JPS6057395A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device that enables a machine to perform operations in accordance with human voice.

Conventional configurations and their problems Speech recognition devices that recognize human voices are effective in various fields such as input into computers and word processors, and work such as sorting items by voice. It is attracting attention as an input method.

In conventional word speech recognition devices, single word speech is input through a microphone, acoustic analysis is performed, feature parameters are extracted, and comparisons are made with academic language standard patterns stored as word dictionaries. , degree of certainty (
The word with the highest likelihood) was output as the recognition result.

However, with this method, as the number of words to be recognized increases, the proportion of similar words in the dictionary increases.
In order to ensure a practical level of recognition rate, it is necessary to limit the number of words to be recognized to a decimal number, or to equip the device itself with a response function. There are problems in that it is necessary to output a word with a high likelihood, or to output multiple candidates and have the user select one, which limits its use and is time-consuming.

OBJECT OF THE INVENTION The present invention solves the above-mentioned problems with conventional word voice input devices, and provides a more practical voice input device that can handle a large number of words, has a high recognition rate, and is easy to operate. The purpose is to realize a recognition device.

Structure of the Invention The present invention has been made to achieve the above object, and includes n word dictionary strings each having at least one word dictionary consisting of a plurality of word dictionary entries, and the first column (i=1.・・・
...-, n-1) word dictionary items constituting the word dictionary of the word dictionary string and the word dictionary in the word dictionary string of the (i+1)th column are arranged so that there is a correspondence relationship. A voice consisting of n words is input to the group storage section, and (i+
1) A word dictionary selection unit that selects a corresponding word dictionary in the word dictionary column of the column, compares the (i+1)th human-generated word with the selected corresponding word dictionary, and selects ( The present invention provides a speech recognition device characterized by having at least a word recognition unit that recognizes an i+1)th input word.

DESCRIPTION OF EMBODIMENTS When we search for a specific item among many items, it takes a lot of time to search from just one end. In such a case, if each object is systematically classified and organized according to its own attribute metal base, by tracing the attributes of a particular object in order from the larger classification,
It can be easily searched for.

Even in a word speech recognition device, if recognition can be performed using such a method, not only more efficient recognition can be performed, but also more reliable recognition results can be obtained.

In the case of a word speech recognition device that handles large words, if the main goal is to more accurately recognize one word from among many words, this can be achieved by using the method described above. It is possible to do so. For example, if Japanese city name words are to be recognized, one city name word group is created for the ward and streetcar to which the city belongs, and the city (ward) name word group is created by collecting the names of the cities and wards to which each city belongs. In the same way, create a tree-type hierarchical structure that collects the names of prefectures to which each ward and city belongs to form a prefecture name word group, and input - times to input prefecture name, city (ward) name, etc. , three words are uttered consecutively in the order of the city name, the recognition device obtains multiple recognition candidates in each hierarchy in this order, and calculates the positional relationship of the recognition candidates in each hierarchy on the tree and each By determining the city name based on the likelihood for the word, it becomes possible to recognize large words without significantly reducing the recognition rate.

Furthermore, if such a method is used, utterance can be performed with a single input, so there is no need for the troublesome procedures required in conventional methods that mainly involve question and answering.

Here, the likelihood value used for word recognition can be determined as follows.

If Ω is a word dictionary and D is a dictionary item, the recognition result when input information X is given is given by D, which maximizes the posterior probability of occurrence of the dictionary item when input information X is given. . This relationship is expressed as DeQ P(x) according to Bayes' theorem. Pa) is constant <, P(X) is D
Since it is common regardless of the result, it is only necessary to find D that satisfies the final condition, and the value of P(xID) at that time becomes the likelihood for the dictionary item.

As mentioned in the above description, the present invention utters a plurality of words in one input, recognizes the first attribute (for example, prefecture name) with the first utterance, and recognizes the second word. The second attribute (e.g. city name) represented by the recognition result of the first word is recognized by mouth, and the third attribute (e.g. city name) represented by the recognition result of the second word is recognized between the third words. Recognition is performed hierarchically, such as recognizing from the ward name).

The details of this embodiment will be explained below with reference to the drawings.

FIG. 1 is a diagram showing an example of the structure of a word dictionary tree used for recognition in this embodiment. FIG. 2 is a diagram showing an example of the configuration of the single-word speech recognition device according to this embodiment. The present embodiment will be described with reference to FIGS. 1 and 2, taking as an example a case in which city name words are recognized.

In Figure 1, 1 to 6 each represent a word dictionary,
6 represents a word dictionary string of the first layer, 7 represents a word dictionary string of the second layer, and 8 represents a word dictionary string of the third layer. Each word dictionary is
It consists of word names that have attributes indicated by words in higher hierarchy. The city name words to be recognized are in the word dictionary string 8 in the third layer, and the word dictionary string 7 in the second layer contains word dictionaries consisting of city (ward) names to which each city belongs.
The word dictionary row 6 in the first layer contains word dictionaries consisting of the names of prefectures to which each city (ward) belongs.

When recognizing a city name word, in one input, say the name of the prefecture to which the city belongs, the name of the city (ward), and the name of the city, dividing all three words in that order, and separate the input word and dictionary for each layer. Words are matched, and the words with the highest likelihood are selected as recognition candidates.

First, in the simplest case, the branches of the first recognition are sequentially traced from the top layer of the tree, and the recognized city name is output.
OOKYOOTO (Tokyo) / CHIYODAKU (Chiyoda-ku) 7 KANDA (Kanda) / and the case where each word is uttered separately will be explained.

In FIG. 2, in the initial state, the value of the layer counting section 13 is 1, and the word dictionary selection section 12 selects the prefecture name dictionary in the word dictionary string 6 of the first layer. When the first word is input from the voice input unit 9, the word recognition unit 1o retrieves the prefecture name dictionary selected by the word dictionary selection unit 12 from the word dictionary group storage unit 11, and selects one word from it. (for example, Tokyo) is recognized, and the result is sent to the word determination section 14 and word dictionary selection section 12. At this time, the value of the layer counting section is +
1, indicating that the next recognition is for word dictionary string 7 in the second layer. The word selection unit 12 selects the word dictionary to which the next word to be input belongs based on the recognition result of the first idiom and the value of the hierarchy counting unit 13. When the second word is input from the voice input unit 9, the word recognition unit 1o
The ward name dictionary of the word dictionary row 7 in the second @ layer selected by the word dictionary selection section 12 is retrieved from the word dictionary group storage section 11, one word (for example, Chiyoda Ward) therein is recognized, and the result is is sent to the word determination section 14 and the word dictionary selection section 12. At this time, the value of the layer counting unit 13 is further incremented by 1 to indicate that the next recognition is for the word dictionary string 8 of the third layer. The word dictionary selection unit 12 uses the recognition result of the second word and the hierarchy counting unit 10.
The word dictionary to which the next word to be input belongs is selected depending on the value of . When the third word interval is input to the voice input unit 9, the word recognition unit 1o retrieves the city name dictionary of the word dictionary row 8 in the third layer selected by the word dictionary selection unit 12 from the word dictionary storage unit 11. One of the words (for example, Kanda) is recognized, and the result is sent to the word determination section 14. When the word recognition for all layers is completed in this way, the word judgment unit 14 obtains the recognition result /Chiyoda-ku, Tokyo 7/Kanda/, and some or all of these are the final recognition results. is output as In this case, /KANDA/
In order to be recognized, /TOO is required in the first layer word dictionary string.
KYOOTO/, second layer word dictionary string/CHIYO
DAKU/ and /KANDAZ must be correctly recognized in the word dictionary string of the third layer. but,
The number of dictionary items in the word dictionary at each level is much smaller than the number of dictionary items in a conventional dictionary that includes all city names in one word dictionary, so the word recognition rate at each level is extremely high. Therefore, the value of their product, which is the final word recognition rate, also becomes high.

Note that, in the case where the number of dictionary entries in each word dictionary is limited to sufficiently reduce the rate of erroneous recognition as in this embodiment, the word determination unit 14 is not necessarily necessary. FIG. 3 shows a word recognition flowchart when the input speech is three words and the word dictionary string is three layers, and the number of recognition candidates is narrowed down to one.

Next, in each layer, if the number of recognized word candidates is not limited to one but is allowed up to several candidates, the probability that a correct word is included among them is further increased. In this case, the second
In the example in the figure, when the first prefecture name word is recognized,
The word recognition unit 10 recognizes, for example, the third most likely dictionary entry between the input word and the prefecture name word dictionary as recognition candidates, and the results are sent to the word determination unit 14 and the word dictionary selection unit 12. sent to. The word dictionary selection unit 12 selects three famous word dictionaries corresponding to the top three recognition candidates. Next, when the city (ward) name word as the second word is recognized, the word recognition unit 10 selects a word with a high likelihood between the input word and the dictionary item for each of the three city (ward) name word dictionaries. For example, the top three words are sent to the word determination section 14 and word dictionary selection section 12 as recognition candidates. The word dictionary selection unit 12 selects nine city name word dictionaries for each recognition candidate. Finally, the third
When a city name word in the word category is recognized, the word recognition unit 10 selects the word with the highest likelihood between the input word and the dictionary entry for each of the nine city name word dictionaries and sends it to the word determination unit 14 as a recognition candidate. Sent.

When word recognition for all layers is completed in this way, the word determination unit 14 creates nine combinations of word tree sequences that are connected to prefecture names, dragon prisoner names, and city names. Here, the likelihood of the dictionary entry for the input of the first word is set as Si (-1, 2, 3) in descending order, and the likelihood of the dictionary entry of the i-th dictionary for the input of the second word is set as Si (-1, 2, 3). 8 (j = 1.2.3) in descending order of magnitude, and the maximum likelihood of the dictionary entry of the i-th dictionary for the input of the third word is 5iJ1, then the judgment by the word judgment unit 14 The rule is, for example, the sum of likelihoods for each of nine word tree sequences L L-31+S is +S i) 1 ('””'
1 + 2 + 3 : ) = 1 + 2 + 3 ) or the sum of likelihoods weighted by the number of dictionary entries.

Lω=cc+iS,+cc+1)S11+Qliil
5ij1(1-1+2+3+]=142+
3) All the words of the word tree series with the largest value (ω1.ωij+ωij1 is the weight based on the number of dictionary entries in each hierarchy) or the city name word therein may be used as the recognition result. In this way, by creating multiple recognition candidates in each layer and determining the recognition result in units of word tree series from the likelihood value in each layer, the correct word can be determined at the recognition stage in the upper layer. It is possible to prevent the branches of the tree from falling off and further improve the recognition rate.

FIG. 4 shows a word recognition flowchart in the case where the input speech is three words, the word dictionary string has a three-layer structure, and a plurality of recognition candidates are allowed. In this case, the number of recognition candidates is uniformly determined to be m for the first input word and n for the second input word.

In this example, if we assume that each word dictionary in the word dictionary strings from the first layer to the third layer consists of 100 words and is a general word with no special recognition tendency, A word speech recognition device that has a recognition rate of 96.6% for 100 words targeted at a speaker can achieve a recognition rate of 90% or more for 100,000 words. Further, the amount of newly created word dictionaries for creating a hierarchical structure at this time increases by only 1.1%, which is sufficient for practical use.

In addition, the number of matching operations between the symbol pattern of the dictionary item and the input cover pattern is calculated when the number of dictionary items in each word dictionary is n.
In the case of 3 layers, the ratio between selecting only the first candidate with a hierarchical structure and not using a hierarchical structure is approximately 3n:n
3 Therefore, by using this embodiment, the calculation time can be significantly reduced. The word dictionary items constituting the word dictionary of the i-th column (i = 1..., n-1) word dictionary string and < i + 1 are provided.
A word dictionary group storage part L arranged so as to have a correspondence relationship with the word dictionary in the '>'-y-th word dictionary string, inputs speech consisting of n words, and inputs the sound consisting of n words. a word dictionary selection unit that selects a corresponding word dictionary in the (l+1)th word dictionary string of the word dictionary group storage unit based on the word recognition result; and an (i+1)th input word and the selected corresponding word dictionary. Based on the degree of certainty, (i+
1) A speech recognition device is provided which is characterized by having at least a word recognition unit that recognizes the first input word, and by layering tens to hundreds of words as one word dictionary, The increase in the number of dictionary entries to be created is kept to a few inches or less, and the two-layer structure can handle thousands to tens of thousands of words, and the three-layer structure can reduce the number of words from tens of thousands to millions of words. It has become possible to recognize words with a high recognition rate, and it has become possible to easily realize a speech recognition device for large words, which has not been put into practical use in the past. Since it is not necessary to perform matching with patterns for all dictionary items, calculation time is significantly reduced.

[Brief explanation of drawings]

Fig. 1 is a diagram showing an example of the configuration of the word dictionary tre'e of the present invention, Fig. 2 is a block diagram of a word speech recognition device which is an embodiment of the present invention, and Figs. 4ζ°゛Aj showing an example flowchart. 1 to 6... Word dictionary, 6... First layer word dictionary string, 7... Second layer word dictionary string,
8...3rd layer word dictionary string, 9...
Voice input section, 10... Word recognition section, 11...
. . . word dictionary group storage section, 12 . . . word dictionary selection section, 13 . . . layer side section, 14 . . . word determination section.

Claims (2)

[Claims] (1) Provide n columns of word dictionary strings each having at least one word dictionary consisting of a plurality of word dictionary entries, and set the i-th column (i=1
．． ......, arranged so that there is a correspondence relationship between the word dictionary items constituting the word dictionary of the word dictionary string of (n-1) and the word dictionary in the word dictionary string of the (i+1)th column. Input the word dictionary group storage unit and the audio consisting of n words,
a word dictionary selection unit that selects a corresponding word dictionary in the (i+1)th word dictionary string of the word dictionary group storage unit based on the recognition result of the first input word; and the (i+1)th input word and the selection; A speech recognition device comprising at least a word recognition unit that compares the selected corresponding word dictionary and recognizes the (i+1)th input word based on its degree of certainty. (2) The recognition results of the word recognition unit are ranked, a plurality of recognition result candidates are selected, and a word dictionary is selected corresponding to each of the candidates, so that a group of recognition result candidates for all input speech is selected. 2. The speech recognition device according to claim 1, wherein the input word is recognized based on the degree of certainty among the input words.