JP2833995B2 - Large vocabulary speech recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a large vocabulary speech recognition apparatus.

[0002]

2. Description of the Related Art Conventionally, this kind of large vocabulary speech recognition device has
This is a large vocabulary speech recognition device in the sense that the number of words to be recognized at the time of one utterance is large.

A conventional large vocabulary speech recognition apparatus of this kind will be described with reference to the drawings.

FIG. 7 is a block diagram showing an example of a conventional speech recognition apparatus.

In FIG. 7, reference numeral 1 denotes a word group for describing and storing a word group to be recognized, and is referred to as a word set herein. Before starting the voice recognition operation, the operator converts the plurality of word sets 1 into a voice feature sequence in a dictionary generation unit 3a in advance to generate a plurality of dictionary files 4a. When a dictionary loading instruction is given to the dictionary loading unit 6a, the dictionary loading unit 6a expands the contents of the plurality of dictionary files 4a in the dictionary memory 10. An operation instruction for starting a voice recognition operation is input to an operation instruction input unit 8.
The control unit 9a transmits an operation instruction to each unit of the apparatus, the voice input from the microphone 11 is converted into a voice feature by the voice feature extraction unit 12, stored in the voice feature memory 13, and the comparison / matching unit 14 This is compared with the contents of the dictionary memory 10 to select the closest word, and this is output to the recognition result output unit 15.

Here, in the application for switching the recognition target word for each utterance, according to the conventional large vocabulary speech recognition apparatus, a plurality of word sets 1 to be recognized at each time are prepared. A plurality of dictionary files 4a in advance
And a dictionary load instruction is given so that another dictionary is developed in the dictionary memory 10 each time one utterance ends. In addition, when a plurality of word sets 1 are to be recognized at the same time, a word set in which the plurality of word sets 1 are combined is prepared, and is treated as a different one from the individual word sets. Further, in order to reduce the number of times of loading the dictionary, a word set is created by combining a plurality of word sets 1 within a range allowed by the capacity of the dictionary memory 10, converted into one dictionary, and loaded into the dictionary memory 10, Comparison and collation unit 1
No. 4 had a function of limiting the range on the dictionary memory 10 to be compared.

[0007]

In this conventional large vocabulary speech recognition apparatus, since information on the dictionary configuration as to which dictionary file contains which recognition target word is not stored as an apparatus, the recognition target word is not stored. The instruction for expanding the dictionary file including the character string in the dictionary memory is left to the operator, and as a result, there are the following problems.

First, when there is a plurality of word sets, there is no means for checking even if a dictionary corresponding to a word set different from the recognition target is erroneously loaded at the time of recognition.

Second, if the operator does not memorize which dictionary file is currently loaded in the dictionary memory, a useless dictionary load instruction is given, resulting in a poor response of the apparatus.

Third, when a plurality of recognition target word groups are loaded into the dictionary memory at a time in order to reduce the number of dictionary loads for each utterance as much as possible and to increase the response of the apparatus, the dictionary file must be loaded for each combination. The number increases and file management becomes more complicated.

[0011]

A large vocabulary speech recognition apparatus according to the present invention comprises: storage means for storing a plurality of words to be subjected to speech recognition in a plurality of genres defined in advance; Each voice feature is extracted and classified into a plurality according to the input of predefined grouping information for grouping the plurality of words, and is stored in a plurality of files as a dictionary of voice information including the voice features. Dictionary file generating means for generating and storing dictionary configuration information indicating a file configuration according to the plurality of classifications, and when grouping information related to input voice is input, the dictionary input information is input with reference to the dictionary configuration information. After checking whether the voice information of the words of the group related to the grouping information is stored in the dictionary memory, the voice information is selected from the plurality of files and the dictionary is selected. A word dictionary generating means to be stored in the memory; a voice selecting means for extracting a voice feature from the input voice and comparing and collating with the voice information extracted from the dictionary memory to select a word having the most similar voice feature; It has.

The large vocabulary speech recognition apparatus of the present invention stores a plurality of words to be subjected to speech recognition in a plurality of genres defined in advance, and extracts a speech feature of each of the plurality of words. And categorizing the plurality of words into a plurality according to an input of predefined grouping information for grouping the plurality of words, accumulating the plurality of words in a plurality of files as a dictionary of audio information including the audio features, and storing the files according to the plurality of classifications. Dictionary file generating means for generating and storing dictionary configuration information indicating a configuration; and when grouping information related to input speech is input, the dictionary file generation unit refers to the dictionary configuration information and relates to the input grouping information. After checking whether the voice information of the words of the group is stored in the dictionary memory, the voice information is selected from the plurality of files and stored in the dictionary memory. Word dictionary generating means, and voice selecting means for extracting a voice feature from the input voice and comparing and collating with the voice information extracted from the dictionary memory to select a word having the most similar voice feature, The storage unit has a plurality of word sets for storing a plurality of words to be subjected to speech recognition for each of a plurality of predefined genres, and the dictionary file generation unit includes the plurality of word sets related to the input speech. Grouping information input unit for inputting grouped information, and a dictionary generating unit for selecting a word from the plurality of word sets according to the input of the grouping information, extracting voice characteristics of the word, and storing the extracted voice features in a plurality of dictionary files And the plurality of dictionary files storing the voice characteristics of the plurality of words, and the grouping information input to the grouping information input unit. And a dictionary configuration information storage unit that stores the dictionary configuration information generated,
The word dictionary generation unit checks whether or not the speech characteristics of the words selected based on the dictionary configuration information in advance and classified by the grouping information are not stored in the dictionary memory before the input of the input speech. A dictionary loading unit that selects from a plurality of dictionary files and stores the dictionary in the dictionary memory; and a dictionary memory that stores voice characteristics of the plurality of words selected from the dictionary file, wherein the voice selecting unit includes a microphone. A voice feature extraction unit for extracting a voice feature of the input voice input from the voice memory, a voice feature memory for storing the extracted voice feature, and a voice feature stored in the voice feature memory and selected from the dictionary memory. A comparison / matching unit for comparing / collating with a voice feature to recognize a word corresponding to the input speech and outputting a result thereof; In the case of generating the plurality of dictionary files for the input voice, the plurality of word sets to be simultaneously subjected to voice recognition are always written into one dictionary file as a first condition, and the same word set is used as a second condition. In some cases, writing to the plurality of dictionary files, storing a word set so that each of the plurality of dictionary files does not exceed the capacity of the dictionary memory as a third condition, and performing the dictionary based on the grouping information The dictionary file is configured to minimize the number of files.

[0013]

Next, the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In FIG. 1, in this embodiment, a plurality of word sets 1 for storing a plurality of words to be subjected to speech recognition for each of a plurality of predefined genres, and a predefined group related to input speech A grouping information input unit 2 for inputting grouping information, and a dictionary generating unit 3 for selecting words from a plurality of word sets 1 according to the input of the grouping information, extracting voice characteristics of the words, and storing the extracted voice features in a plurality of dictionary files 4. A plurality of dictionary files 4 for storing voice characteristics of a plurality of words; a dictionary configuration information memory 5 for storing dictionary configuration information generated based on the grouping information input to the grouping information input unit 2; Before the input of the input voice, it is checked whether or not the voice characteristics of the words previously selected based on the dictionary configuration information and classified by the grouping information are stored in the dictionary memory 10. Tsu dictionary loading section 6 to be stored in the dictionary memory 10 elected from multiple dictionary file 4 after click
A dictionary memory 10 for storing voice features of a plurality of words selected from the dictionary file 4, a voice feature extraction unit 12 for extracting voice features of an input voice input from a microphone (microphone) 11, and A voice feature memory 13 for storing voice features, and a voice feature stored in the voice feature memory 13 and a voice feature selected from the dictionary memory 10 are compared and collated to recognize a word corresponding to the input voice, and the result is output as a recognition result. A comparison / matching unit 14 that outputs to the unit 15;
The control unit 9 includes a dictionary loading unit 6, a voice feature extraction unit 12, and a control unit 9 that controls the comparison and collation unit 14 according to an operation instruction input from the operation instruction input unit 8.

In the case of generating a plurality of dictionary files for an input voice, the dictionary generating unit 3 of the present embodiment always converts a plurality of word sets 1 to be simultaneously subjected to voice recognition into one dictionary file 4 as a first condition. Writing, as the second condition, the same word set 1 may be written to a plurality of dictionary files 4 as the case may be. As the third condition, each of the plurality of dictionary files 4 may be written so as not to exceed the capacity of the dictionary memory 10. Is stored, and the dictionary file 4 is configured based on the grouping information so that the number of dictionary files 4 is minimized.

FIG. 2 is a flowchart showing the operation of the dictionary loading unit in FIG. 1, FIG. 3 is a diagram showing an example of a word set for each genre in the present embodiment, and FIG. FIG. 5 is a diagram showing an example, and FIG. 5 shows five configuration examples of a dictionary file in the present embodiment, wherein (a), (b),
(C), (d) and (e) are diagrams showing respective configurations,
FIG. 6 is a diagram illustrating an example of the dictionary configuration information according to the present embodiment.

Next, the operation of this embodiment will be described with reference to FIGS. 1, 2, 3, 4, 5 and 6.

The word set 1 is a minimum unit of a word group to be recognized. Specifically, it refers to a group of place name words, a group of product name words, and the like as shown in FIG. The grouping information input from the grouping information input unit 2 is information indicating a group to be simultaneously recognized among the plurality of word sets 1.

The dictionary generation unit 3 converts each word included in each word set 1 into a speech feature sequence and converts the words into a plurality of dictionary files 4.
At this time, the dictionary file 4 is configured with the following logic. First, the first logic is that the result of converting each of the word sets belonging to the same group into the voice feature sequence based on the input grouping information must be written in one dictionary file 4. The second logic is that the word set 1 belonging to another group may be written in one dictionary file 4. The third logic is that the capacity of each of the plurality of dictionary files 4 must be smaller than the capacity of the dictionary memory 10. The fourth logic is a logic for configuring the dictionary file 4 so as to minimize the number of dictionary files 4 among the above three logics. Further, the dictionary generation unit 3 generates dictionary configuration information indicating which group of the word set 1 has been written in which dictionary file 4 and stores the dictionary configuration information in the dictionary configuration information memory 5.

The voice recognition operation is performed by first receiving an operation instruction from the operation instruction input unit 8 and transmitting the operation instruction to each unit by the control unit 9. The operation instruction includes information indicating which group of the word set 1 is to be recognized. The control unit 9 first informs the dictionary loading unit 6 of this group. The dictionary loading unit 6 refers to the dictionary configuration information in the dictionary configuration information memory 5 and knows a dictionary file name to be loaded (S1). Next, the dictionary loading unit 6 checks the loaded dictionary file name stored in the dictionary memory 10 to check whether the dictionary file has already been loaded (S2). If not already loaded, the contents of the dictionary file 4 are loaded into the dictionary memory 10 (S3).

Next, the control unit 9 activates the voice feature extraction unit 12. The voice feature extraction unit 12 extracts voice features from voice data input from the microphone 11 and stores the voice features in the voice feature memory 13. Further, the control unit 9 activates the comparison and collation unit 14. The comparison / matching unit 14 compares and matches the content of the speech feature memory 13 with the content of the dictionary memory 10 and selects a word having a feature most similar to the input speech, and outputs the word to the recognition result output unit 15 as a recognition result. Output.

Next, a specific example will be described.

For example, assume that there are six word sets as shown in FIG. Also, consider the information shown in FIG. 4 as the grouping information. Here, since the grouping information designates a set of word sets 1 to be recognized at the same time, in FIG. 4, the group A includes “article name” such as “television” and “radio” and “B101” such as “B101”. This means that any of the "control words" such as "article code" and "yes" and "redo" are to be recognized at the same time. Similarly, group B means that “person in charge” and “control word” are to be recognized simultaneously, and group C means that “place name”, “business office name” and “control word” are to be recognized simultaneously.

The dictionary generation unit 3 receives the word set 1 and the grouping information and generates a dictionary file 4 so as to satisfy the above four logics. Considering the first logic and the second logic, (a) to (e) shown in FIG.
The following five dictionary configuration plans can be considered. That is, the word set 1 belonging to the same group is contained in one dictionary file 4, and the word set 1 is also allowed to be contained in a plurality of dictionary files 4. The numbers in FIG. 5 indicate the memory capacity when each word set is converted into a speech feature sequence (units are omitted). Here, for example, if the capacity of the dictionary memory 10 is 1000, the third logic is applied, and in FIG. 5C and FIG. It is excluded because it is not appropriate. Further, when the fourth logic is applied, (a) is excluded because the number of dictionary files is large. Eventually, one of the configurations (b) and (d) will be selected. Here, the description is continued assuming that (b) is selected. The dictionary generation unit 3 generates dictionary configuration information indicating the configuration of FIG. 5B and stores the dictionary configuration information in the dictionary configuration information memory 5. This is, for example, an information table as shown in FIG.

After the dictionary file 4 and the dictionary configuration information have been created in advance, the recognition operation is started.

First, when a word belonging to group A shown in FIG. 6 is to be recognized, an operation instruction "recognize in group A" is given from the operation instruction input unit 8 to the control unit 9; Informs the dictionary loading unit 6 of the loading instruction of the group A. Since nothing is loaded in the dictionary memory 10 at first, the dictionary loading unit 6 loads the dictionary file “DIC01” based on the dictionary configuration information. "
DIC01 "includes a voice feature sequence of a word group belonging to both the group A and the group B. The comparison / collation unit 14
Can be recognized by comparing the voice feature converted based on the voice input from the microphone 11 with the dictionary portion of the group A.

Next, from the operation instruction input unit 8, "group B
Control unit 9 ".
When the instruction to load the group B is transmitted from the dictionary loading unit 6, the dictionary loading unit 6 also includes the group B in the dictionary file 4 named “DIC01” based on the dictionary configuration information from the dictionary configuration information memory 5, which has already been loaded into the dictionary memory 10. Know that you are doing nothing in this case. The comparison / collation unit 14 can recognize the input voice by comparison / collation with the dictionary part corresponding to the group B.

Further, assuming that an operation instruction "recognize in group C" is given from the operation instruction input unit 8 this time, the dictionary loading unit 6 uses "DIC02" based on the dictionary configuration information from the dictionary configuration information memory 5. Is to be loaded, and this is loaded into the dictionary memory 10.

In the present embodiment, the word set 1 has been described as a simple set of words, but it may be a continuous word set in which a plurality of words are connected. For example, when the “article code” is formed by a combination of one letter of the alphabet and three letters of the number, it may be interpreted as a continuous word set including a basic word group of the alphabet and a basic word group of the number.

When the dictionary generation unit 3 does not find a combination satisfying the above-mentioned logical condition, for example, when the conversion capacity of a word set belonging to a certain group is larger than the capacity of the dictionary memory 10, an appropriate alarm is notified. Having a function is a matter of course.

As an example of application to a large vocabulary word dictionary,
For example, when considering a large vocabulary of "place names" nationwide, if all the "place names" cannot be recognized at one time due to the capacity limitation of the dictionary memory 10, the "prefecture names" Next, if the word set is hierarchically divided into "city names" within the prefecture, the control for loading the dictionary for each utterance by the application of the present invention will be significantly simplified. Keep it.

As described above, in the present embodiment, the dictionary generation unit 3 refers to the grouping information from the grouping information input unit 2 that represents a set of a plurality of word sets 1 to be simultaneously recognized, and The word set 1 is input, converted and output to the minimum dictionary file 4, and at the same time, dictionary configuration information is generated and stored in the dictionary configuration information memory 5. When receiving a dictionary loading request, the dictionary loading unit 6 determines the dictionary file 4 to be loaded by referring to the dictionary configuration information and checks whether the dictionary file 4 has already been loaded. Is read into the dictionary memory 10 to prevent the operator from erroneously loading the dictionary file 4 different from the recognition target into the dictionary memory 10, and the dictionary already loaded is wasted. It is possible to eliminate the need for reloading, and furthermore, it is not necessary for the operator to know the contents of the plurality of generated dictionary files 4, so that the management work of the dictionary files 4 by the operator can be significantly reduced. There is.

Furthermore, configuring the minimum number of dictionary files 4 means that when the same word set 1 is stored in a plurality of dictionary files 4 in duplicate, the degree of duplication is minimized. As a result, there is an effect that the total capacity of the dictionary file 4 can be reduced.

Further, in this embodiment, for example, in a word set of "article name", if only a specific word may be recognized at a certain point in time, those words are referred to as a word set of "article name 1". In addition, by dividing the remainder into a word set of "article name 2" and rewriting the grouping information in accordance with this, the grouping can be easily changed without increasing the total dictionary file capacity accompanying the change. There is also an effect.

[0036]

As described above, according to the present invention, the storage means for storing a plurality of words to be subjected to voice recognition in a plurality of predefined genres, and extracting the voice characteristics of each of the plurality of words. In addition, a plurality of words are classified into a plurality according to input of pre-defined grouping information for grouping a plurality of words, and stored in a plurality of files as a dictionary of voice information including voice features, and a file configuration by the plurality of classifications is shown. Dictionary file generating means for generating and storing the dictionary configuration information, and when grouping information related to the input voice is input, the voice of the word of the group related to the input grouping information with reference to the dictionary configuration information Word dictionary generating means for checking whether information is stored in the dictionary memory, selecting from a plurality of files and storing the information in the dictionary memory This has the effect of preventing the operator from accidentally loading a dictionary file different from the recognition target, and the effect of preventing the already loaded dictionary from being reloaded unnecessarily. There is.

Further, since the operator does not need to know the contents of the plurality of generated dictionary files, there is an effect that the management work of the dictionary files by the operator can be remarkably reduced.

Further, by configuring the minimum number of dictionary files, when the same word set is stored in a plurality of dictionary files redundantly, the degree of the duplication is minimized. Has the effect of reducing the total capacity of the dictionary file.

[Brief description of the drawings]

FIG. 1 is a block diagram showing one embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of a dictionary loading unit in FIG.

FIG. 3 is a diagram illustrating an example of a genre-specific word set in the embodiment.

FIG. 4 is a diagram illustrating an example of grouping information according to the present embodiment.

FIG. 5 is a diagram illustrating five configuration examples of a dictionary file according to the present embodiment, and (a), (b), (c), (d), and (e) are diagrams respectively illustrating the configurations.

FIG. 6 is a diagram illustrating an example of dictionary configuration information according to the present embodiment.

FIG. 7 is a block diagram showing an example of a conventional voice recognition device.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 word set 2 grouping information input unit 3 dictionary generation unit 4 dictionary file 5 dictionary configuration information memory 6 dictionary loading unit 8 operation instruction input unit 9 control unit 10 dictionary memory 11 microphone (microphone) 12 audio feature extraction unit 13 audio feature memory 14 Comparison / collation unit 15 Recognition result output unit

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yasuo Tomooka 5-7-1 Shiba, Minato-ku, Tokyo Within NEC Corporation (56) References JP-A-4-291398 (JP, A) JP-A-4 JP-A-160636 (JP, A) JP-A-3-119397 (JP, A) JP-A-62-1103699 (JP, A) JP-A-60-2997 (JP, A) (58) Fields investigated (Int. . ^6, DB name) G10L 3/00 571 G10L 3/00 531 G06F 9/24

Claims (5)

(57) [Claims] 1. A storage means for storing a plurality of words to be subjected to voice recognition in a plurality of predefined genres, extracting voice characteristics of each of the plurality of words, and grouping the plurality of words. According to the input of the pre-defined grouping information for converting into a plurality of categories, the dictionary configuration information indicating the file configuration according to the plurality of classifications is stored in a plurality of files as a dictionary of the voice information including the voice features and stored in a plurality of files. Dictionary file generating means for generating and storing, and when the grouping information related to the input voice is input, the dictionary information is referred to the dictionary configuration information, and the voice information of the words of the group related to the input grouping information is converted to the dictionary. A word dictionary generating means for selecting from the plurality of files after checking whether the file is stored in a memory and storing the word dictionary in the dictionary memory; A large vocabulary speech recognition device comprising: a speech selection unit for extracting a speech feature from speech, comparing and collating with the speech information extracted from the dictionary memory, and selecting a word having the most similar speech feature. . 2. The storage means has a plurality of word sets for storing a plurality of words to be subjected to speech recognition for each of a plurality of genres defined in advance, and the dictionary file generation means has a function related to the input speech. A grouping information input unit for inputting the predefined grouping information; selecting a word from the plurality of word sets according to the input of the grouping information; extracting a voice feature of the word to a plurality of dictionary files; A dictionary generation unit to store, the plurality of dictionary files storing voice characteristics of the plurality of words, and the dictionary configuration information generated based on the grouping information input to the grouping information input unit. 2. The large vocabulary speech recognition device according to claim 1, further comprising: a dictionary configuration information storage unit that performs the processing. 3. A method according to claim 1, wherein said word dictionary generating means selects a speech feature of a word previously selected based on said dictionary configuration information and classified by said grouping information before inputting said input speech, and whether said speech feature is stored in said dictionary memory. A dictionary loading unit that selects from the plurality of dictionary files after checking and stores the dictionary memory in the dictionary memory, and the dictionary memory that stores speech characteristics of the plurality of words selected from the dictionary file. The large vocabulary speech recognition device according to claim 1, wherein 4. A voice feature extraction unit for extracting voice features of the input voice input from a microphone, a voice feature memory for storing the extracted voice features, and a voice feature memory. 4. A comparison / comparison unit for comparing and collating the selected speech feature with a speech feature selected from the dictionary memory, recognizing a word corresponding to the input speech, and outputting the result. Large vocabulary speech recognition device described. 5. The method according to claim 1, wherein the dictionary generation unit writes the plurality of word sets to be simultaneously subjected to voice recognition as a first condition to the one dictionary file when the plurality of dictionary files are generated for the input voice. The same word set as the second condition is sometimes written to the plurality of dictionary files, and the third condition is to store the word set so that each of the plurality of dictionary files does not exceed the capacity of the dictionary memory, 3. The large vocabulary speech recognition apparatus according to claim 2, wherein the dictionary file is configured such that the number of the dictionary files is minimized based on the grouping information.