JP2003295888A - Speech recognition device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To obtain a speech recognition device and a program therefor which make it possible to vocally input a symbol such as a punctuation mark in a document vocal input device. <P>SOLUTION: The device is equipped with a symbol character grammar storage means 101 which acoustically analyzes an input speech and stores symbol character grammar as a rule of connection between a word and the symbol character and a collating means 1102 which selects an acoustic pattern between a word acoustic pattern and a symbol character acoustic pattern according to the symbol character grammar, collates the acoustic pattern with a speech feature quantity, and outputs a word or symbol character as to a matching acoustic pattern as a speech recognition result as to the input speech. <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a voice recognition program for recognizing a voice uttered by a human (user) and displaying a recognition result.

[0002]

2. Description of the Related Art Speech recognition is a technology in which a machine automatically extracts linguistic information from a human voice, and is highly practical such as document creation by voice input. Regarding the speech recognition device, for example, "Speech Information Processing" by Sadahiro Furui, June 1998,
Chapter 5 of Morikita Publishing (hereinafter referred to as Reference 1) and “Basics of Speech Recognition” L.A. RABINER, B.R. H. Translated by JUANG, Sadahiro Furui, November 1995, NTT Advanced Technology (hereinafter referred to as Reference 2) and "Speech recognition by probabilistic model" by Seiichi Nakagawa, July 1988, Corona Publishing (hereinafter referred to as Reference 3) ) For details.

A voice recognition device described in Document 1 as a conventional example will be described below with reference to a block diagram shown in FIG.

In FIG. 14, reference numeral 1001 denotes a user who uses the voice recognition device, and 1002 denotes a user 1001.
It is the input voice uttered by. A voice recognition unit 1003 extracts a voice feature amount from the input voice 1002, and performs matching using a standard pattern of a recognition target of a predetermined word or sentence. Reference numeral 1004 denotes a voice recognition result collated by the voice recognition means 1003, which is normally output as text. A voice recognition display unit 1005 displays the voice recognition result 1004 output by the voice recognition unit 1003.

FIG. 15 is a block diagram for explaining the inside of the voice recognition means 1003. In FIG. 15, 1
Reference numeral 101 denotes a voice feature amount extraction means, which is an input voice 100.
The second audio signal is A / D-converted, the A / D-converted signal is cut out in frames at constant time intervals of about 5 ms to 20 ms, and acoustic analysis is performed to extract the audio feature amount. 1
A recognition target word dictionary 105 stores the notation and reading of a predetermined recognition target word. Also, 110
Reference numeral 6 is a standard pattern table, which stores standard patterns in subword units and their label notations. Reference numeral 1104 denotes a word acoustic pattern generation means, which is a recognition target word dictionary 1
The reading of the recognition target word stored in 105 is converted into the label notation of the subword voice unit, the standard pattern of the subword voice unit corresponding to the label notation is extracted from the standard pattern table 1106, and this standard pattern is recognized. The acoustic pattern of the recognition target word is generated by connecting the sub-word voice units that constitute the reading of the. 1103 is a normal grammar, which stores rules for connecting words to each other.
Here, the normal grammar 1103 includes only connection rules between words, and does not include connection rules between symbol characters such as “,” and “.” And words. Reference numeral 1102 denotes a collating unit, which selects an acoustic pattern from the acoustic patterns of the plurality of recognition target words generated by the word acoustic pattern generating unit 1104 according to the normal grammar 1103, and outputs the acoustic pattern and the speech feature amount extracting unit 1101. The generated speech feature amount is collated, and the word corresponding to the matching acoustic pattern and the document in which the word is connected are output as the speech recognition result 1004.
In the following description, the matching unit 1102 matches the speech feature amount and the acoustic pattern of the recognition target word according to the grammar, that is, the matching unit 1102 selects the recognition target word from the context of the words specified in the grammar. It means narrowing down and then selecting the acoustic pattern of the recognition target word to be compared with the acoustic feature amount.

Next, the operation will be described. When the input voice 1002 is input, the voice feature amount extraction means 110
1 performs acoustic analysis to extract the voice feature amount of the input voice 1002. Further, the word acoustic pattern generation means 1104 includes a recognition target word dictionary 1105 and a standard pattern table 11.
From 06, a recognition target word acoustic pattern is generated. Next, the collation unit 1102 collates the speech feature amount with the recognition target word acoustic pattern generated by the word acoustic pattern generation unit 1104 using the normal grammar 1103, and outputs a speech recognition result 1004.

As described above, the conventional speech recognition apparatus recognizes the uttered voice by matching the voice feature amount of the input voice uttered by the user with a predetermined recognition target word by using the normal grammar. Is displayed.

[0008]

When creating a document, it is necessary to appropriately use symbol characters such as the punctuation mark "," and the punctuation mark ".". However, in the document input by the conventional voice recognition device, even if you say "Toten" to input these symbol characters, the voice recognition result becomes "reading point",
There is a problem that "," is not recognized. It is also possible to set “,” in advance as the recognition result of the utterance “Toten”, but if you do so, the recognition result of homonyms such as “,”, “reading point”, and “our shop” will be listed. It is necessary to select an optimum word from the candidates, which makes the selection operation complicated. SUMMARY OF THE INVENTION It is an object of the present invention to solve the above problems and to obtain a voice recognition device and a voice recognition program in which a symbol character can be easily input in the voice recognition device.

[0009]

A voice recognition apparatus according to the present invention comprises a voice feature quantity extraction means for acoustically analyzing an input voice and extracting a voice feature quantity indicating a feature of the voice, and a voice stored in a standard pattern table. The word acoustic pattern generation means for generating a word acoustic pattern for this word from the standard pattern for obtaining the acoustic score in recognition and the word stored in the recognition target word dictionary, and the dictionary for notation and reading of symbol characters are stored. A symbol character acoustic pattern generating means for generating a symbol character acoustic pattern for this symbol character from the symbol character and the standard pattern, and a symbol character grammar storing means for storing a symbol character grammar which is a connection rule between a word and a symbol character, According to this symbolic grammar, an acoustic pattern is selected from the word acoustic pattern and the symbolic character acoustic pattern, and the acoustic pattern and the speech feature amount are selected. A word or symbol characters for acoustic pattern matching engaged those comprising matching means for outputting a speech recognition result for said input speech.

Further, in the voice recognition device according to the present invention, the voice recognition device stores a normal grammar storing means for storing a normal grammar which is a connection rule between words, and one of the normal grammar and the symbol character grammar. A grammar changeover switch for selecting a grammar is further provided, and when the grammar changeover switch selects the normal grammar, the collating means selects an acoustic pattern from the word sound pattern according to the normal grammar to select the voice. It is configured such that a word about an acoustic pattern that matches the feature amount and matches is output as a voice recognition result that does not include a symbol character for the input voice.

Further, the voice recognition apparatus according to the present invention comprises a voice feature amount extraction means for acoustically analyzing an input voice to extract a voice feature amount indicating a voice feature, and a standard for obtaining an acoustic score in voice recognition. According to this normal grammar, a word sound pattern generation means for generating a word sound pattern for this pattern from the words stored in the pattern and recognition target word dictionary, a normal grammar storage means for storing a normal grammar which is a connection rule between words, Connection between the word and the symbol character, and a matching means for matching the sound pattern selected from the word sound pattern with the sound feature amount and outputting the word for the matched sound pattern as an intermediate speech recognition result for the input sound. A symbol character grammar storing means for storing a symbol character grammar which is a rule, and a symbol character is inserted into the intermediate speech recognition result according to this symbol character grammar. Generating a document containing the symbol characters Te and those having the mark character inserting means for outputting the document.

Further, in the voice recognition device according to the present invention, the voice feature amount extracting means acoustically analyzes the input voice including the utterance of the symbol character to extract the voice feature amount of the utterance of the symbol character, and the collating means, An intermediate speech recognition result including a collation result portion for the utterance of a symbol character is output, and the symbol character inserting means identifies the collation result portion according to the symbol character grammar and replaces the collation result portion with a symbol character to generate a symbol. It is configured to generate a document including characters.

Further, in the voice recognition apparatus according to the present invention, the voice feature amount extraction means acoustically analyzes an input voice including a pause time of a constant interval to extract a voice feature amount including information about a pause position, and the symbol is used. The character insertion means is configured to identify the pause position according to the symbol character grammar and insert the symbol character at the pause position to generate a document including the symbol character.

In the voice recognition device according to the present invention, the voice feature amount extraction means extracts prosody information from the input voice,
The symbol-character grammar storage means stores the relation between prosody information and symbol characters in the symbol-character grammar, and the symbol-character insertion means stores the intermediate according to the prosody information extracted by the phonetic feature extraction means and the symbol-character grammar. The configuration is such that a symbol character is inserted into a voice recognition result and a document including the symbol character is generated.

Further, the speech recognition apparatus according to the present invention has a normal grammar storing means for storing a normal grammar which is a connection rule between words, and a grammar switching for selecting one of the normal grammar and the symbol character grammar. A switch is further provided, and the symbol character inserting means outputs the intermediate speech recognition result as a document containing no symbol character when the grammar change switch selects the normal grammar.

A voice recognition program according to the present invention is
The voice feature quantity extraction procedure for acoustically analyzing the input voice and extracting the voice feature quantity indicating the feature of the voice, and the standard pattern for recognizing the acoustic score in the voice recognition stored in the standard pattern table and the recognition target word dictionary are A procedure for generating a word acoustic pattern of this word from a stored word, and a symbol for generating a symbolic acoustic pattern of this symbolic character from the symbolic character stored in the dictionary for notation and reading of symbolic characters and the above standard pattern. A character-acoustic pattern generation procedure, a symbol-character grammar storage procedure that stores a symbol-character grammar that is a connection rule between a word and a symbol character, and an acoustic pattern is selected from the word acoustic pattern and the symbol-character acoustic pattern according to this symbol-character grammar. This acoustic pattern is collated with the above-mentioned speech feature amount, and a word or symbol character for a matching acoustic pattern is added to the input speech. Is intended to execute the the verification procedure is output as the speech recognition result to the computer.

The voice recognition program according to the present invention is
The computer further executes a normal grammar storing procedure for storing a normal grammar, which is a connection rule between words, and a grammar switching procedure for selecting one of the normal grammar and one of the above symbol character grammars. , If the grammar selection switch selects the normal grammar, a sound pattern is selected from the word sound patterns according to the normal grammar, the sound pattern is compared with the sound feature amount, and a word having a matching sound pattern is input to the input sound. Is output as a speech recognition result that does not include the symbol character of.

A voice recognition program according to the present invention is
A speech feature quantity extraction procedure that performs acoustic analysis of the input speech to extract a speech feature quantity indicating a speech feature, and a standard pattern for obtaining an acoustic score in speech recognition and a word stored in the recognition target word dictionary. The word sound pattern generation procedure for generating the word sound pattern, the normal grammar storage procedure for storing the normal grammar that is the connection rule between words, the sound pattern selected from the word sound patterns according to the normal grammar, and the speech feature amount. And a matching character string grammar storing procedure for storing a word for the matched acoustic pattern as an intermediate speech recognition result for the input speech and a symbol character grammar that is a connection rule between the word and the symbol character. And a symbol that inserts a symbol character into the intermediate speech recognition result according to this symbol character grammar to generate a document containing the symbol character and outputs this document. It is intended to execute the character insertion procedure into the computer.

The voice recognition program according to the present invention is
The voice feature amount extraction procedure acoustically analyzes an input voice including a utterance of a symbol character to extract a voice feature amount about the utterance of a symbol character, and the matching procedure includes an intermediate result including a match result portion about the utterance of a symbol character. The speech recognition result is output, and the symbol character insertion procedure is configured to identify the matching result portion according to the symbol character grammar and replace the matching result portion with the symbol character to generate a document including the symbol character. is there.

The voice recognition program according to the present invention is
The voice feature amount extraction procedure acoustically analyzes an input voice including a pause time at a constant interval to extract a voice feature amount including information about a pause position, and the symbol / character insertion procedure determines the pause position by the symbol / character grammar. According to the above-mentioned method, the symbol character is inserted into the pause position to generate a document including the symbol character.

A voice recognition program according to the present invention is
The voice feature amount extraction procedure extracts prosody information from the input voice, the symbol character grammar storage procedure stores the relation between the prosody information and the symbol characters in the symbol character grammar, and the symbol character insertion procedure According to the prosody information extracted by the speech feature extracting means and the symbol character grammar, symbol characters are inserted into the intermediate speech recognition result to generate a document including a symbol document.

A voice recognition program according to the present invention is
The computer is further caused to execute a normal grammar storing procedure for storing a normal grammar which is a connection rule between words and a grammar switching procedure for selecting one of the normal grammar and one of the above symbol character grammars to insert the above symbol characters. When the grammar switching procedure selects the normal grammar, the procedure outputs the intermediate speech recognition result as a document containing no symbol characters.

[0023]

DETAILED DESCRIPTION OF THE INVENTION

First Embodiment FIG. 1 is a block diagram showing the configuration of a voice recognition device according to a first embodiment of the present invention. In FIG. 1, reference numeral 101 denotes a symbolic character grammar, which stores connection rules for words and symbolic characters. Rules for connecting words and symbols are described in, for example, Chapter 2 of Document 3, P. Finite state automata described in No. 21, N-gram and HMM
It is expressed by a statistical language model such as (Hidden Markov Models). For the statistical language model, see “Statistical Language Model” Kenji Kita, The University of Tokyo Press, November 1999 (hereinafter referred to as Reference 4).
Have been introduced in detail.

Reference numeral 102 is a symbol character input selection signal,
This is a signal transmitted when the user causes the voice recognition means to input a symbol character. For example, when the user presses a specific key (for example, the Shift key of the keyboard) of the terminal used to create the document for selecting the symbol character input, the symbol character input selection signal 102 is generated and the grammar changeover switch 103 is generated. Is transmitted. In addition, when a specific key of the terminal is released (pressing is stopped), the symbol / character input selection signal 102 is generated and transmitted to the grammar changeover switch 103, and when it is pressed, the symbol / character input selection signal 102 is transmitted. There may be no configuration.

Reference numeral 103 denotes a grammar changeover switch, and when the symbol character input is selected and the symbol character input selection signal 102 is transmitted, the grammar changeover switch 103 of FIG. The grammar used in 1102 is the symbol character grammar 101. When the user does not input the symbol character and the symbol character input selection signal 102 is not transmitted, the grammar change switch 103 is connected to the contact A, so that the grammar used by the collating means 1102 is the normal grammar 11
It becomes 03.

Reference numeral 105 is a dictionary of notation and reading of symbol characters, which stores predetermined notation and reading of symbol characters. For example, as shown in FIG. 4, the notation and reading of symbol characters are stored. . Here, the symbol character refers to a character that is not normally read out, such as ".", "," Or "?". As the reading of the symbol characters in the dictionary 105 for notation and reading of the symbol characters, it is preferable to assign in advance a symbol that allows the user to easily associate the symbol. In the present embodiment, the reading of “.” Is “kuteN” and the reading of “,” is “tooteN”. It is also possible to assign one reading to a plurality of symbol character strings. For example, it is possible to register the reading of the symbol character string “(^ _ ^)” as “kao” (face).

Reference numeral 104 denotes a symbol / character acoustic pattern generation means, which generates a symbol / character acoustic pattern from the symbol / character reading and standard pattern table 1106 stored in the symbol / character notation and reading dictionary 105. It should be noted that, hereinafter, the same functional blocks as those of FIG. 14 and FIG.

Next, the operation will be described with reference to FIG. First, in step ST101, the word acoustic pattern generation means 1104 generates a word acoustic pattern using the recognition target word dictionary 1105 and the standard pattern table 1106. The recognition target word dictionary 1105 stores the notation and reading of a predetermined recognition target word. Here, the notation is text represented by kanji, kana, numbers, and the like, and the reading means pronunciation when reading kanji, kana, and numbers, and is represented by phoneme symbols and the like. For example, in the case where news speech is the recognition target, the recognition target word stores a combination of the notation and its reading (eg, the reading "koQkaigiiN" for the notation "diet of parliament") as shown in FIG. . Further, the standard pattern table 1106 stores standard patterns for obtaining an acoustic score in voice recognition. The standard pattern is, for example, a pattern [λl1, λl2, ..., λlM] (l1, l2,
... lM is a label name, M is the total number of labels), and H is obtained by performing parameter learning on speech data of many speakers who normally utter.
MM (Hidden Markov Models) is used.

A method of generating a word acoustic pattern will be described by taking a case of using a standard pattern of a sub-word voice unit (a voice piece unit such as a phoneme or a syllable) as an example. First, the word acoustic pattern generation means 1104 uses the recognition target word dictionary 1105.
The reading of the recognition target word memorized by [wr (1), wr (2),
..., wr (N)] (word numbers in parentheses) are converted to label representations in subword voice units. Next, the word acoustic pattern generation means 1104 selects a standard pattern of a subword voice unit corresponding to the label description of the subword voice unit from the standard patterns stored in the standard pattern table 1106, and the label representation in the word. , Each standard pattern is connected in the order of appearance, and word acoustic patterns [Λ (1), Λ (2), ...,
Λ (N)] (word numbers in parentheses) are generated.

A method of generating the word acoustic pattern Λ (n) will be described by taking a standard pattern in which phonemes depending on the front-back environment are used as subword voice units. As the n-th recognition target word stored in the recognition target word dictionary 1105, “Tomorrow (as
u) ”exists. in this case,
In continuous word speech recognition in which words are connected before and after "tomorrow", "tomorrow" is expressed as / $ asu * / in the phoneme sequence. Note that $ is the last phoneme of the preceding word, and * is the beginning phoneme of the following word. In this example, the sub-word voice unit label is (a)
The central phoneme is / a /, the preceding phoneme is the last phoneme $ of the preceding word, the subsequent phoneme is the label {$ as}, and (b) the central phoneme is / s.
/ With a leading phoneme of / a / and a subsequent phoneme of / u /
It can be converted into {asu} and (c) a label {su *} in which the central phoneme is / u /, the preceding phoneme is / s /, and the following phoneme is the first phoneme / * / of the following word. Standard patterns λ $ as, λasu, and λsu * corresponding to the subword voice unit label are extracted from the standard pattern table 1106, and a standard pattern Λ obtained by concatenating them is extracted.
(n) becomes the standard pattern for the word "tomorrow". Recently, a speech recognition system has been studied, in which a subword speech unit standard pattern of phonemes depending on the surrounding phoneme environment is used and the number of words to be recognized is tens of thousands or more.

Next, in step ST102, the symbol / character acoustic pattern generation means 104 generates a symbol / character acoustic pattern using the symbol / character notation and reading dictionary 105 and the standard pattern table 1106.

The method of generating the symbolic character acoustic pattern in step ST102 will be described by taking the case of using the standard pattern of the subword voice unit as an example. In the symbol character acoustic pattern generation means 104, the reading of symbol characters [wr '(1), wr' set in the dictionary 105 for notation and reading of symbol characters is performed.
(2), ..., wr '(M)] (word numbers in parentheses) are converted into sub-word voice unit label notation, and standard pattern table 1
A standard pattern of subword voice units corresponding to the label is selected from the standard patterns stored in 106, and the standard patterns are concatenated in the order in which each subword voice unit appears in reading of the symbol character, whereby the symbol character acoustic pattern [Λ '( 1),
Λ '(2), ..., Λ' (M)] (word numbers in parentheses) are generated.

The method of generating the symbolic character acoustic pattern Λ '(n) will be described by taking a standard pattern in which phonemes depending on the surrounding environment are used as subword speech units. As the nth symbol character stored in the dictionary 105 for notation and reading of symbol characters, “.
(KuteN) ”exists. In this case, in continuous word speech recognition in which words are connected before and after ".", "." Is represented as / $ kuteN * / in the phoneme sequence. Note that $
Is the last phoneme of the preceding word and * is the first phoneme of the following word. In this example, the label of the sub-word phonetic unit is (a) the central phoneme is / k /, the preceding phoneme is the last phoneme $ of the preceding word,
Label {$ ku} with subsequent phoneme / u /, and (b) label {kut} with central phoneme / u /, preceding phoneme / k /, and subsequent phoneme / t /.
And (c) the central phoneme is / t /, the preceding phoneme is / u /, the subsequent phoneme is / e / the label {ute}, and (d) the central phoneme is / e / and the preceding phoneme is / t. /, Label {teN} whose subsequent phoneme is / N /, (e) the central phoneme is / N /, the preceding phoneme is / e /, and the subsequent phoneme is the first phoneme of the following word / * / Label {eN * } And can be converted to. Standard pattern λ $ ku, λku corresponding to this subword voice unit label
Standard pattern table 110 for t, λute, λteN, λeN *
The standard pattern Λ '(n) extracted from No. 6 and connecting them becomes the standard pattern of the symbol character ".".

In step ST103, the audio feature amount extraction means 1101 converts the audio signal of the input audio 1002 into A /
The D-converted and A / D-converted signal is cut out in frames with a fixed time interval of about 5 ms to 20 ms, acoustic analysis is performed, and a voice feature vector O = [o (1), o (2) , ..., o
(T)] (T is the total number of frames). Here, the voice feature amount is a feature amount vector that can express a voice feature with a small amount of information, and is, for example, a feature amount vector configured by a physical amount of a dynamic feature of the cepstrum. The voice feature amount is described in detail in, for example, 5-2 sound processing of Document 1.

In step ST104, the grammar changeover switch determines whether or not the user selects the symbol / character input by the presence / absence of the symbol / character input selection signal 102. When the user selects the symbol character input, the symbol character input selection signal 1
02 is transmitted, and the grammar change switch 103 selects a symbol character grammar. Therefore, in this case, step ST
In 106, the collation means 1102 determines that the symbolic character grammar 1
01 is used to select a sound pattern from the symbolic character sound pattern and the word sound pattern, and this sound pattern is compared with the sound feature amount output from the sound feature amount extraction means 1101 to correspond to the matched sound pattern. The symbol characters or the recognition target words are sequentially output as a speech recognition result 1004 in a text format.

Next, the matching method in step ST106 will be described in detail. The collating unit 1102 outputs the voice feature amount output from the voice feature amount extracting unit 1101.
With respect to O, the word sequence W ′ of the recognition result is extracted by Expression 1. In Equation 1, the first term P (O | W) is the acoustic probability. Here, the word acoustic pattern generation means 110
The word acoustic pattern [Λ (1), Λ (2),
, Λ (N)] and the symbol character acoustic pattern [Λ '(1), Λ' (2), ..., which is the output of the symbol character acoustic pattern generation means 104.
Λ '(M)] is connected according to the rules for connecting words and symbol characters defined in the symbol character grammar 101 to assume the word sequence W and evaluate the probability P (O | W). Recently, in order to calculate acoustic probability, HMM (Hid
Den Markov Models) are often used. The method of calculating the acoustic probability is described in detail in Documents 1, 2, and 3.

[0037]

[Formula 1]

The second term P (W) is a linguistic probability, and is the occurrence of a sequence W of words and symbols assumed according to the rules for connecting words and symbol characters defined in the symbol-character grammar 101. It represents the probability. Recently, in order to obtain the linguistic probability, a statistical language model that gives the probability of word chain is often used. For details of the statistical language model, refer to Reference 4.

Next, the symbolic character grammar 101 is N-gram.
A method of obtaining the occurrence probability of a sequence including a symbol character will be described by taking a case of a model as an example. The N-gram model is one of statistical language models. N-gram
The probability of occurrence of L words (represented by Equation 2) according to the model (represented by Equation 3) is given by Equation 5.
As shown in Equation 5, N-gram is the immediately preceding (N-1)
The conditional probability that a current word occurs from a word (represented by Equation 4) is held as a parameter, and the probability of a word string is calculated. For this parameter, the probability is generally calculated in a large amount of text corpus. N is 2 2-
In the case of gram, as a parameter of the symbolic character grammar 101, for example, conditional probability Ps (, (tooteN) | I am (watasiw
a)), Ps (. (kuteN) | go (iku)) etc. are held.

[0040]

[Formula 2]

[0041]

[Formula 3]

[0042]

[Formula 4]

[0043]

[Formula 5]

A sequence including symbol characters, for example, "Go to the station.
The occurrence probability by 2-gram of "(ekie iku kuteN)" is calculated by the following formula 6. In addition, Formula 6
In, # represents the beginning and end of a sentence.

[0045]

[Formula 6]

The matching means 1102 converts the text notation of the word sequence W'obtained by Equation 1 into the speech recognition result 100.
Output as 4.

Since the symbol / character input selection signal 102 is not transmitted when the user 1001 does not select the symbol / character input, the grammar selection switch selects the normal grammar in step ST104. As a result, step ST105
In the above, the matching means 1102 selects an acoustic pattern from the word acoustic pattern generated in step ST 101 using the normal grammar 101, collates this acoustic pattern with the voice feature amount, and recognizes the recognition target word corresponding to the matched acoustic pattern. Are sequentially output as a speech recognition result 1004 in a text format.

The normal grammar 1103 is a recognition target word dictionary 1
The connection rule between words stored in 105 is stored. The rules of connection between words are finite state automata, N-gram and HMM (Hidden Marko).
v Models). N-gram, which is one of the statistical language models
When (N = 2) is used, the conditional probability P (To station | I), P (Go | To station), etc. is a normal grammar 11 as a parameter.
It is stored in 03.

Next, the matching method will be described in detail. The voice feature amount extraction means 1101 extracts an acoustic feature amount, and the feature amount vector of this acoustic feature amount is set to O. The matching means 1102 extracts the word sequence W ′ of the recognition result by the mathematical expression 1 with respect to this feature amount vector O. In Expression 1, the first term P (O | W) is an acoustic probability, and is a word acoustic pattern [Λ (1), Λ (2), ..., Which is an output from the word acoustic pattern generation means 1104. Λ (N)] and regular grammar 110
The probability is evaluated assuming the word sequence W according to the connection rules between words defined in 3. HMM (Hidden Markov M) is used to calculate the acoustic probability.
Often used.

The second term P (W) in the mathematical expression 1 is a linguistic probability, and represents the probability of the word sequence W assumed according to the connection rule between words defined in the normal grammar 1103. In order to obtain such a linguistic probability, a statistical language model that gives the probability of word chain is often used.

Sequences containing symbolic characters, such as "go to station
(ekie iku) ”, the occurrence probability by 2-gram is calculated as in the following formula 7.

[0052]

[Formula 7]

The matching means 1102 converts the text notation of the word sequence W'obtained by Equation 1 into the speech recognition result 100.
Output as 4.

In step ST107, it is determined whether or not the utterance has ended. If the utterance has not ended, the process returns to step ST103 and is repeated. on the other hand,
If the utterance is finished, the process is finished.

Although the present embodiment has been described in the case where the selection / non-selection of the symbol / character input is switched for each sentence, the symbol / character input selection signal 102 may be switched for each word or symbol / character. For example, "I go to the station. (Wa
tasiwa tooteN ekie iku kuteN ) ”, the symbol character input selection signal 10 is underlined.
If 2 is transmitted, the occurrence probability in this case is
Calculated by Conditional probability Ps in Equation 8
(• | •) is stored in the symbolic character grammar 101. The conditional probability P (・ | ・) is the normal grammar 110
3 is stored.

[0056]

[Formula 8]

In the first embodiment, the voice feature quantity extracting means 1101, the matching means 1102, the word acoustic pattern generating means 1104, and the symbolic character acoustic pattern generating means 104.
May be configured by hardware, but a voice recognition program for performing these processes may be created and the computer may execute the voice recognition program.

As described above, according to the speech recognition apparatus of the first embodiment, when a symbol character is input, the user selects the symbol character input and utters the reading of the symbol character to input the symbol character. As a result, the document can be efficiently created.

[0059]

[Embodiment 2] FIG. 5 is a block diagram showing a configuration of a speech recognition apparatus according to Embodiment 2 of the present invention. In the figure, reference numeral 201 denotes a symbol / character insertion changeover switch, and when the user selects symbol / character input, a symbol / character input selection signal 1
02 is transmitted, and as a result, the symbol character insertion changeover switch 2
01 is connected to the contact A. When the user does not select the symbol character input and the symbol character input selection signal 102 is not transmitted, the symbol character input changeover switch 201 is connected to the contact B.

A symbol character inserting means 202 inserts a symbol character at the most suitable position between word strings using the symbol character grammar 101. The same functional blocks as those in the first embodiment are designated by the same reference numerals and the description thereof will be omitted.

Next, the operation will be described with reference to FIG. First, in step ST201, the word acoustic pattern generation means 1104 generates a word acoustic pattern using the recognition target word dictionary 1105 and the standard pattern 1106.

In step ST202, the voice feature amount extraction means 1101 converts the voice signal of the input voice 1002 into A /
The D-converted and A / D-converted signal is cut out in frames with a fixed time interval of about 5 ms to 20 ms, acoustic analysis is performed, and a voice feature vector O = [o (1), o (2) , ..., o
(T)] (T is the total number of frames).

Collating means 11 in step ST203
Reference numeral 02 denotes the word acoustic pattern and the voice feature amount extraction means 110 generated in step ST201 according to the normal grammar 101.
The speech feature amount O extracted in 1 is collated, and the speech recognition result not including the symbol character is output as the intermediate speech recognition result Wc by the mathematical expression 1.

In step ST204, it is determined whether or not the user selects the symbol / character input by the symbol / character input selection signal 102. When the user selects the symbol character input, the symbol character input selection signal 102 is transmitted. Therefore, in step ST204, the symbol character insertion changeover switch 201 determines that the symbol character input is selected and connects to the contact A. In this case, next, in step ST205, the symbol character inserting means 202 determines the position of inserting the symbol character between the words of the intermediate speech recognition result Wc according to the symbol character grammar 101, inserts the symbol character, and performs the voice recognition. The result 1004 is output.

The symbol character inserting process in the symbol character inserting means 202 will be described by using an example. It is assumed that the symbol character grammar 101 is 2-gram, and the speech recognition result Wc that does not include the symbol character output from the matching unit 1102 is “I am going to the station”. The symbol character inserting means 202,
The most appropriate symbol character is inserted between the words of the word string Wc using the symbol character grammar 101. Here, in order to determine whether to insert the symbol character between “Iwa” and “To the station”, the value of the symbol character that maximizes the value of Expression 9 is larger than a predetermined threshold value. Or not. For the symbol character “,”, when the value of the mathematical expression 9 is the maximum and the value is larger than the threshold value, the symbol character is inserted and “I am to the station”.

[0066]

[Formula 9]

Since the symbol character input selection signal 102 is not transmitted when the user does not select the symbol character input,
In step ST204, the symbol / character insertion changeover switch 201 is connected to the contact B. As a result, step ST
At 206, the intermediate speech recognition result Wc is set to the speech recognition result 10
Output as 04.

In step ST207, it is determined whether or not the utterance has ended. If utterance has not ended, the process returns to step ST202 and is repeated. on the other hand,
If the utterance has ended, the processing ends.

In the second embodiment, the voice feature quantity extracting means 1101, the collating means 1102, the word sound pattern generating means 1104, and the symbol character inserting means 202 may be configured by hardware, but these processes are performed. A voice recognition program may be created and the computer may execute the voice recognition program.

As described above, according to the voice recognition device of the second embodiment, when a symbol character is input,
Since the user can automatically input the symbol characters by selecting the symbol character input and uttering a word, there is an effect that a document can be efficiently created.

[0071]

[Third Embodiment] FIG. 7 is a block diagram showing the structure of a speech recognition apparatus according to a third embodiment of the present invention. In the figure, 301 is a symbol character insertion means at the pause position,
According to the symbol character grammar 101, the most appropriate symbol character is inserted at the position of the pause in the word string. Here, the pause position refers to a silent section that exists between the utterances of words for a certain period of time or longer. Note that the same functional blocks as those of the first and second embodiments are designated by the same reference numerals and the description thereof will be omitted.

Next, the processing method will be described with reference to FIG. First, in step ST301, the word acoustic pattern generation means 1104 generates a word acoustic pattern using the recognition target word dictionary 1105 and the standard pattern 1106.

The voice feature quantity extraction means 1101 executes step S.
At T302, the audio signal of the input audio 1002 is A /
The D-converted and A / D-converted signal is cut out in frames with a fixed time interval of about 5 ms to 20 ms, acoustic analysis is performed, and a voice feature vector O = [o (1), o (2) ,…, O
(T)] (T is the total number of frames).

Collating means 11 in step ST303
02 is step ST301 according to the normal grammar 101.
And the sound feature amount extraction means 1
By comparing with the voice feature amount O extracted in 101, the voice recognition result not including the symbol character is output as the intermediate voice recognition result Wpc by the mathematical expression 1, and the silent interval between words is detected to determine the pause position information. Will also be output.

In step ST304, it is determined whether or not the user has selected the symbol character input by the symbol character input selection signal 102. When the user selects the symbol character input, the symbol character input selection signal 102 is transmitted. Therefore, in step ST304, the symbol character insertion changeover switch 201 determines that the symbol character input is selected and connects to the contact A.

In this case, next, in step ST305, the symbol character inserting means 301 at the pause position determines the position of inserting the symbol character between the words of the intermediate speech recognition result Wpc according to the symbol character grammar 101, and at this position. A symbol character is inserted and a voice recognition result 1004 is output.

Next, the symbol character insertion means 3 at the pause position
The symbol character insertion processing at the pause position in 01 will be described using an example. Symbolic character grammar 101 is 2-g
Suppose that the intermediate speech recognition result Wpc is “I am going to <pause> station”. The symbol character insertion means 301 at the pause position inserts the most appropriate symbol character at the pause position of the word string Wpc according to the symbol character grammar 101. Here, in order to decide whether to insert the symbol character in the pose position that exists between "I am" and "To the station",
For the symbol character that maximizes the value of Expression 9, it is determined whether the value is larger than a predetermined threshold value. For the symbol character “,”, when the value of the mathematical expression 9 is the maximum and the value is larger than the threshold value, the symbol character is inserted and “I am to the station”.

Since the symbol character input selection signal 102 is not transmitted when the user does not select the symbol character input, the symbol character insertion changeover switch 201 is connected to the contact B in step ST304. As a result, in step ST306, the intermediate speech recognition result Wpc is output as the speech recognition result 1004.

In the present embodiment, the case where a pause position exists in the middle of utterance has been described, but it is also possible to insert a symbol character after the last word of utterance. In the example of “go”, in the symbol character that maximizes Expression 10, the symbol character is inserted when its maximum value is larger than a predetermined threshold value. In the case of the symbol character “.”, The numerical formula 10 becomes maximum, and when the maximum value is larger than the threshold value, “go.”.

[0080]

[Formula 10]

In step ST307, it is determined whether or not the utterance has ended. If utterance has not ended, the process returns to step ST302 and repeats. on the other hand,
If the utterance has ended, the processing ends.

In the present embodiment, the voice feature quantity extraction means 1101, the collation means 1102, the word acoustic pattern generation means 1104, and the symbol character insertion means 3 at the pause position.
01 may be configured by hardware, but a voice recognition program for performing these processes may be created and the computer may execute the voice recognition program.

As described above, according to the voice recognition device of the present embodiment, the user selects the symbol character input to input the symbol character and speaks the word, so that the symbol character is automatically placed at the pause position. Since it can be input manually, there is an effect that a document can be efficiently created.

[0084]

Fourth Embodiment FIG. 9 is a block diagram showing the structure of a voice recognition device according to a fourth embodiment of the present invention. In the figure, reference numeral 401 denotes a prosody information extraction means, which is means for extracting prosody information from the user's voice. Here, the prosody information is information on accent and intonation. The prosody information is described in Chapter 4 of Document 1.

Reference numeral 402 is a symbol character inserting means using prosody information. Note that the same functional blocks as those of the first to third embodiments are designated by the same reference numerals and the description thereof will be omitted.

Next, the operation will be described with reference to FIG. In step ST401, the word acoustic pattern generation means 1104 generates a word acoustic pattern using the recognition target word dictionary 1105 and the standard pattern 1106.

In step ST402, the voice feature amount extraction means 1101 converts the voice signal of the input voice 1002 into A /
The D-converted and A / D-converted signal is cut out in frames with a fixed time interval of about 5 ms to 20 ms, acoustic analysis is performed, and a voice feature vector O = [o (1), o (2) , ..., o
(T)] (T is the total number of frames).

In step ST403, the collating means 11
02 is step ST401 according to the normal grammar 101.
Word acoustic pattern and speech feature amount extraction means 11 generated in
Matching with the voice feature amount O extracted in 01
Outputs a speech recognition result that does not include a symbol character as an intermediate speech recognition result Wc.

In step ST404, it is determined by the symbol / character input selection signal 102 whether the user has selected the symbol / character input. When the user selects the symbol character input, the symbol character input selection signal 102 is transmitted. Therefore, in step ST404, the symbol character insertion changeover switch 201 determines that the symbol character input is selected and connects to the contact A.

In this case, in step ST405, the prosody information extracting means 401 extracts prosody information from the user's voice.

In step ST406, the symbol character inserting means 402 using the prosody information uses the symbol character grammar 101 and the prosody information output from the prosody information extracting means 401 in step 405, to generate a symbol between the words of the intermediate speech recognition result Wc. The position at which the character is inserted is determined, the symbol character is inserted at this position, and the voice recognition result 1004 is output.

Next, the symbol character insertion processing in the symbol character insertion means 402 using the prosody information will be described using an example. It is assumed that the symbol / character grammar 101 is 2-gram and the intermediate speech recognition result Wc is “I am going to the station”. The prosody information extraction means 401 extracts prosody information,
Here, the pitch frequency is extracted as the most dominant factor of the prosody. The pitch frequency is p. 71
As shown in (3), the pitch frequency is high at the beginning of the vocalization period, that is, in the exhalation stage, which is surrounded by a pause, and gradually decreases due to a decrease in hypoglottic pressure. FIG. 11 qualitatively shows the change situation of the pitch frequency, and shows an example in which "I am" is uttered at once and then "go to the station" is uttered. Generally, since "," is inserted in the breathing portion, the portion where a gap occurs in the pitch frequency, such as the point X in FIG. 11, is the target of symbol character insertion. Symbolic character insertion means 4 using prosody information
02 inserts the most suitable symbol character at the position of the pitch frequency gap in the word string Wc based on the symbol character grammar 101. In this example, in order to determine whether to insert a symbol character in the pitch frequency gap existing between “I am” and “to the station”, the value of the symbol character that maximizes Equation 9 is previously set. It is determined whether or not it is larger than a predetermined threshold. For the symbol character “,”, when the value of Expression 9 is the maximum and the value is larger than the threshold value,
After inserting the symbolic characters, it becomes "I'm going to the station".

If the user has not selected the symbol / character input, the symbol / character input selection signal 102 is not generated and transmitted to the symbol / character insertion changeover switch 201. Therefore, in step ST404, the symbol / character insertion changeover switch 201 is selected.
Is connected to contact B. As a result, step ST407
, The intermediate speech recognition result Wc is converted to the speech recognition result 1004
Output as.

In step ST408, it is determined whether utterance has ended. If the utterance has not ended, the process returns to step ST402 and is repeated. On the other hand, if the utterance has ended, the process ends.

In the present embodiment, the prosody information extraction means 401 is configured as a block separate from the voice feature quantity extraction means 1101, but the voice feature quantity extraction means 110 is used.
1 may execute the process of extracting prosody information.

Further, in this embodiment, the voice feature quantity extraction means 1101, the matching means 1102, the word acoustic pattern generation means 1104, the prosody information extraction means 401, and the symbol character insertion means 402 using prosody information are configured by hardware. Alternatively, a voice recognition program that performs these processes may be created and the computer may execute the voice recognition program.

As described above, according to the voice recognition device and the voice recognition method of the fourth embodiment, when the user inputs a symbol character, the user selects the symbol character input and utters the voice. Since the symbol characters can be automatically input using the prosody information, there is an effect that a document can be efficiently created.

[0098]

Fifth Embodiment FIG. 12 is a block diagram showing the structure of a voice dialogue apparatus according to a fifth embodiment of the present invention. Figure 1
In No. 2, 501 is a symbol character input selection means. Figure 1
The same functional blocks as those in FIG. 14 of the voice recognition device of the first embodiment and the conventional voice dialogue device shown in FIG.

Next, the operation will be described with reference to FIG. In step ST501, the user 1001 selects the symbol character input using the symbol character input selection means 501. Here, for example, a configuration is possible in which when the Shift key of the keyboard is pressed, selection of symbol character input is performed. When the user selects the symbol character input, the symbol character input selection signal 102 is transmitted. Voice recognition means 100
3 judges by the symbol character input selection signal 102 whether the user 1001 has selected the symbol character input.

If the insertion of symbol characters is selected in step ST501, the voice recognition means 1003 outputs the voice recognition result 1004 in step 503.
Here, the voice recognition unit 1003 is, for example, the voice recognition unit described in the first embodiment. In this case, the speech recognition unit 1003 acoustically analyzes the input speech, selects an acoustic pattern from the word acoustic pattern and the symbolic character acoustic pattern according to the grammatical character grammar, and compares this acoustic pattern with the extracted speech feature amount. Result of speech recognition 100
4 is output.

Although the voice recognition means 1003 has been described in the case of the first embodiment, it may be any one of the voice recognition means of the second to fourth embodiments.

When the symbol character insertion is not selected, the voice recognition means 1003 in step ST502.
Outputs a voice recognition result 1004. Here, the voice recognition unit 1003 is, for example, the voice recognition unit described in the first embodiment. In this case, the speech recognition unit 1003 acoustically analyzes the input speech, selects a word acoustic pattern according to the normal grammar 1103, matches the acoustic pattern with the speech feature amount, and the speech recognition result 100 that does not include a symbol character.
4 is output.

In step ST504, step S
The voice recognition result 1004 output by the voice recognition means 1003 in T502 or step ST503 is displayed on the display device.

In step ST505, it is determined whether the utterance has ended. If the utterance has not ended, the process returns to step ST501 and is repeated. on the other hand,
If the utterance has ended, the processing ends.

In the fifth embodiment, the symbol / character input selection means 501, the voice recognition means 1003, and the voice recognition result display means 1005 may be configured by hardware, but a program for performing these processing is created. The computer may execute the voice recognition program.

As described above, according to the voice recognition device and the voice recognition method of the fifth embodiment, the user can input the symbol characters by giving an instruction from the auxiliary input means such as the keyboard and uttering. Since the recognition result is displayed, there is an effect that a document can be efficiently created.

[0107]

Since the apparatus and program according to the present invention are configured to match the utterance portion corresponding to the symbol character included in the input voice with the reading of the symbol character by using the symbol character grammar, the symbol character is recognized by the voice recognition. This has the effect of enabling the input of documents including.

The apparatus and program according to the present invention are
The grammar selection switch is used to switch between the normal grammar and the symbolic character grammar, the word acoustic pattern and the symbolic character acoustic pattern are selected according to the grammar selected here, and the acoustic pattern and the acoustic feature amount extracted from the input speech are collated. Since the configuration is adopted, an appropriate grammar can be switched according to the document that the user is trying to input, and as a result, the literacy rate is improved.

The apparatus and program according to the present invention are
Regarding the intermediate speech recognition result obtained by matching the acoustic feature amount extracted from the input speech according to the normal grammar, the symbol character insertion position is determined according to the symbol character grammar,
Since the configuration is such that the symbol characters are inserted, the output result of the document input device using the speech recognition according to the conventional configuration that does not use the symbol character grammar is used as the intermediate speech recognition result of the present invention.
Further, there is an effect that a symbol character can be inserted into the intermediate speech recognition result.

The device and program according to the present invention are
The silent segment is extracted from the input voice as a pause position, and the symbol position is configured to insert the symbol character into the intermediate speech recognition result according to the symbol character grammar. Therefore, the user can recognize the voice without uttering the symbol character. This has an effect that symbol characters can be input in the document input by.

The apparatus and program according to the present invention are
Since the prosody information is extracted from the input speech and the symbol character is inserted into the intermediate speech recognition result according to the symbol character grammar for the gap position in the prosody information, the user does not need to utter the speech corresponding to the symbol character. This has an effect that symbol characters can be input in document input by recognition.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a first embodiment of the present invention.

FIG. 2 is a flowchart of processing according to the first embodiment of the present invention.

FIG. 3 is a diagram showing an example of a recognition target word dictionary according to the first embodiment of the present invention.

FIG. 4 is a diagram showing an example of a dictionary for notation and reading of symbol characters according to the first embodiment of the present invention.

FIG. 5 is a configuration diagram of a second embodiment of the present invention.

FIG. 6 is a flowchart of processing according to the second embodiment of the present invention.

FIG. 7 is a configuration diagram of a third embodiment of the present invention.

FIG. 8 is a flowchart of processing according to the third embodiment of the present invention.

FIG. 9 is a configuration diagram of a fourth embodiment of the present invention.

FIG. 10 is a flowchart of processing according to the fourth embodiment of the present invention.

FIG. 11 is a time-series graph showing pitch frequency variation in the fourth embodiment of the present invention.

FIG. 12 is a configuration diagram of a fifth embodiment of the present invention.

FIG. 13 is a flowchart of processing according to the fifth embodiment of the present invention.

FIG. 14 is a configuration diagram of a conventional technique.

FIG. 15 is a configuration diagram of a voice recognition device according to a conventional technique.

[Explanation of symbols]

101: Symbolic character grammar 102: Symbolic character input selection signal 103: Grammar changeover switch 104: Symbolic character acoustic pattern generation means 105: Symbolic character notation and reading dictionary 201: Symbolic character changeover switch 202: Symbolic character insertion means 301: Pause Symbol / character insertion means 401 to position: Prosody information extraction means 402: Symbol / character insertion means 501 using prosody information 501: Symbol / character input selection means 1001: User 1002: Input voice 1003: Speech recognition means 1004: Speech recognition result 1005: Speech recognition result display means 1101: Speech feature amount extraction means 1102: Matching means 1103: Normal grammar 1104: Word acoustic pattern 1105: Recognition target word dictionary 1106: Standard pattern table

Claims (14)

[Claims] 1. A voice feature amount extraction means for acoustically analyzing an input voice to extract a voice feature amount indicating a voice feature, and a standard pattern for obtaining an acoustic score in voice recognition stored in a standard pattern table. A word acoustic pattern generation means for generating a word acoustic pattern for this word from a word stored in the recognition target word dictionary, a symbol character stored in the dictionary for notation and reading of symbol characters, and a symbol for this symbol character from the above standard pattern. A symbol / character acoustic pattern generating means for generating a character / acoustic pattern, a symbol / character grammar storing means for storing a symbol / character grammar which is a connection rule between a word and a symbol / character, and the word / acoustic pattern and the symbol / character according to the symbol / character grammar. A sound pattern is selected from the sound patterns, the sound pattern is compared with the sound feature amount, and a word or a sound pattern that matches the sound pattern or A voice recognition device comprising collating means for outputting a symbol character as a voice recognition result of the input voice. 2. The speech recognition device comprises: a normal grammar storing means for storing a normal grammar which is a connection rule between words; and a grammar changeover switch for selecting one of the normal grammar and the symbol character grammar. When the grammar selector switch selects the normal grammar, the matching means selects a sound pattern from the word sound patterns in accordance with the normal grammar, matches the sound feature amount with the sound feature quantity, and matches the sound. The speech recognition apparatus according to claim 1, wherein the word for the pattern is output as a speech recognition result that does not include a symbol character for the input speech. 3. A voice feature amount extraction means for acoustically analyzing an input voice to extract a voice feature amount indicating a voice feature, a standard pattern for obtaining an acoustic score in voice recognition, and a recognition target word dictionary are stored. A word sound pattern generation means for generating a word sound pattern for this word from a word, a normal grammar storage means for storing a normal grammar which is a connection rule between words, and a sound selected from the word sound patterns according to this normal grammar. Collating means for collating the pattern and the voice feature amount, and outputting a word for the matched acoustic pattern as an intermediate speech recognition result for the input voice,
A symbol character grammar storing means for storing a symbol character grammar which is a connection rule between a word and a symbol character, and a symbol character is inserted into the intermediate speech recognition result according to the symbol character grammar to generate a document including the symbol character, and this document is generated. A voice recognition device comprising a symbol character insertion means for outputting. 4. The speech feature quantity extracting means acoustically analyzes the input speech including the utterances of the symbol characters to extract the speech feature quantity of the utterances of the symbol characters, and the collating means is a collation result about the utterances of the symbol characters. Output intermediate speech recognition result including parts,
4. The symbol character inserting means is configured to identify the collation result portion according to the symbol character grammar and replace the collation result portion with the symbol character to generate a document including the symbol character. The voice recognition device described. 5. The voice feature amount extracting means acoustically analyzes the input voice including a pause time at a constant interval to extract a voice feature amount including information about a pause position, and the symbol / character inserting means determines the pause position. 4. The speech recognition apparatus according to claim 3, wherein the speech recognition apparatus is configured to identify the character according to the symbol character grammar, and insert the symbol character at the pause position to generate a document including the symbol character. 6. The voice feature quantity extraction means extracts prosody information from the input voice, and the symbol character grammar storage means stores the relation between the prosody information and the symbol characters in the symbol character grammar. The inserting means is configured to insert a symbol character into the intermediate speech recognition result according to the prosody information extracted by the speech feature extracting means and the symbol character grammar to generate a document including the symbol character. The voice recognition device according to 3. 7. The speech recognition device comprises a normal grammar storage means for storing a normal grammar which is a connection rule between words, and a grammar changeover switch for selecting one of the normal grammar and the symbol character grammar. The symbol character inserting means is configured to output the intermediate speech recognition result as a document containing no symbol character when the grammar selector switch selects the normal grammar. The voice recognition device according to any one of claims 3 to 6. 8. A voice feature amount extraction procedure for acoustically analyzing an input voice to extract a voice feature amount indicating a voice feature, and a standard pattern for obtaining an acoustic score in voice recognition stored in a standard pattern table. The word sound pattern generation procedure that generates the word sound pattern of this word from the words stored in the recognition target word dictionary, and the sign character sound of this sign character from the sign character and the above standard pattern stored in the dictionary of notation and reading of sign characters A symbol / character acoustic pattern generation procedure for generating a pattern, a symbol / character grammar storage procedure for storing a symbol / character grammar that is a connection rule between words and symbol characters, and the above-mentioned word / acoustic pattern and the symbol / character acoustic pattern according to this symbol / character grammar. Select an acoustic pattern from the above, check this acoustic pattern against the above-mentioned voice feature amount, and enter the word or symbol character for the matching acoustic pattern above. A voice recognition program for causing a computer to execute a verification procedure for outputting a voice recognition result of a force voice. 9. The speech recognition program includes a normal grammar storing procedure for storing a normal grammar, which is a connection rule between words,
The normal grammar and the grammar switching procedure for selecting one of the symbolic grammars are further caused to be executed by the computer, and the collation procedure is performed when the grammar switching switch selects the normal grammar. According to the normal grammar, an acoustic pattern is selected from the word acoustic patterns, collated with the speech feature amount, and a word for the matching acoustic pattern is output as a speech recognition result that does not include the symbol characters of the input speech. 9. The voice recognition program according to claim 8. 10. A voice feature amount extraction procedure for acoustically analyzing an input voice to extract a voice feature amount indicating a voice feature, a standard pattern for obtaining an acoustic score in voice recognition, and a recognition target word dictionary are stored. A word acoustic pattern generation procedure for generating a word acoustic pattern of this word from a word, a normal grammar storing procedure for storing a normal grammar that is a connection rule between words, and an acoustic pattern selected from the word acoustic patterns according to this normal grammar. And the speech feature amount are collated, and a matching procedure for outputting a word for a matched acoustic pattern as an intermediate speech recognition result for the input speech, and a symbol character grammar that is a connection rule between the word and the symbol character are stored. According to the symbol character grammar storage procedure and the symbol character grammar, the symbol character is inserted into the intermediate speech recognition result to generate a document including the symbol character, and this document is generated. A speech recognition program for causing a computer to execute a symbol character insertion procedure for outputting. 11. The speech feature quantity extraction procedure acoustically analyzes an input speech including a utterance of a symbol character to extract a speech feature quantity about a utterance of a symbol character, and the above matching procedure includes a collation about an utterance of a symbol character. The intermediate speech recognition result including the result portion is output, and the symbol character insertion procedure identifies the matching result portion according to the symbol character grammar and replaces the matching result portion with the symbol character to generate a document including the symbol character. The voice recognition program according to claim 10, wherein the voice recognition program is configured. 12. The voice feature amount extraction procedure acoustically analyzes an input voice including a pause time at a constant interval to extract a voice feature amount including information about a pause position, and the symbol character insertion procedure includes the pause position. 11. The speech recognition program according to claim 10, wherein the program is identified according to the symbol character grammar, and the symbol character is inserted at the pause position to generate a document including the symbol character. 13. The voice feature quantity extraction procedure extracts prosody information from an input voice, and the symbol character grammar storage procedure comprises:
The relation between the prosodic information and the symbolic character is stored in the symbolic character grammar, and the symbolic character inserting step inserts the symbolic character into the intermediate speech recognition result according to the prosodic information extracted by the speech feature extracting means and the symbolic character grammar. 11. The speech recognition program according to claim 10, wherein the speech recognition program is configured to generate a document including a symbol document. 14. The speech recognition program comprises a normal grammar storing procedure for storing a normal grammar which is a connection rule between words, and a grammar switching procedure for selecting one of the normal grammar and the symbol character grammar. And the symbol character insertion procedure is configured to output the intermediate speech recognition result as a document not containing symbol characters when the grammar switching procedure selects the normal grammar. 14. The voice recognition program according to claim 10, characterized in that.