JPH05119793A - Method and device for speech recognition - Google Patents

Abstract

PURPOSE:To provide a speech word processor which creates a document with a free vocabulary by speech input while attaining a high-precision recognition rate by limiting an object of recognition. CONSTITUTION:In a standard pattern storage part 6, the recognition vocabulary is classified into plural partial gatherings and registered for respective speech standard patterns while partial gathering flags are added together with label names. A speaker's speech is inputted from a speech input part 1 and a speech analysis part 2 extracts the feature pattern of the speech. The speaker, on the other hand, inputs a code indicating the partial gathering of the voiced speech from a key input part 5. A standard pattern selection part 7 selects the standard pattern group of the partial gathering corresponding to the input code and its label name in the standard pattern storage part 6. A pattern matching part 8 selects the feature pattern having the highest similarity in the selected standard pattern group and outputs its label name through a recognition output part 9.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition method and apparatus suitable for a voice word processor for inputting voice and outputting characters.

[0002]

2. Description of the Related Art Various studies have been conventionally made on a technique for creating a sentence using voice recognition, that is, a so-called voice word processor. The reason is that it is easy to compose sentences according to the words spoken by humans, and the input speed of voice is faster than that of character input. In a voice word processor, for example, it is an important technique to understand the meaning of a voice uttered in a phrase unit and write it down into a sentence containing kanji and kana. In order to realize this technology, it is important to establish a large vocabulary speech recognition technology and a recognition technology capable of coping with inflectional parts such as verbs and adjectives that are difficult to recognize.

In the prior art, in order to realize large vocabulary recognition, a recognition method for significantly increasing the memory capacity of words has been adopted. However, the memory capacity is limited on the hardware, the processing time for recognizing a large vocabulary increases in proportion to the number of vocabularies, and the recognition rate decreases as the number of similar words increases. It is considered that this method is not effective as a method for recognizing a large vocabulary word with high accuracy. In addition, there is a method of recognizing vocabulary at the syllable level and converting a syllable string into a vocabulary, but since it is necessary to determine the conversion rule and the word recognition accuracy depends on the syllable recognition accuracy,
Even now, research is being conducted to obtain highly accurate results.

In addition, the technology of (1) is based on a knowledge processing, a method of estimating the ending change from the context of spoken words or between phrases and improving recognition accuracy, and based on a large amount of sentence database. A method of estimating the succeeding word ending from the stem portion of the word and improving the recognition accuracy is adopted. However, since the knowledge processing and the method using the database require complicated processing on the hardware, there is a problem that the device scale becomes large, and it is not suitable for application to a simple system.

[0005]

The speech recognition method for improving the uncertain portion of the recognition result by applying the conventional knowledge processing or large-scale database, etc. requires a large processing scale.
It is not suitable for building a simple voice recognition device. A solution to this is expected to be achieved by assisting the input of a portion where the recognition result is unclear in voice recognition with another reliable input means such as a keyboard. In the method, in the large vocabulary recognition, the recognition target word is divided into subsets by specifying the key input and only the specific vocabulary is recognized as the recognition target, and a plurality of recognition result candidates are listed, and the result is selected by the key input. A method and a plurality of label names corresponding to inflection changes are attached to the recognized voice, and the recognized voice is
A combination of a method of selecting a label name by key input, etc. can be considered. These methods can be achieved by limiting the number of keys to be input without significantly reducing the convenience or speed of voice input.

It is an object of the present invention to combine a supplementary input method such as a human key input with voice recognition to achieve a highly accurate recognition rate without significantly reducing the convenience or speed of voice input. It is possible to construct a large vocabulary speech recognition device, and to provide a speech word processor that creates a sentence of free vocabulary for speech input.

[0007]

In order to achieve the above object, the speech recognition method of the present invention classifies a recognition vocabulary into a plurality of subsets, and for a standard pattern of speech, a subset together with its label name. Is registered in the standard pattern storage unit, a subset is designated by human code input, the standard pattern and label name of the relevant subset are selected from the standard pattern storage unit, and the selected The feature is that the similarity between the standard pattern of the subset and the unknown voice pattern is calculated, and the label name of the standard pattern with high similarity is output.

Further, the speech recognition method of the present invention outputs label names for a plurality of standard patterns determined to have a high similarity, such as a standard pattern having the highest similarity and a standard pattern having the second highest similarity. Then, one of the label names is finally selected by a human code input from among them.

Further, in the voice recognition method of the present invention, a plurality of label names are given to one standard pattern and registered in the standard pattern storage unit, and the standard pattern judged to have a high degree of similarity is It is characterized by outputting a plurality of label names and finally selecting one label name by a human code input.

Further, the speech recognition apparatus of the present invention which implements these methods classifies the recognition vocabulary into a plurality of subsets, and adds a flag indicating the subset together with its label name to the standard speech pattern. The standard pattern storage unit that has been stored in memory, the voice input unit that inputs the voice signal, the signal analysis unit that detects the voice section from the voice signal and extracts the feature pattern of the voice by signal analysis, and outputs the code by human key operation. A key input section, a standard pattern selection section for selecting and capturing a standard pattern and label name of the corresponding subset from the standard pattern storage section by a subset specifying code input through the key input section; and the signal analysis section. Pattern matching section for calculating the similarity between the characteristic pattern output from the standard pattern and the standard pattern selected by the standard pattern selection section , The calculation result of the pattern matching unit, characterized in that it is composed of a label name output section that outputs the label of the standard pattern similarity between the feature pattern is determined to be high.

[0011]

In the present invention, when classifying the recognized vocabulary into subsets, the recognition vocabulary is divided into a field-based set, a part-of-speech-based set, and the like, and the vocabulary having similar utterances in one subset is reduced. By doing, even if the population is a large vocabulary, it is possible to perform recognition with less deterioration in recognition accuracy. Furthermore, by giving a plurality of label names corresponding to word ending changes to one voice, it is possible to freely select a word ending change that was difficult to recognize in the conventional technique with the aid of human key input.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a block diagram of a voice recognition apparatus which is an embodiment of the present invention. The voice input unit 1 is a speaker 11
The voice signal is input by a microphone, a transmitter, or the like. The voice analysis unit 2 detects a voice signal input from the voice input unit 1 and analyzes the detected signal period. As a method of section detection,
For example, there is a method using short-time power. Moreover, as an analysis method, for example, there is an LPC cepstrum. The label input unit 3 inputs a label name for the analyzed voice section, and its means is, for example, a keyboard of a computer. Here, it is assumed that the number of label names for one voice is good, for example, from a minimum of 1 to a maximum of 9. It is also possible to give a name to the label name that is mixed with kanji and kana. The label output unit 4 displays a label name as a recognition result, and also displays the timing of voice input to the speaker 11, which means is, for example, a display device of a computer or a character. A voice synthesizer that synthesizes voice from data and outputs it. The key input unit 5 is used in combination with the standard pattern selection unit 7 and the recognition output unit 9 to select a recognition vocabulary and a recognition result, and also, to puncture a sentence when proofreading a sentence only by key input, a reading point,
It enables line breaks, deletion of words, etc., and its means is, for example, a PB tone recognition output device in a touch-tone phone. The key input section 5 and the label input section 3 may be the same.

In the registration mode, the standard pattern storage unit 6 stores
The voice pattern analyzed by the voice analysis unit 2 (this is called a standard pattern) is given the label name input by the label input unit 3 and held. The standard pattern selection unit 7 extracts only a subset corresponding to the input code from the key input unit 5 from all the standard patterns stored in the standard pattern storage unit 6 and uses it for calculation in the pattern matching unit 8. To set. In the recognition mode, the pattern matching unit 8 includes the unknown voice feature pattern of the label name currently analyzed by the voice analysis unit 2 (this is referred to as an unknown pattern with respect to the standard pattern) and the plurality of patterns selected by the standard pattern selection unit 7. The similarity calculation is performed with the standard pattern of.
D, which is known as a pattern matching method for voice recognition
There is P matching. The recognition output unit 9 outputs one or more label names attached to one or more standard patterns determined to have the highest similarity with the unknown pattern as a result of the similarity calculation in the pattern matching unit 8. It is output. The recognition output unit 9 can also select and output only a specific recognition result or label name according to the input code from the key input unit 5. The sentence output unit 10 outputs a sentence including a word string determined by the recognition output unit 9.

FIG. 2 is a detailed explanatory view of the standard pattern storage unit 6. The standard pattern storage unit 6 stores 1 for each standard pattern.
In addition to the standard pattern dictionary that holds individual or multiple label names, it also has a large vocabulary dictionary that holds a subset name and a code uniquely assigned to it for classifying recognized vocabularies into subsets. There is.

FIG. 2A is a specific example of a large vocabulary dictionary, which is an example in which the recognized vocabulary is classified into sets by part of speech. here,
The particle group, the noun group, the verb group, ... Are subsets, each of which is given a unique code and registered in the standard pattern storage unit 6 as a large vocabulary dictionary. As shown in FIG. 2A, here, 01 # is a particle group, 02 # is a noun group, 03
# Indicates a group of verbs.

FIG. 2 (b) shows one record in the standard pattern dictionary. A minimum of one label name is required for the standard pattern, but a plurality of label names (for example, nine label names at the maximum) can be attached to the standard pattern. In FIG. 2B, the flag indicates to which subset the voice pattern of the record is classified. Here, the subset name in the vocabulary dictionary shown in FIG. 2A is entered. The voice pattern is a standard (voice) pattern used for recognition, and the result analyzed by the voice analysis unit 2 is expressed in bits. Label 1 contains the label name that is first output as a recognition result. Labels 2 to N contain the label names of the next candidates that are output one after another when the next candidate is designated by the key input unit 5. The end label represents the end of the record for one voice pattern.

2 (c) and 2 (d) are concrete examples of the standard pattern dictionary. FIG. 2 (c) shows a case of a noun group and (d) shows a case of a verb group. It should be noted that "*****" means a voice pattern (standard pattern) expressed in bits.

The operation of FIG. 1 will be described below. The operation is divided into a registration mode for registering a standard pattern used for recognition and a recognition mode for recognizing an unknown pattern. Each will be described below. In the case of the registration mode In the registration mode, the voice uttered by the speaker 11 is registered in the standard pattern storage unit 6. First, speaker 11
When the user designates a voice registration operation by operating the key input unit 5, the voice recognition device outputs a message "Please input voice" from the label output unit 4 on the screen or in the synthesized voice, and the voice of the speaker 11 is output. Prompt for input. Here, when the speaker 11 utters a word voice such as “I”, “Tokyo”, “go”, “you”, “wait”, “thank you”, the voice is captured from the voice input unit 1 and the voice is input. The analysis unit 2 detects a section and analyzes the signal in word units, and sends the signal to the standard pattern storage unit 6. When the speaker 11 inputs the label name of the voice from the label input unit 3, the standard pattern storage unit 6 registers the voice feature pattern as a labeled standard pattern in the standard pattern dictionary. The label name is, for example, the name of the uttered voice such as "go" or "wait" as it is.
As shown in, for example, "Go", "Don't go", "Go", "Go", and "Wait" are "Wait" , "I can't wait", "I can't wait", "Please wait"
With multiple labels such as. In addition, a sentence such as "Thank you every time" may be registered for the voice "Thank you." Similarly,
Register particles such as "ha", "de", and "he". At the same time, the standard pattern storage unit 6 gives a subset name to these stored patterns by input from the label input unit 3. For example, FIG. 2 (c) and (d)
As shown in, register the "noun group" as a subset in patterns such as "Tokyo", "I", "you",
Register "verbs" in patterns such as "go" and "wait". Similarly, "particle groups" are registered for patterns such as "ha", "de", and "he".

Further, it is determined which numerical value (code) from the key input unit 5 the subset name in the standard pattern dictionary is to be designated and is registered in the large vocabulary dictionary. For example, as shown in FIG. 2A, “particle group” = 01 #,
"Noun group" = 02 # and "verb group" = 03 #. These registrations are performed by the label input unit 3 and the key input unit 5.

In this way, the standard pattern dictionary and the large vocabulary dictionary are registered in the standard pattern storage unit 6. Also,
PB tone input from the key input unit 5
For those that directly control a sentence without using voice recognition, such as reading point, line feed, and cancellation, for example, "punctuation mark" = 91 #, "reading mark" = 92 #, "line feed" = 99 #, "cancel" = 90
#, Etc. In the case of recognition mode In recognition mode, based on the result of recognition of the input unknown pattern and a subset of standard patterns,
Create sentences. For example, "I'm going to Tokyo."
If you want to create a sentence by voice input, do as follows. Here, recognition of the part of the word "I" will be shown. A conceptual diagram of this voice recognition is shown in FIG.

First, when the speaker 11 specifies a voice recognition operation by operating the key input unit 5, the voice recognition device
A message such as "Please input PB tone" is output from the label output unit 4 on the screen or in the synthesized voice to prompt the speaker 11 to input PB. The speaker 11 inputs 02 # in the key input unit 5 and designates a "noun group" as a subset. In the standard pattern selection unit 7, when 02 # is input from the key input unit 5, 02 # = noun group is first obtained from the large vocabulary dictionary of the standard pattern storage unit 6, and then this "noun group" is keyed. Search the standard pattern dictionary for "noun group"
Take in a subset of the standard pattern with the flag as a pair with its label name. Here, again the label output unit 4
A message such as "Please input voice" is displayed to prompt the speaker 11 to input voice. When the speaker 11 inputs a voice such as "I", the voice is analyzed by the voice analysis unit 2 and the result is sent to the pattern matching unit 8 as an unknown pattern. Pattern matching unit 8
Then, similar calculation is performed for a plurality of standard patterns of the subset of the "noun group" selected by the standard pattern selection unit 7 and unknown patterns, and the result is sent to the recognition output unit 9. The recognition output unit 9 rearranges the recognition results in descending order of similarity, and outputs the results to the label output unit 4 on a screen or by voice.

At this time, the valid result is 1 as the recognition result.
When there are only words or when only one word result is requested, the result of the recognition output unit 9 is fixed as it is, but when the result is fixed from a plurality of candidates, first,
The first rank candidate is output on the screen or voice of the label output unit 4, and the key input unit 5 is set in the input waiting state. When the speaker inputs "#" (decision instruction) from the key input unit 5, the result is confirmed, and when "*" (indication of next candidate) is input, the next candidate is output and the input of the key input unit 5 is awaited. When the state is entered and "0" (voice input instruction) is input, the voice input is resumed. The correspondence between these key inputs and instructions is only an example, and the combination is arbitrary. In this way, each word is input by voice, the candidates are confirmed by the key input of the key input unit 5, and the sentence is generated.

Here, for words such as "verbs" in which a word having some inflections is registered for one voice, the stem is first determined, and the next inflection is determined. A conceptual diagram of this voice recognition is shown in FIG.

For example, when inputting the voice "go" to generate the word "go", first enter 03 # by key input to select "verb group", and then "go" Is input by voice, and the candidate for the stem of the recognition result is “Go”.
From "Wait", etc., select "Go" by pressing the "#" key. Then, select "Go" from the multiple label names given to the voice "Go", such as "Go", "Do not go", and "Go". This is also performed by a key input process as described above. In other words, if you enter "#" in the input waiting state of the key input unit 5, the result is confirmed, if you enter "*", the next candidate is displayed and the key input unit 5 is in the input waiting state, and if you input "0". Try to return to voice input again. In this way, even if there are a plurality of recognition candidates and each candidate has a plurality of label names, the recognition result can be confirmed by the input from the key input unit 5.

The word string generated by voice input is finalized as a sentence by supplementing "punctuation marks", "reading marks" and the like with key input, and finally a key corresponding to "line feed" is input from the key input unit 5. As a result, the text output unit 10 outputs the text as a finalized text.

Although one embodiment of the present invention has been described above, the configuration of FIG. 1 can also be realized by combining a terminal device and a center device. In this case, the terminal side has only a voice input unit 1 using a handset of a telephone, a key input unit 5 using a PB tone output device of a telephone, and a label output unit 4 such as a display device. The other label input unit 3, standard pattern storage unit 6, storage pattern selection unit 7, pattern matching unit 8, recognition output unit 9, etc. are all placed on the center side, and the terminals and the center are connected by a communication line.

[0028]

As is apparent from the above description, the following effects can be achieved by the voice recognition method and apparatus of the present invention. (1) High-accuracy large vocabulary speech recognition can be achieved by limiting recognition to a subset. (2) First by selecting a result from multiple recognition candidates
Even if the rank recognition result is incorrect, it can be corrected if it is listed as a recognition candidate. (3) By giving multiple label names to one voice and selecting to output a specific label name for the recognition result, it is possible to create a free sentence even for voice input of a small vocabulary. it can.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a voice recognition device according to the present invention.

FIG. 2 is a detailed explanatory diagram of a standard pattern storage unit in FIG.

FIG. 3 is an operation conceptual diagram when the word “I” is recognized in FIG. 1.

FIG. 4 is an operation conceptual diagram when recognizing the word “go” in FIG. 1.

[Explanation of symbols]

1 voice input unit 2 voice analysis unit 3 label input unit 4 label output unit 5 key input unit 6 standard pattern storage unit 7 standard pattern selection unit 8 pattern matching unit 9 recognition output unit 10 sentence output unit 11 speaker

Claims (4)

[Claims] 1. A recognition vocabulary is classified into a plurality of subsets, a standard pattern storage unit is added to a standard pattern with a flag indicating the subset together with its label name, and is registered in a standard pattern storage unit. By selecting the standard pattern and label name of the corresponding subset from the standard pattern storage unit, calculating the similarity between the standard pattern of the selected subset and the unknown voice pattern, and a standard with a high degree of similarity. A voice recognition method characterized by outputting a label name of a pattern. 2. The voice recognition method according to claim 1, wherein
A speech recognition method characterized by outputting the label names of a plurality of standard patterns judged to have a high degree of similarity and finally selecting one label name by a human code input. 3. The voice recognition method according to claim 1, wherein
A plurality of label names are given to one standard pattern and registered in the standard pattern storage unit. For standard patterns determined to have a high degree of similarity, the plurality of label names are output, and the A voice recognition method characterized in that one label name is finally selected by code input. 4. A recognized standard vocabulary is classified into a plurality of subsets, and a standard pattern storage unit that stores a standard pattern of voice with a label name and a flag indicating the subset is input, and a voice signal is input. A voice input unit, a signal analysis unit for detecting a voice section from a voice signal and extracting a voice characteristic pattern by signal analysis, a key input unit for outputting a code by a human key operation, and an input through the key input unit. A standard pattern selecting unit that selects and loads a standard pattern and label name of the corresponding subset from the standard pattern storage unit by a subset specifying code, and a characteristic pattern output from the signal analyzing unit and the standard pattern selecting unit The pattern matching unit that calculates the degree of similarity with the standard pattern that has been generated, and the calculation result of the pattern matching unit, A voice recognition device comprising a label name output unit that outputs a label name of a standard pattern that is determined to have a high degree of similarity to the signature pattern.