral analysis; A kanji dictionary that stores kanji in association with its registered kanji character string, registered kanji phoneme string, and kanji appearance probability, 11 compares, for each input phoneme string, the input phoneme string and the registered kanji phoneme string, A kanji predicting means for setting all the matched parts to kanji conversion sections; 7 a random access memory (RAM) for storing the phonological lattice and the kanji conversion sections together with corresponding appearance probabilities;
Selecting means for selecting, as an optimal input phoneme divided sequence, an input phoneme divided sequence having the highest overall appearance probability from a plurality of input phoneme divided sequences which is an aggregate of the kanji conversion sections,
Reference numeral 9 denotes output means for generating a recognized character string by associating a registered kanji character string corresponding to each divided section of the optimal input phoneme divided string.

Next, the operation will be described. When a voice is input, the microphone 1 outputs an analog signal corresponding to the voice, and the phoneme probability calculation means 2 outputs a plurality of input phoneme strings as a phoneme lattice based on the analog signal. The kanji prediction means 11 sets a kanji conversion section with reference to the kanji dictionary for each input phoneme sequence.

The selecting means 8 selects an input phoneme string having the highest overall appearance probability as a result of the setting of the kanji conversion section in this way, and the output means 9 outputs the divided section of each of the optimal input phoneme strings. A recognition character string is generated by associating a Chinese character with.

[0005]

Since the conventional speech recognition apparatus is configured as described above, the recognizable (convertible) words are limited to the words registered in the kanji dictionary 12. was there.

Accordingly, when trying to recognize a foreign word having a different way of chaining sounds (the relationship between the arrangement of sounds and the arrangement of written characters) with a predetermined language, the foreign word is converted into a kanji dictionary. 12 must be registered. However,
In a situation where foreign words are rapidly increasing or changing, as in Japanese in recent years, it is impossible to register all foreign words in the kanji dictionary 12, and such a language cannot be registered. This significantly impairs the practicality of the speech recognition technology in.

In particular, in continuous word recognition in which foreign words are mixed, not only cannot the word be converted, but also
Because that part is not converted correctly, the part of the input speech code string that should be correct is also incorrectly separated, and the overall appearance probability is unduly lowered,
The result is that a recognized character string is generated based on other incorrect speech code strings.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and a speech recognition device capable of highly recognizing speech including foreign words without registering all foreign words in a dictionary. The purpose is to obtain.

[0009]

A speech recognition apparatus according to the present invention is provided with a pre-processing means for receiving a speech in a predetermined language and generating an input speech code sequence in the language based on the speech; A word dictionary that stores words in association with a registered word character string and a registered word speech code string, and a foreign word segmented character string obtained by dividing a foreign word into a certain length, and the foreign word segmented speech A foreign word segmented character string storage unit that stores the input speech code string in association with the code string, and compares the input speech code string with the registered word speech code string and the foreign word segmented speech code string, and converts the input speech code string for each matching portion. It is provided with dividing means for generating an input code division sequence to be divided, and character string generation means for generating a recognition character string by associating character strings corresponding to respective divided portions of the input code division sequence.

The speech recognition apparatus according to the present invention is characterized in that a speech in a predetermined language is input, and based on the speech, a plurality of input speech code strings in the language are generated together with respective input appearance probabilities, A word dictionary that stores words in the above-described language in association with the registered word character strings, registered word speech code strings, and word appearance probabilities, and a foreign word divided character string obtained by dividing a foreign word into a certain length. A foreign word divided character string storage unit that stores the foreign word divided speech code string and the divided character string appearance probability in association with each other, and for each input speech code string, stores the input speech code string in the registered word speech code string and the Compared with a foreign word segmented speech code string,
A dividing means for generating an input code division sequence for dividing the input speech code sequence for each matching portion; and a total appearance probability of each input code division sequence based on an input appearance probability, a word appearance probability, and a divided character string appearance probability. And character string generating means for generating a recognized character string by associating a character string corresponding to each divided part of the input code divided string with the combination having the highest appearance probability.

The speech code of the speech recognition device according to the present invention is a phoneme, phoneme or syllable in a predetermined language.

The foreign word segmented character string of the speech recognition apparatus according to the present invention is a character string of a foreign language written in katakana or an alphabet.

[0013] The dividing means of the speech recognition apparatus according to the present invention compares a part which does not match the registered word speech code string with the foreign word divided speech code string, and divides a part which matches the foreign word divided speech code string. Things.

In the speech recognition apparatus according to the present invention, the word dictionary stores the word appearance probability for each word, and the foreign word divided character string storage means stores the divided character string appearance probability for each foreign word divided character string. The dividing unit generates the input code division sequence such that the overall appearance probability becomes highest when the input speech code sequence is divided.

In the speech recognition apparatus according to the present invention, the foreign word divided character string storage means stores the divided character string appearance probability for each foreign word divided character string. If multiple foreign word segmented strings are intermittently continuous, compare the probability of converting the entire intermittent continuous section into a foreign word segmented string with the probability of intermittently converting Then, the input code division sequence is generated by using the one having the higher section probability.

In the speech recognition apparatus according to the present invention, the dividing means compares the section probabilities after weighting the probabilities obtained when the entire intermittent continuous section is converted into a foreign word divided character string with correction weights. .

The speech recognition apparatus according to the present invention is provided with an example sentence storing means for storing a sentence in a predetermined language, and the dividing means divides an input speech code string at a portion corresponding to a word extracted from the example sentence. is there.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Embodiment 1 FIG. FIG. 1 is a block diagram showing a configuration of a speech recognition device for Japanese input according to Embodiment 1 of the present invention. In the figure, reference numeral 1 denotes a microphone (pre-processing means) for outputting an analog signal in accordance with a voice input, and 2 denotes a plurality of input phonemes which are obtained by subjecting the analog signal to A / D conversion, quantization, and spectrum analysis and separated into phoneme units. Phoneme probability calculating means (preprocessing means) for generating a phoneme lattice having a network structure composed of a sequence and input appearance probabilities indicating the likelihood of each phoneme, 3 registers words such as kanji and foreign words as registered word strings, 4 is a word dictionary that stores word phonological sequences and word appearance probabilities in association with each other, and n is a foreign word in katakana character notation (n is an integer).
An n-gram dictionary (foreign word divided character string storage means) for storing a foreign word divided character string obtained by dividing each character in association with the foreign word divided phoneme sequence and the divided character string appearance probability, For each phoneme string, the input phoneme string and the registered word phoneme string are compared, and word matching means (segmentation means) for setting all matched parts as word conversion sections, 6 is a part which does not match the registered word phoneme string. Is compared with a foreign word segmented phoneme string, and a katakana unknown word correcting unit (segmenting unit) for setting a part that matches the foreign word segmented phoneme string as an unknown foreign word conversion section, 7 is the above phoneme lattice, word conversion section and unknown A random access memory (RAM) 8 for storing a foreign language conversion section together with a corresponding appearance probability has the highest overall appearance probability among a plurality of input phoneme division sequences which are an aggregate of the conversion sections. Selecting means for selecting an input phoneme splitting column as the optimum input phoneme dividing column (string generation means)
Reference numeral 9 denotes output means (character string generation means) for generating a recognition character string by associating character strings corresponding to the respective divided sections of the optimal input phoneme division string.

[0019]

(Equation 1)

The above equation 1 is an equation for calculating the appearance probability (hereinafter referred to as predicted word probability) P (W | Y) of each word in the input phoneme sequence according to the first embodiment of the present invention. In the equation, W is a part or all of the input phoneme sequence, Y is a word phoneme sequence, and P (Y | W) is a probability of the phoneme sequence when the word W is given. 2, P (W) is the appearance probability of the word string, and can be obtained by, for example, Equation 3 below, and P (Y) is the appearance probability of the phoneme string. Then, to simply compare the appearance probabilities between words for the same input phoneme sequence, P (Y |
W) Only P (W) needs to be calculated and compared. And
In the first embodiment, W having the maximum predicted word probability P (W | Y) is selected as a conversion candidate. Although the multiplication operation is performed in Expressions 2 and 3 below, each probability may be a logarithmic probability and may be a sum operation.

[0021]

(Equation 2)

[0022]

(Equation 3)

[0023]

(Equation 4)

Equation (4) is an equation for calculating the appearance probability of unknown katakana words in the input phoneme sequence according to Embodiment 1 of the present invention. In the above equation, Ci is a katakana notation character, and the appearance probability of the katakana unknown word is obtained based on this arithmetic expression, and the predicted word probability corresponding to the katakana unknown word is obtained by substituting this appearance probability into the above equation 1. be able to.

Next, the operation will be described. FIG. 2 is a flowchart showing a speech recognition operation according to Embodiment 1 of the present invention. In the figure, ST1 is a voice input step for inputting a voice to the microphone 1, ST2 is an analog signal output step for the microphone 1 to output an analog signal corresponding to the voice, and ST3 is a phonological probability calculation means 2 for the analog input. This is a phoneme lattice generation step in which a signal is analyzed to generate a phoneme lattice composed of a plurality of input phoneme sequences and input appearance probabilities, and stored in the RAM 7.

ST4 is a comparison section extraction step of sequentially extracting one from a plurality of input phoneme strings one by one for each section of a predetermined n characters of the phoneme lattice, and ST5 is the same phoneme string as this section input phoneme string. ST6 is a failure position storage step of storing the position of the section input phoneme string on the phoneme lattice when there is no such word. ,
ST7 is a predicted word probability calculation step of calculating a predicted word probability based on the input appearance probability of the word conversion section and the word appearance probability of the matching word for the word matched in the section match determination step, and ST8 is an input phoneme sequence Is a word perfect match determination step of determining whether or not all of the words match all of the word phoneme strings. ST9 determines the word conversion section to the completely matched word and the occurrence probability of the word by RA
This is a solution word storing step to be stored in M7.
Is a confirmation step for confirming whether or not all the words in the word dictionary 3 have been compared. ST11 is whether or not the section comparison has been completed for all the input phoneme strings included in the section for the predetermined n characters of the phoneme lattice. This is a section word comparison completion judgment step of judging whether or not each word is basically a solution of each input phoneme sequence and is stored in the RAM 7 together with each occurrence probability.

ST12 is a solution checking step for determining whether or not a word serving as a solution is stored in the RAM 7.
T13 is an unknown word generation step of generating an unknown katakana word when the katakana unknown word correction means 6 matches with reference to the n-gram dictionary 4 when there is no solution. As a result, the katakana unknown word is stored in the RAM 7 together with the appearance probability for a section where the words do not match (a section where a solution cannot be obtained in all input phoneme strings).

ST14 is a selection step of selecting and reading out a combination of conversion candidates having the maximum appearance probability from combinations of word strings and katakana unknown words for each section stored in the RAM 7, and ST15 is a selection step. This is an output step of generating a recognized character string by associating a character string corresponding to each divided section of the input phoneme divided string corresponding to the converted candidate.

Next, a specific example will be described in detail. When the user speaks "... Dictate" toward the microphone 1, a phoneme lattice (only a part is shown) as shown in FIG. 3 is generated. In the figure, “de”, “di”, etc. are input phonemes predicted based on the input, and are arranged in the order of the phonemes uttered sequentially in the horizontal direction (from left to right in the figure). The plurality of input phonemes arranged vertically are a plurality of input phonemes predicted based on the same sound, and the value attached to each connected line segment is the appearance probability of the phoneme. Each phoneme string formed by connecting these line segments is one input phoneme string (for example, “de, ku, w
a, ... "," de, ku, te, ... "," de, re,
ku, ... ").

Next, each input phoneme sequence of the phoneme lattice is compared with a word by prefix matching. For example, when words are stored in the word dictionary 3 as shown in FIG. 4, a match search with each of these words is performed. And, for example, "di, na,
To explain the input phoneme sequence of “a”, P (Y | W) is obtained by the following equation 5, and P (W) is the word appearance probability itself when n = 1, so that the predicted word probability is It is obtained as in the following Expression 6.

[0031]

(Equation 5)

[0032]

(Equation 6)

In the first embodiment, this calculation is performed for all words having the same phoneme sequence (for example, a match determination with "director", "disk", "detail", etc. and a probability calculation are performed). A candidate having a high probability is selected as a word conversion candidate. Here, for the sake of simplicity, all input phoneme strings are compared with all word phoneme strings. However, processing can be performed at high speed using a dynamic programming method such as Viterbi.

Next, a specific example of katakana unknown word processing when such a conversion candidate is not found will be described. For simplicity of explanation, "Y = di, ku, te, e, to,
An example of the input phoneme sequence “su, ru” will be described.

FIG. 5 is a diagram showing n-type semiconductor devices according to the first embodiment of the present invention.
5 is a data list showing a specific example of a part of the gram dictionary 4. In the figure, “#” indicates a pause, and is a specific example when n = 3. For example, the character 3-gram (conditional probability) in which the character "te" appears after the character "dic" is obtained by dividing the frequency of appearance of three characters of "dict" by the frequency of appearance of two characters of "dic". Approximate
Count as letters).

The word appearance probability of the input phoneme string can be obtained by a character string "dictate" according to the following equation (7). In the notation character string thus predicted, up to "dictate" is in katakana notation. Therefore, this range is regarded as katakana unknown word, and its appearance probability is obtained by the following equation (8). That is, the word “dictate” is replaced with the above “di, ku, te, e,
The conversion candidate of the section “to” is used, and the word appearance probability is treated as 2.95e−08 (= 2.95 × 10 ⁻⁸ ). If the probability of occurrence of a character string containing the word "dictate" is the highest, the character string containing the katakana unknown word is output as a recognized character string.

[0037]

(Equation 7)

[0038]

(Equation 8)

As described above, according to the first embodiment, the foreign words are divided into a fixed length together with the word dictionary 3 storing the words in association with the registered word character strings and the registered word phoneme strings. An n-gram dictionary 4 is provided for storing a foreign word segmented character string obtained in association with the foreign word segmented phoneme sequence, and the word predicting means 5 and the katakana unknown word correcting means 6 convert the input phoneme string into a registered word phoneme string. And the input code division sequence that divides the input phoneme sequence for each matching portion is generated by comparing the input code division sequence with the input word division phoneme sequence. There is an effect that can be.

In particular, when a foreign language is compared with a predetermined language,
In many cases, the way of chaining the sounds is different, but in the first embodiment, even in such a case, undefined foreign words are estimated at the time of word division based on speech. For example, even when compared with a case where a foreign language is estimated after being once converted into a character string of a predetermined language, the conversion error can be significantly suppressed, and an undefined foreign language can be highly estimated. effective.

Further, since the undefined foreign words are estimated at the time of the word segmentation based on the speech as described above, it is difficult to adversely affect the word segmentation in other input phoneme strings, and the speech including the undefined foreign words is reduced. Has the effect of preventing failure of the entire recognition process.

As a result, since foreign words rapidly increase or change as in recent Japanese, it is substantially difficult to register all foreign words in the word dictionary 3. Even in such a situation, there is an effect that speech including foreign words can be highly recognized.

According to the first embodiment, the katakana unknown word correcting means 6 compares the registered word phoneme sequence with all the input phoneme sequences with the foreign word segmented phoneme sequence, and performs the foreign word segmentation. Since the part that matches the phoneme sequence is divided,
There is an effect that speech including undefined foreign words can be quickly recognized without performing extra processing.

In the first embodiment, phonemes are used as speech codes. However, similar effects can be obtained by using phonemes or syllables as speech codes.
In addition, although the katakana notation character string of a foreign word is used as the foreign word division character string, the same effect can be obtained even if this is an alphabet notation character string.

Embodiment 2 FIG. 6 is a flowchart showing a speech recognition operation according to Embodiment 2 of the present invention.
In the figure, ST16 is an unknown word generation step of generating an unknown katakana word for each input phoneme sequence when the katakana unknown word correction means 6 matches with reference to the n-gram dictionary 4. As a result, a word or katakana unknown word is assigned to each phoneme sequence. Other configurations and operations are the same as those in the first embodiment, and a description thereof will be omitted.

As described above, according to the second embodiment, the foreign words are divided into a fixed length together with the word dictionary 3 storing the words in association with the registered word character strings and the registered word phoneme strings. An n-gram dictionary 4 is provided for storing a foreign word segmented character string obtained in association with the foreign word segmented phoneme sequence, and the word predicting means 5 and the katakana unknown word correcting means 6 convert the input phoneme string into a registered word phoneme string. And an input code division sequence that divides the input phoneme sequence for each matching part by comparing the input dictionary with the phonetic string divided by the foreign words.
The input code division sequence can be generated by estimating and generating a foreign word that is not registered in the Japanese language, and as a result, the number of foreign words suddenly increases or changes as in Japanese in recent years. Even in a situation where it is practically difficult to register all foreign words in the dictionary 3, there is an effect that speech including foreign words can be highly recognized.

According to the second embodiment, the katakana unknown word correcting means 6 generates a katakana unknown word for each input phoneme sequence and makes a selection between other words based on the appearance probability. Compared with the first embodiment, there is an effect that the processing can be performed with a smaller RAM capacity while performing the same high-speed speech recognition processing.

In the second embodiment, phonemes are used as speech codes. However, similar effects can be obtained by using phonemes and syllables as speech codes.
In addition, although the katakana notation character string of a foreign word is used as the foreign word division character string, the same effect can be obtained even if this is an alphabet notation character string.

Embodiment 3 FIG. 7 is a flowchart showing a voice recognition operation according to Embodiment 3 of the present invention.
In the figure, ST17 is an unknown word prediction step in which the katakana unknown word correcting means 6 determines whether or not a plurality of foreign word divided character strings are intermittently continuous in the word string of the solution.
Reference numeral 18 denotes an unknown word generation step for calculating a character string and its appearance probability when the katakana unknown word correction means 6 converts the entire intermittent continuous section into a foreign word divided character string,
ST19 is a solution correction output step in which the appearance probability of the unknown word is compared with the probability in the case of intermittent conversion, and the word string of the solution is corrected using the higher section probability and output to the output means 9 is there. Other configurations and operations are the same as those in the first embodiment, and a description thereof will be omitted.

Next, a specific example will be described. "Di, k
When an input phoneme sequence of “u, te, e, to” is input, for example, when words such as “di” and “te” are registered in the word dictionary 3, “di / ku / Erroneous recognition character string "te / to" is generated. Even in such a case, the katakana unknown word processing means converts the entirety of "di, ku, te, e, to" into "dictate" and calculates the appearance probability. As a result, it is possible to generate a recognition character string converted to “dictate” instead of “day / day / day / to”.

As described above, according to the third embodiment, the n-gram dictionary 4 stores the split character string appearance probability for each foreign word split character string, and When the input phoneme string is divided and a plurality of foreign word divided character strings are intermittently continuous, the probability that the entire intermittent continuous section is converted to the foreign word divided character string and the intermittent Since the input code division sequence is corrected using the one with the higher section probability by comparing the probability of conversion, the part that should be converted as an undefined foreign word is erroneously converted to a plurality of foreign word division character strings. Even if it is determined that the part is intermittently continuous, there is an effect that it can be converted into an original undefined foreign word.

Embodiment 4 FIG. In the unknown word generation step ST18 in FIG. 7, the same as in the third embodiment except that the appearance probability of a character string when the entire intermittent continuous section is converted into a foreign word divided character string is weighted in a predetermined manner. And the description is omitted.

As described above, according to the fourth embodiment, the katakana unknown word correcting means 6 performs correction weighting on the probability when the entire intermittent continuous section is converted into a foreign word divided character string. Since the section probabilities are compared above, even if the part that should be converted as an undefined foreign word is mistakenly determined to be a part where multiple foreign word divided character strings are intermittently continuous Has the effect that it can be converted into the original undefined foreign word.

Embodiment 5 FIG. FIG. 8 is a flowchart showing a speech recognition operation according to Embodiment 5 of the present invention.
In the figure, ST20 refers to the word dictionary 3 to determine the presence or absence of a word phoneme string including the same phoneme string as the section input phoneme string, and refers to the n-gram dictionary 4 to determine the same Generate matching katakana unknown words and RA
This is a batch processing step to be stored in M7, and ST21 is a batch selection step of selecting and reading out a combination of conversion candidates with the highest occurrence probability from combinations of words and katakana unknown words corresponding to a plurality of input phoneme strings. . Other configurations and operations are the same as those in the first embodiment, and a description thereof will not be repeated.

Next, a specific example will be described. For example, when an input phoneme sequence of “e, N, zi, N” is input and a word such as “circle, eNziN, 4e-07” is registered in the word dictionary 3, “circle formation” is performed based on the input phoneme sequence. And “engine” (undefined word) are generated. If the probability of occurrence of the undefined word "engine" is higher than the probability of occurrence of "circle", an undefined katakana character string is generated despite the fact that the same word is registered in the word dictionary 3. Can be.

As described above, according to the fifth embodiment, the word dictionary 3 stores the word appearance probabilities of each word, and the n-gram dictionary 4 stores the divided characters of each foreign word divided character string. The input code division sequence is generated so that the sequence appearance probability is stored and the overall appearance probability becomes highest when the input phoneme sequence is divided. For example, an undefined foreign word is registered in the word dictionary 3. Even when the pronunciation is the same as that of a word, there is an effect that the appearance probability of the undefined foreign word is compared with the occurrence probability of the word to convert the word into an undefined foreign word.

Embodiment 6 FIG. FIG. 9 is a block diagram showing a configuration of a speech recognition apparatus for Japanese input according to a sixth embodiment of the present invention. In the figure, reference numeral 10 denotes an example sentence storage means for storing a sentence. Reference numeral 11 denotes a portion that does not match the registered word phoneme sequence, and uses a foreign word in a portion that matches a foreign word in katakana notation characters extracted from the above example sentence. Means for generating katakana unknown words, and a katakana unknown word correction means (division means) for setting a portion of the n-gram dictionary 4 that matches the foreign word divided phoneme sequence as an unknown foreign word conversion section.
The other configuration is the same as that of the first embodiment, and the same reference numerals are given and the description is omitted.

Next, the operation will be described. FIG. 10 is a flowchart showing a speech recognition operation according to Embodiment 6 of the present invention. In the figure, in step ST22, if there is no solution, the katakana unknown word correcting means 11 first generates an unknown katakana word by referring to the example sentence, and then generates an n-gram if it does not match this example sentence. This is an example sentence unknown word generation step of generating a katakana unknown word when a match is made with reference to the dictionary 4. The other operations are the same as those of the first embodiment, and the same reference numerals are given and the description is omitted.

Next, a specific example will be described. For example, in the state where the example sentence "Apricot is sold by Mitsubishi Electric" is stored in the example sentence storage means 10, "a, p
When a phoneme sequence such as "u, ri, ko, Q, to" is input, the katakana unknown word correcting means 11 compares the input phoneme sequence with the katakana character string of the above example sentence converted into a phoneme sequence. Then, the input phoneme sequence is converted into an unknown katakana word “apricot”.

The word appearance probability of the katakana unknown word based on this example sentence may be calculated using, for example, the following equation (9). Further, in Expression 9, the denominator is the total number of read characters of the example sentence, but a simple multiplication of the number of characters of the example sentence by a predetermined coefficient may be used.

[0061]

(Equation 9)

As described above, according to the sixth embodiment, the example sentence storage means 1 for storing a sentence in a predetermined language.
0, and the katakana unknown word correcting means 11 divides the input phoneme string at a part corresponding to the word extracted from the example sentence. Therefore, a word not previously registered in the word dictionary 3 or a foreign word divided character string storage means The user can freely learn undefined foreign words, such as coined words and technical terms, that cannot be fully estimated in the sentence form without damaging the estimation function.
There is an effect that it can be recognized by voice.

In the sixth embodiment, the first embodiment is used.
Although the example in which the example sentence storage unit 10 is provided has been described on the premise of the above, the same effect can be obtained also on the premise of the second and fifth embodiments.

[0064]

As described above, according to the present invention, a word dictionary for storing a word in association with a registered word character string and a registered word speech code string and a foreign word are divided into a predetermined length. A foreign word segmented character string storage means for storing the foreign word segmented character string obtained by the above in association with the foreign word segmented speech code string is provided, and the dividing means stores the input speech code string in the registered word speech code string and the foreign word segment. Compared with the word segmentation speech code string,
Since the input code division sequence for dividing the input speech code sequence for each matching portion is generated, there is an effect that a foreign word not registered in the word dictionary is estimated and generated to generate the input code division sequence.

In particular, a foreign language often differs from a predetermined language in the manner of chaining sounds, but in the present invention, even in such a case, undefined words are not defined in word division based on speech. Since the foreign words are estimated, even if it is compared with, for example, a case where the foreign words are estimated after being once converted into a character string of a predetermined language, the conversion error can be remarkably suppressed, and the undefined There is an effect that a foreign word can be highly estimated.

In addition, since the undefined foreign words are estimated at the time of speech-based word segmentation, the word segmentation in other input speech code strings is not adversely affected.
When a speech including an undefined foreign word is recognized, there is an effect that it is possible to prevent a failure in the entire recognition processing.

As a result, it is substantially difficult to register all the foreign words in the word dictionary because the foreign words suddenly increase or change as in Japanese in recent years. Also, there is an effect that speech including foreign words can be highly recognized.

According to the present invention, the dividing means compares the part which does not coincide with the registered word speech code string with the foreign word divided speech code string and divides the part which coincides with the foreign word divided speech code string. There is an effect that a voice including an undefined foreign word can be quickly recognized without performing a complicated process.

According to the present invention, the word dictionary stores the word appearance probability for each word, and the foreign word divided character string storage means stores the divided character string appearance probability for each foreign word divided character string. Since the dividing means generates the input code division sequence so that the overall appearance probability becomes highest when the input speech code sequence is divided, for example, the undefined foreign word has the same pronunciation as the word registered in the word dictionary. Even in such a case, there is an effect that the appearance probability of the undefined foreign word is compared with the occurrence probability of the word, and conversion to the undefined foreign word can be performed.

According to the present invention, the foreign word divided character string storage means stores the divided character string appearance probability for each foreign word divided character string, and the dividing means stores a plurality of divided character strings when the input speech code string is divided. If the foreign word segmentation string is intermittent,
By comparing the probability of converting the entire intermittent continuous section into a foreign word segmented character string and the probability of intermittently converting it, an input code division sequence is generated using the higher one of the section probabilities. Therefore, even if the dividing means once mistakenly determines that a part that should be converted as an undefined foreign word is a part where a plurality of foreign word divided character strings are intermittently continuous, Can be converted into the original undefined foreign word.

According to the present invention, the dividing means compares the section probabilities after weighting the probabilities obtained when the entire intermittent continuous section is converted into the foreign word divided character string with correction weights.
Even if the dividing means mistakenly determines that the part that should be originally converted as an undefined foreign word is a part where a plurality of foreign word divided character strings are intermittently continuous, There is an effect that it can be converted into an original undefined foreign word.

According to the present invention, there is provided an example sentence storage means for storing a sentence in a predetermined language, and the dividing means divides the input speech code string at a portion corresponding to a word extracted from this example sentence. The user can freely learn additional words, such as unregistered words and undefined foreign words such as coined words and technical terms that cannot be estimated by the foreign word segmented character string storage means, without impairing the estimation function. There is an effect that voice recognition can be performed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a speech recognition device for Japanese input according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing a speech recognition operation according to the first embodiment of the present invention.

FIG. 3 is an explanatory diagram of a phoneme lattice according to the first embodiment of the present invention.

FIG. 4 is a data list showing a specific example of a part of the word dictionary according to the first embodiment of the present invention.

FIG. 5 is an n-gram according to the first embodiment of the present invention.
6 is a data list showing a specific example of a dictionary.

FIG. 6 is a flowchart showing a speech recognition operation according to Embodiment 2 of the present invention.

FIG. 7 is a flowchart showing a speech recognition operation according to Embodiment 3 of the present invention.

FIG. 8 is a flowchart showing a voice recognition operation according to Embodiment 5 of the present invention.

FIG. 9 is a block diagram showing a configuration of a speech recognition device for Japanese input according to a sixth embodiment of the present invention.

FIG. 10 is a flowchart showing a speech recognition operation according to Embodiment 6 of the present invention.

FIG. 11 is a block diagram showing a configuration of a conventional voice recognition device.

[Explanation of symbols]

1 microphone (preprocessing means), 2 phoneme probability calculating means (preprocessing means), 3 word dictionary, 4n-gram dictionary (foreign word divided character string storage means), 5 word prediction means (dividing means), 6, 11 Katakana unknown word correction means (division means), 8 selection means (character string generation means), 9 output means (character string generation means), 10 example sentence storage means.

Claims (9)

[Claims] 1. A pre-processing means for inputting speech in a predetermined language and generating an input speech code string in the language based on the speech, a word in the language, a registered word character string and a registered word speech code A word dictionary that is stored in association with a sequence, and a foreign word segmented character string that stores a foreign word segmented character string obtained by dividing a foreign word into a fixed length in association with the foreign word segmented speech code string. Storage means, and a dividing means for comparing the input speech code string with the registered word speech code string and the foreign word segmented speech code string, and generating an input code segment sequence for dividing the input speech code string for each matching portion; A speech recognition apparatus comprising: a character string generation unit that generates a recognition character string by associating character strings corresponding to respective divided portions of an input code division string. 2. Pre-processing means for receiving a voice in a predetermined language, generating a plurality of input voice code strings in the language together with respective input appearance probabilities based on the voice, and registering words in the language. A word dictionary that stores word character strings, registered word speech code strings and word appearance probabilities in association with each other, and a foreign word segmented character string obtained by dividing a foreign word into a certain length, the foreign word segmented speech code string And a foreign word divided character string storage unit that stores the input speech code string in association with the divided character string appearance probability, and for each input speech code string, stores the input speech code string with the registered word speech code string and the foreign word divided speech code string. Comparing means for generating an input code division sequence for dividing an input speech code sequence for each matching portion; and inputting character codes based on an input appearance probability, a word appearance probability, and a divided character string appearance probability. A character string generating means for obtaining a total appearance probability of the divided sequence and generating a recognized character string by associating a character string corresponding to each divided portion of the input code divided sequence with the combination having the highest appearance probability Voice recognition device. 3. The phonetic code is a phoneme in a predetermined language,
3. The speech recognition device according to claim 1, wherein the speech recognition device is a phoneme or a syllable. 4. The speech recognition apparatus according to claim 1, wherein the foreign word division character string is a katakana notation character string or an alphabet notation character string of a foreign word. 5. The method according to claim 1, wherein the dividing unit compares a part that does not match the registered word speech code string with the foreign word divided speech code string, and divides the part that matches the foreign word divided speech code string. The speech recognition device according to claim 1 or 2. 6. The word dictionary stores the word appearance probability for each word, and the foreign word divided character string storage means stores the divided character string appearance probability for each foreign word divided character string. 3. The speech recognition apparatus according to claim 1, wherein, when the input speech code string is divided, the input code division string is generated such that the overall appearance probability is highest. 7. A foreign word divided character string storage means stores a divided character string appearance probability for each foreign word divided character string, and the dividing means stores a plurality of foreign word divided character strings when an input speech code string is divided. If the character string is intermittently continuous, the probability that the entire intermittent continuous section is converted into a foreign word segmented character string is compared with the probability that the intermittent conversion is performed intermittently. 7. The speech recognition device according to claim 5, wherein an input code division sequence is generated using a higher one. 8. The dividing means according to claim 7, wherein the dividing means compares the section probabilities after applying correction weighting to the probabilities when the entire intermittent continuous section is converted into the foreign word divided character string. Voice recognition device. 9. An apparatus according to claim 1, further comprising an example sentence storing means for storing a sentence in a predetermined language, wherein the dividing means divides the input speech code string at a portion corresponding to a word extracted from the example sentence. Any one of claims 8
The speech recognition device according to the item.