JPH07230299A - Voice recognition device - Google Patents

Abstract

PURPOSE:To roughly discriminate which one of the word groups beforehand prepared inputted voice belong to employing a group discrimination section which is made up of neural network. CONSTITUTION:Feature patterns of the voice signals inputted through a voice input section 1 is generated by a feature pattern generation section 2. Based on the feature patterns, a group discrimination section 3 constituted of neural network discriminates which one of input items 1, 2 or 3 the word belongs to. Then, the words stored in the input items are discriminated by word recognition sections 4, 5 or 6 of discriminated input items 1, 2 or 3.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention recognizes an input voice by a group identification unit for identifying which group the input voice belongs to and a word identification unit for identifying which word in each group it corresponds to. Voice recognition device.

FIG. 5 is a block diagram showing the structure of a conventional voice recognition apparatus, and reference numeral 11 in the figure denotes a voice input unit.
The voice input unit 11 includes a microphone, a microphone amplifier, and an AD.
The voice input through the converter is input to the characteristic pattern creating unit 12 that extracts a characteristic pattern as a voice signal by frequency spectrum analysis or the like. The characteristic pattern creating unit 12 extracts a frequency spectrum pattern that is a characteristic of the audio signal and inputs it to each word identifying unit 13, 14 or 15.

To each of the word identifying parts 13, 14 and 15, for example, a group of contents representing characteristics of "person", that is, input items such as sex, birthplace and age are assigned, and in the case of sex, male, Words for women, men, women, etc., in the case of birthplace, words for Tokyo, Osaka, etc., and for older people, 1
Word recognition unit 1 for words in their 0s, 20s, 30s, etc.
It is stored separately in the memories assigned to 3, 14, and 15, and can be operated sequentially in accordance with instructions from the key input unit 16 and the recognition result control unit 17.

For example, when the word input unit 11 inputs the voice of "male" when the word identifying unit 13 for the input item 1 in which a word regarding gender is first stored in the memory is operated. The word identifying unit 13 compares the feature pattern input from the feature pattern creating unit 12 with a previously stored word representing gender, identifies a word corresponding to the feature pattern, and when a corresponding word exists. Outputs a signal indicating that the corresponding word has been recognized to the recognition result control unit 17.

Next, the word input section 1 for the input item 2 in which the word concerning the place of birth is stored in the memory is used as the key input section 1.
The operation of 6 is set to the operating state, and the word identifying unit 14 is set to the standby state. When the voice relating to the place of birth is input and the word corresponding to the input voice relating to the place of birth is recognized, the word identifying unit 15 for the input item 3 in which the word relating to the age is stored in the memory is input by the key input unit 16. When the voice is input in the operating state and the word corresponding to the age is input, and the word corresponding thereto is recognized, the key input unit 16 again sets the word recognition unit 13 in the operating state and waits.

By the way, in such a conventional apparatus, since the word identification units 13, 14, 15 are sequentially operated to identify the word corresponding to the input voice, there is a problem that the operation is troublesome. It was As a countermeasure against this, a technique has been proposed in which the characteristic pattern of the input word speech is analyzed, rough identification is performed for each group prepared in advance, and then the speech recognition is performed by identifying the words prepared in each group. (Japanese Patent Publication No. 2-52278: G1)
0L 3/00).

In this prior art, words are preliminarily divided into groups by using a clustering method, the center coordinates of each group are obtained, and this is used as a representative standard pattern, and the analysis pattern of the input voice is collated with the representative standard pattern. The group of the representative standard pattern having the minimum mutual distance is identified as the corresponding group.

FIG. 6 is an explanatory diagram showing a group classification mode by the above-mentioned conventional apparatus. There are now six types of words, and if these are indicated by symbols ○, ●, ◇, ◆, □, △, etc., the three words ○, ●, and ☆ are group G ₁ , due to the physical characteristics of each word. The two words ◇ and ◆ are grouped into a group G ₂ , the two words □ and # are grouped into a group G ₃ , and the two words Δ and ▽ are grouped into a group G ₄ , respectively. Then, the analysis pattern of the input voice is collated with the representative standard pattern of each group, the group of the representative standard pattern having the smallest mutual distance is identified, and then input for each group as in the conventional technique shown in FIG. The voice is matched with the words in the group for identification.

[0009]

However, in such a method, the groups are grouped by using a clustering method, and as is apparent from FIG. 6, it is often erroneously recognized that the physical feature similarity is low. Therefore, the grouping is unavoidable because it depends on the physical characteristics of the target word, and the grouping is based on the voice itself, and there is a problem that the subdivision of the group cannot be avoided.

The present invention has been made in view of the above circumstances, and an object of the present invention is to identify a group by using a neural network capable of having an identification function according to learning content, thereby physically inputting speech. An object of the present invention is to provide a voice recognition device that enables grouping even when there are few characteristic features. Another object of the present invention is to allow the control section to sequentially operate the word identifying section for each group based on the identification result of the group identifying section composed of a neural network, thereby enabling the conventional key input section to operate. It is to provide a voice recognition device that does not require operation.

[0011]

According to a first aspect of the present invention, a voice recognition device sequentially inputs a voice corresponding to each predetermined group of words, and recognizes a word corresponding to a voice input in the group. In the voice recognition device,
A voice input unit, a feature pattern creation unit that creates a feature pattern of an input voice, and a neural network that is trained to enable identification of which group the input voice belongs to based on the feature pattern And a word identifying unit that identifies a corresponding word from each of the groups identified by the group identifying unit based on the characteristic pattern. Another voice recognition device according to the second aspect of the invention is characterized by including a control unit that switches a group to be input next voice based on the identification result of the group identification unit.

[0012]

According to the first aspect of the present invention, by setting the learning pattern for the neural network according to the application, even if the physical characteristics of the input voices are low in similarity, the input voices can be classified by group. Can be identified. In the second invention, the control unit can sequentially and selectively operate the word identifying units according to the identification result of the group identifying unit configured by the neural network.
It is possible to automatically identify the input voice.

[0013]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be specifically described below with reference to the drawings showing the embodiments. FIG. 1 is a block diagram showing the configuration of a voice recognition device according to the present invention, in which 1 denotes a voice input unit. The voice input unit 1 is composed of a microphone, a microphone amplifier, an AD converter, and the like, and the voice input through the voice input unit 1 is output to the characteristic pattern creating unit 2 as a voice signal. The characteristic pattern creating unit 2 obtains the frequency spectrum of the input voice signal, extracts the feature, and outputs this to the group identifying unit 3 configured by a neural network.

As the characteristic pattern creating section 2, for example, 1
A short-time spectrum method for extracting a frequency characteristic of a section of about ˜20 ms, a method for extracting a coefficient sequence representing this, or the like is used, but not limited to this, another conventionally known method is used. May be.

The group identification unit 3 composed of a neural network is a word identification unit for each group in which words grouped in advance are input based on the input characteristic pattern, that is, word identification for the input item 1. The part 4, the word identifying part 5 for the input item 2 and the word identifying part 6 for the input item 3 are matched to which of the word identifying parts 6 the voice signal belongs,
It discriminates and outputs a signal to any of the corresponding word discriminating units 4, 5 or 6 and also outputs a signal to the recognition result control unit 7.

FIG. 2 is an explanatory diagram showing the relationship between the output layer neurons of the neural network and the word identifying units 4, 5, and 6. The neural network is usually configured in a hierarchical structure composed of three layers of an input layer neuron, an intermediate layer neuron (none of which are shown) and an output layer neuron,
Each layer includes one or a plurality of neurons, and the input layer neurons and the intermediate layer neurons are coupled with different coupling coefficients between the intermediate layer neurons and the output layer neurons. In the embodiment, the word identifying units 4, 5 and 6 are respectively associated with three output layer neurons O ₁ , O ₂ and O ₃ .

[0017] Figure 3 is an explanatory view showing an example of grouping in the case of using the group identification portion 3 composed of a neural network, now word ○ as shown in FIG. 3, for example, × the group K _1, also Words □ and ☆ are group K
₂ , the words ●, △, and ▽ are grouped into the group K ₃ , and the ◇ and ◆ are grouped into the group K ₄ , and the above-mentioned words are given to the group identification unit 3 configured by the neural network. Are repeatedly learned so that it can be discriminated which _one of the groups K _{1 to} K ₄ belongs to, respectively. As is clear from comparison between FIG. 3 and FIG. 6, the groups G ₁ , G _{2 to} G ₄ shown in FIG. 6 are divided based on extremely similar physical characteristics, while the groups shown in FIG. It is understood that K ₁ , K _{2 to} K ₄ are not necessarily limited to the case where the physical characteristics are similar.

Then, for example, the words belonging to the group K ₁ are stored in the memory assigned to the word identification unit 4 for the input item 1,
The words belonging to group K ₂ are stored in the memory allocated to the word identifying unit 5 for input item 2, and the words belonging to group K ₃ are stored in the memory allocated to the word identifying unit 6 for input item 3 respectively. . As a result, when the output from the characteristic pattern creating unit 2 is input to an input layer neuron (not shown) of the group identifying unit 3 configured by a neural network, the output is input to each intermediate layer neuron, and the output of each intermediate layer neuron is The output layer neurons are given a group identification in this process, and the output layer neurons O _{1 to} O ₃ respectively output “100”, “010” or “0”.
A collation / identification signal such as "01" is output to the word identification units 4, 5, and 6.

For example, when a signal "100" is output from the output layer neurons O _{1 to} O ₃ , the word identifying unit 4 for the input item 1 is output, and when a signal "010" is output, the signal for the input item 2 is input. When the word identifying unit 5 further outputs the signal "001", the word identifying units 6 for the input items 3 are made to operate respectively, and the word identification based on the feature pattern output from the feature pattern creating unit 2 is performed. It

That is, each word identifying unit 4, 5 or 6 collates the characteristic pattern of the input voice with each word stored in advance, and when the corresponding word is recognized, an identification signal is output to the recognition result control unit 7. Output to. Next, the control operation of the recognition result control unit 7 will be described with reference to the flowchart shown in FIG.

It is assumed that the following words are stored in advance in the memories assigned to the word identifying sections 4, 5, 6 for the input items 1, 2, and 3, respectively. Input item Target word 1 Male, Female, Male, Female 2 Hokkaido, Tohoku, Kanto, Chubu, Kinki, Chugoku, Shikoku, Kyushu 3 10's, 30's, 40's, 50's, 60's

In FIG. 4, first, FLAG [1] = false,
Initialization is performed with FLAG [2] = false and FLAG [3] = false, and the code indicating the input item number is set to I, and this is set to the numerical value "1" corresponding to input item 1 (S1),
It is determined whether I = 3 (S2). Since I is not 3, voice input is performed, a feature pattern is created by the feature pattern creation unit 2 from the input voice signal, and identification is performed by the group identification unit 3 configured by a neural network (S3). .

This identifies a group of input audio signals, that is, input item 1. Now, for example, if it is identified as input item i (generalized number of input item), I = i. Whether or not (S6), I = i
Therefore, the word identification is performed, the value RCG [1] indicating which word in the input item is recognized is j, and the flag FLAG [1] indicating that some input has already been made in the input item. Is performed (S9), and the process returns to step S2. Further, in the determination of step S6, I ≠
In the case of i, that is, when the input item identified by the group identification unit 3 configured by the neural network is, for example, “2” with respect to the precondition I = i, the flag FLAG [1] is true. ， False is judged (S7), FLA
When G [1] is true, I = I + 1 is set (S8),
The process proceeds to step S9, and if FLAG [1] is false, the process returns to step S2.

For example, in the case of I = 1, when the voice input is “male”, the input item identified by the group identification unit 3 composed of the neural network is “1”, that is, i = 1, In the judgment as to whether or not I = i, I = i, RCG [1] = male, FLAG [1]
= True is substituted, and the process returns to step S3. In this case, since the value of I remains 1, the change input of the input item 1 becomes possible.

Next, for example, when I = 1, that is, when there is a voice input of "Kinki" with "1" which is the input item regarding sex, the input item identified by the neural network is "2". Therefore, in step S7, it is determined whether the flag FLAG [1] indicating whether or not an input item 1 has already been input is true or false (S7). FL
If AG [1] is true, that is, if there is already an input regarding gender, the input item is changed to I = I + 1 = 2, and in step S9 FLAG [2] = true and RCG [2] =
Perform the process of "Kinki". On the other hand, if FLAG [1] is false, it is determined that a voice input error has occurred, and the process returns to step S3.
The input item I = 1 remains as it is, and the next voice input waiting state is entered. In the above-described embodiment, the case where the word identification units 4, 5, and 6 of the input items 1, 2, and 3 are provided is three, but the present invention is not limited to this, and the number may be increased or decreased as necessary. Needless to say.

[0026]

As described above, according to the first aspect of the present invention, the identification target word is divided into a plurality of groups in advance, and the group identification unit constituted by the neural network is made to learn this so that it can be identified. This makes it possible to group patterns that do not necessarily have similar physical characteristics and to group the patterns according to the intended use. Further, according to the second aspect of the invention, the control unit can sequentially operate different word identifying units based on the identification result of the group identifying unit configured by the neural network. No operation is required, and the operation is extremely easy.

[Brief description of drawings]

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

FIG. 2 is an explanatory diagram showing a relationship between a group identification unit configured by a neural network and groups.

FIG. 3 is an explanatory diagram showing the contents of each grouped input item.

FIG. 4 is a flowchart showing a processing process of the voice recognition device according to the present invention.

FIG. 5 is a block diagram showing a configuration of a conventional device.

FIG. 6 is an explanatory diagram showing the contents of grouped input items in the conventional apparatus.

[Explanation of symbols]

1 voice input unit 2 feature pattern creation unit 3 group identification unit composed of neural network 4 word identification unit for input item 5 word identification unit for input item 2 word identification unit for input item 3 recognition result control Department

Claims (2)

[Claims] 1. A voice recognition device in which voices corresponding to respective groups of predetermined words are sequentially input, and words corresponding to voice inputs in the group are recognized, the voice input unit and the input voices. A feature pattern creating unit that creates a feature pattern, a group identifying unit configured by a neural network that is learned based on the feature pattern so as to identify which of the groups the input voice belongs to, and the group A voice recognition device, comprising: a word identification unit that identifies a corresponding word from each of the groups identified by the identification unit based on the characteristic pattern. 2. The voice recognition device according to claim 1, further comprising a control unit that switches a group to be input next by voice based on the identification result of the group identification unit.