JPS62219000A - Word 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] [Summary] A syllable candidate string is extracted by comparing the input word sound with a series of standard patterns corresponding to syllables registered in advance, and a syllable candidate string is extracted from the character series of the word to be recognized. Find the distance for each digit between the character sequence corresponding to each word having the number of syllables corresponding to the number of syllables and the syllable candidate string based on an inter-syllable distance memory in which the distance between each syllable is registered, The word with the minimum distance is output as the recognition result. As a result, it is possible to recognize virtually infinite word length or human word speech using a small-capacity storage device, and the amount of recognition processing can be reduced.

[Industrial application field]

The present invention relates to a word speech recognition device using syllables as recognition units, and particularly to a word speech recognition device suitable for recognizing word groups of human speech.

[Conventional technology]

Conventionally, as a typical method for recognizing word sounds,
As shown in Figure 5 (A), a method using words as the recognition unit,
As shown in FIG. 5B, there is a method in which syllables (or phonemes) smaller than words are used as recognition units.

In the word speech recognition method using words as recognition units shown in FIG. 220.

On the other hand, in the word standard pattern section 230, the analysis results of recognition target words are registered in advance as standard patterns.

The word recognition unit 220 compares the standard pattern of each word in the word standard pattern unit 230 with the parameters of the input speech pattern, and outputs the word of the standard pattern with the smallest distance as a recognition result.

In the word speech recognition method using syllables (or phonemes) as recognition units shown in FIG. An input syllable pattern is created and input to the syllable (or phoneme) recognition unit 250. A phoneme is a vowel or a consonant,
A syllable is made up of phonemes, usually formed by combining one vowel and one or two consonants.

In the syllable (or phoneme) standard pattern section 260, the analysis result of the recognition target syllable (or phoneme) is registered in advance as a standard pattern.

The syllable (or phoneme) recognition unit 250 calculates each distance by comparing the parameters of the input syllable (or phoneme) pattern of the input speech and the standard pattern of each syllable (or phoneme) in the syllable (or phoneme) standard pattern unit 260. , is input to the word recognition section 270.

The word dictionary section 280 stores each word to be recognized expressed as a series of its syllables (or phonemes).

The word recognition unit 270 calculates the distance between syllable (or phoneme) sequences for each word in the word dictionary unit 280 based on the distance information for each syllable (or phoneme) sequence manually input by the syllable (or phoneme) recognition unit 250. , the word with the smallest value is output as the recognition result.

The word speech recognition method that uses words as recognition units has a good recognition rate, but on the other hand, the amount of processing and the storage capacity of standard patterns increase rapidly as the number of words increases.

For this reason, it is disadvantageous in the case of large words such as 1000 or more.

On the other hand, word speech recognition methods that use syllables (or phonemes) as recognition units are slightly inferior in terms of recognition rate, but the distance calculation required for recognition is This method has the advantage that even if the number of words increases, the increase in processing amount and storage capacity for standard patterns is much smaller than in the word speech recognition method, which uses words as recognition units.

For this reason, word speech recognition methods that use words as recognition units are generally advantageous when the number of words is small.
When a person speaks a large number of words, a word speech recognition method that uses syllables (or phonemes) as recognition units is said to be advantageous.

[Problem that the invention seeks to solve]

The word speech recognition method that uses syllables as the recognition unit has the characteristic that even if the number of recognized words increases, the increase in processing amount and storage capacity is much smaller than the word speech recognition method that uses words as the recognition unit. ing. However, in the case of recognition of human word speech, the amount of processing and the absolute storage capacity of the word dictionary section are still large. For this reason, in a word speech recognition device that targets word groups of human language, it is a problem to further reduce the amount of processing and the storage capacity of the word dictionary section.

SUMMARY OF THE INVENTION An object of the present invention is to provide a word speech recognition device that uses syllables as a recognition unit, which reduces the amount of processing and the storage capacity of a word dictionary section compared to conventional methods.

[Means for solving problems]

The means taken by the present invention to solve the above-mentioned problems in the conventional word speech recognition method using syllables as recognition units will be explained with reference to FIG.

FIG. 1 is a block diagram showing the basic configuration of the present invention.

In FIG. 1, 110 is a syllable candidate string extraction unit that compares the input word sound with a series of pre-registered standard patterns corresponding to syllables and outputs a set of syllable candidate strings.

Reference numeral 120 denotes a word recognition unit, which includes an inter-syllable distance memory (121) in which the distance between each syllable is stored, and calculates the number of syllables corresponding to the number of syllables in the syllable candidate string among the character series corresponding to the word to be recognized. The distance for each digit between the character sequence corresponding to each word and the syllable candidate string is determined based on the inter-syllable distance memory (121), and the word with the minimum sum of distances is determined as the recognition result. Output.

Note that a character sequence is a sequence of syllables that make up a word arranged in the order of pronunciation of the word. In addition, the digits express the positional relationship of each syllable, and each syllable that makes up a word is numbered 1 from the beginning.
They are called the digit, second digit, etc.

[For production]

When a word sound is input from a microphone or the like, the syllable candidate string extracting unit 110 compares the input word sound with a pre-registered series of standard words corresponding to syllables, extracts a syllable candidate string, and extracts a syllable candidate string. The information is input to the recognition unit 120.

This syllable candidate string is an arrangement of syllables and does not have distance information.

The word recognition unit 120 calculates the distance for each digit between the character sequence corresponding to each word having the number of syllables corresponding to the number of syllables in the input syllable candidate sequence among the character sequences of the word to be recognized and the syllable candidate sequence. The distance is calculated based on the inter-syllable distance memory 121, and the word with the minimum sum of distances is output as the recognition result.

By doing this, only one set of syllable candidate strings is required, so that the amount of processing for extracting syllable candidate strings and the amount of processing for word recognition can be reduced. Furthermore, by connecting to a host system and receiving word set information of a desired category, it is possible to construct a word speech recognition system with virtually unlimited lead or human vocabulary using various small-capacity storage devices.

〔Example〕

Embodiments of the present invention will be described with reference to FIGS. 2 to 4.

FIG. 2 is a block diagram illustrating the configuration of one embodiment of the present invention, FIG. 3 is a diagram illustrating one embodiment of the inter-syllable distance memory, and FIG. 4 is a diagram illustrating another embodiment of the inter-syllable distance memory. be.

(A) Configuration of the Embodiment In FIG. 2, the syllable candidate string extraction section 110, word recognition section 120, and inter-syllable distance memory 121 are as described in FIG.

Reference numeral 130 denotes a microphone, which inputs word sounds pronounced by the speaker to the syllable candidate string extraction section 110.

140 is a host system called an expert system, which performs various processes based on recognized words. or,
The word set information of the category to be recognized is sent to the word recognition unit 120.

In the syllable candidate string extraction unit 110, 111 is a parameter extraction unit that extracts parameters representing the characteristics of a speech pattern from the input speech. Reference numeral 112 denotes a section detecting section, which detects sections for each single syllable based on the parameters extracted by the parameter extracting section 111.

113 is a monosyllable dictionary section in which standard patterns corresponding to syllables extracted from specific word speech or standard patterns corresponding to syllables created from monosyllabic speech are registered in advance. 1
14 is a syllable matching unit that matches the parameters of the word sound pattern corresponding to syllables inputted from the interval detecting unit 112 and each standard pattern series corresponding to syllables from the monosyllable dictionary unit 111, and selects the standard pattern with the smallest distance between them. Output as a syllable candidate string. 115 is a switching circuit that switches the output of the section detection section 112 to the monosyllable dictionary section 113 or the syllable matching section 114;

In the word recognition section 120, reference numeral 122 is a used word dictionary section in which word information for each category to which the input word sound belongs is registered. This word information is sent in advance from the expert system before the word voice is input.

Reference numeral 123 denotes a character string matching unit, which inputs the character sequence of each word having the number of syllables corresponding to the number of syllables in the input syllable candidate string among the character sequences of the word information read out from the used word dictionary unit 122. The distance of each digit to the syllable candidate string is calculated based on the inter-syllable distance memory 121, and the word with the minimum sum of distances is recognized by the expert system 140.
send to

(B) Operation of the Embodiment The operation of the embodiment shown in FIG. 2 will be explained with reference to FIGS. 3 and 4.

Before recognition processing for a word voice uttered by a speaker (not shown) is started, a standard pattern corresponding to a syllable is registered in the monosyllable dictionary section 113, and a standard pattern corresponding to a syllable is registered in the used word dictionary section 122. Word information by category is registered.

When registering a standard pattern corresponding to syllables in the monosyllabic dictionary section 113, the switching circuit 115 is connected to the monosyllabic dictionary section 113 side, and a specific word sound or monosyllabic sound is inputted from the microphone 13G. From the input word speech or monosyllabic speech, standard patterns corresponding to syllables are extracted by the parameter extraction unit 111 and the interval detection unit 112, and the monosyllable dictionary unit 1 extracts standard patterns corresponding to syllables.
Registered on 13th. The configurations and operations of the parameter extracting section 111 and the section detecting section 112 are all well known, so detailed explanation thereof will be omitted.

Since the number of monosyllables is approximately 100 and does not depend on the number of words, the monosyllable dictionary unit 113
capacity will not increase.

The used word dictionary section 122 includes an expert system 14.
Word set information for each category to which the input word sounds belong is sent from 0 and stored. When the category to which the input word voice belongs is changed, the expert system 140 sends word set information for that category, and the contents of the used word dictionary section 122 are rewritten.

By doing so, it is possible to configure a virtually infinite dielectric or human vocabulary word speech recognition system using the used word dictionary section 122 with a small capacity.

Next, when recognizing the input word voice, the switching circuit 1
15 is connected to the syllable matching section 114 side.

The word voice input from the microphone 130 is processed by the parameter extraction unit 11 as in the case of standard pattern registration described above.
1 and the section detection section 112 create an input word sound pattern corresponding to syllables, and input it to the syllable matching section 114 .

The syllable matching unit 114 uses the input word sound pattern corresponding to syllables inputted from the interval detection unit 112 and the monosyllable dictionary unit 11.
The parameters of each standard pattern series corresponding to the three syllables are compared, and the syllable series of the standard pattern with the smallest distance is extracted as a syllable candidate series. Extraction of this syllable candidate string can be performed using, for example, D P (DynaIllicprogra++u
++ing) can be done by matching process,
The syllable candidate string with the smallest distance, that is, the first syllable candidate string is extracted as the desired syllable candidate string. Since it is not necessary to obtain the second and subsequent syllable candidate sequences, the amount of processing can be reduced accordingly.

The character string matching unit 123 inputs the character sequence of each word having the number of syllables corresponding to the number of syllables of the input syllable candidate string among the character sequences of the word information read out from the used word dictionary unit 122. The distance from each digit to the syllable candidate string is determined based on the inter-syllable distance memory 121, and the word with the minimum sum of distances is output as a recognition result.

Next, the operation of the character string matching section 123 will be specifically explained with reference to the inter-syllable distance memories shown in FIGS. 3 and 4.

The inter-syllable distance memory in Figure 3 (A) is for vowels (asi*u
The distance between each syllable consisting of p'3sogN) is shown in a matrix format. For example, the distance between a and i is 50", and the distance between U and 0 is "20". The distance between the same vowels is naturally "O". Note that N stands for "n". shall be expressed.

The inter-syllable distance memory in FIG. 3(B) is based on the consonant (φ).

The distance between each syllable consisting of approximately 20 types of syllables (k, s, t, ... pj, etc.) is also shown in a matrix format. For example, the distance between k and 8 is "5",
The distance between k and n is 10”.The distance between identical consonants is
Naturally, it is "0". Note that "φ" is a virtual consonant set to the vowel "a, i, u, e, o" and expressed as "φ". For example, the vowel “a” is “(φ)a”
It is expressed as As a result, since each syllable is formally composed of a combination of a consonant and a vowel, processing can be performed uniformly and efficiently.

The distance between each syllable is determined by taking into account phonetic similarities and results obtained from actual word recognition. This also applies to the inter-syllable distance memory shown in FIG.

The inter-syllable distance memory in Figure 4 is 'Ah, I, U.

It shows the distances between about 100 types of syllables, such as ``Eh...'', in a matrix format, and the distance between each syllable is the distance calculated from Figure 3 (A) and (B). matches.

For example, for the distance between ``1 (ka)'' and ``U (φU)w,'' the same value ``55'' can be obtained by any inter-syllable distance memory.

The number of syllables is about 100 and does not depend on the number of words, so no matter which of the above-mentioned inter-syllable distance memories are used, it is possible to recognize word sounds in human languages using a small-capacity inter-syllable distance memory. I can do it.

Now, the syllable candidate string input from the syllable matching unit 114 is "
Assume that it consists of three syllables: "ha ki sa".

The character string matching unit 123 reads character sequences corresponding to each word having the same number of three syllables as the number of syllables in the syllable candidate string from the used word dictionary unit 122, and refers to the inter-syllable distance memory. Then, the distance of each digit from the input syllable candidate string is calculated.

Table 1 shows the character sequence "((φ)a kj
ta) ” and the character sequence corresponding to “Sakai” (s
a ka (φ)i)'' and the syllable candidate string “ha
This figure shows an example of finding the distance between "kita".

The distance between the character series "(φ)a ki ta" corresponding to "ha ki sa" and "Akita" is "ha" and "Akita".
(φ) between a″ is 65″, @ki m and ki″″ is “
θ'', the distance between “sa” and ta is “4′”, so the total is “9”.

Similarly, the distance between "ha ki sa" and the character sequence "sa ka (φ)i" corresponding to "Sakai" is "106". These distance values are stored in the inter-syllable distance memory in Figure 4. The same result can be obtained using .

In addition, Okinawa shown in (C) of Table 1, (φ) o ki
It can be considered that words with different numbers of syllables, such as "na wa", have no possibility of becoming recognized words, so a distance or code such as oo (infinity), which is not a recognized word, is given.

By using inter-syllable distance memories such as those illustrated in FIGS. 3 and 4, distances between character sequences corresponding to words read from the syllable candidate string and the used word dictionary section 122 can be calculated using conventional DP matching. The amount of processing required is significantly smaller than that of the method.

The word with the smallest distance determined as above (Table 1
In the example, “Akita”) is extracted as a recognized word and sent to the expert system 140.

Although one embodiment of the present invention has been described above, the present invention includes
It can also be implemented as follows.

That is, without providing the used word dictionary section 122 in the word recognition section 120, the same content as the character sequence corresponding to word information read out from the word recognition section 120 may be directly received from the expert system 140 during word recognition processing. .

In addition, if the number of syllables cannot be detected with high accuracy through the parameter-level matching performed in the syllable candidate string extraction unit 110, multiple syllable sequences with different numbers of syllables are extracted from the unknown input word speech, and the syllable sequence is extracted from the unknown input word speech. The character string matching unit 12 of the word recognition unit 120 calculates the syllable candidate string and the cumulative distance for each different length.
Enter 3.

The character string matching unit 123 performs character string matching for each syllable,
By weighting the distance in parameter level matching and the distance in string matching, the recognition results are obtained by adding values.
Accuracy can be increased.

〔Effect of the invention〕

As explained above, according to the present invention, the following effects can be obtained.

(b) Since it is possible to easily connect to a host device such as an expert system and receive desired word set information from this host device, the words to be recognized can be changed at any time using various storage devices with a small capacity. This makes it possible to construct a speech recognition system with virtually unlimited word usage or human speech.

(b) Since word recognition processing is performed by extracting only the first syllable candidate string with the smallest distance, the amount of processing for extracting syllable candidate strings and the amount of word recognition processing can be reduced.

[Brief explanation of drawings]

Fig. 1...Explanatory diagram of the basic configuration of the present invention, Fig. 2...
An explanatory diagram of one embodiment of the present invention, FIG. 3: An explanatory diagram of an inter-syllable distance memory used in the same embodiment, FIG. 4: An explanation of another inter-syllable distance memory used in the same embodiment Fig. 5: An explanatory diagram of a conventional word speech recognition method. In FIG. 1 and FIG. 2, 110... Syllable candidate string extraction unit, 120... Word recognition unit, 121... Inter-syllable distance memory, 122... Used word dictionary unit, 123... Character Column matching section, 130...
Microphone, 140...Expert system.

Claims (3)

[Claims] (1) In a word speech recognition device whose recognition unit is a syllable, (a) The input word speech is compared with a series of pre-registered standard patterns corresponding to syllables, and a set of syllable candidate sequences is output. (b) an inter-syllable distance memory (121) in which the distance between each syllable is stored; The distance for each digit between the character sequence corresponding to each word having the number of syllables corresponding to the number of syllables and the syllable candidate string is calculated based on the inter-syllable distance memory (121), and the sum of these distances is the minimum. A word speech recognition device comprising: a word recognition unit (120) that outputs a word as a recognition result. (2) The word recognition unit (120) includes a used word dictionary unit (122) that receives and stores word information of the category to which the input word voice belongs from the host device before inputting the word voice;
The distance for each digit between the character sequence corresponding to each word having the number of syllables corresponding to the number of syllables in the voice candidate string read from this used word dictionary section (122) and the syllable candidate string, Determined based on the inter-syllable distance memory (121),
2. The word speech recognition device according to claim 1, wherein the word for which the sum of distances is the minimum is output as a recognition result. (3) The word recognition unit (120) recognizes each word having the number of syllables corresponding to the number of syllables in the voice candidate string among the character series corresponding to the word information of the category to which the input word voice received from the host device belongs. The distance for each digit between the corresponding character sequence and the syllable candidate string is stored in the inter-syllable distance memory (12
1) and outputs the word for which the sum of the distances is the minimum as a recognition result.