JPH08146996A - Speech recognition device - Google Patents

Abstract

PURPOSE: To improve the recognition precision by using, as an input pattern, what causes a misrecognition when it is inputted to a neural network after initial learning to be subjected to speech recognition, and making the neural network to perform additional learning by using unsupervised data. CONSTITUTION: A neural network arithmetic part 4 generates standard speech patterns for initial learning based upon a suitable speech section and standard speech patterns for initial learning based upon speech sections different from the suitable speech section by speeches to be recognized. The neural network initially learns the standard speech patterns for initial learning as input patterns and speech discrimination data representing the speeches corresponding to the respective input patterns as tutor data. When a standard speech pattern for additional learning is inputted to the neural network after the initial learning and speech recognition is performed, a misrecognized pattern is regarded as input data and the neural network performs the additional learning by using the unsupervised data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device for inputting data by voice, for example, a voice recognition device used for a recording device or the like in which reservation of a recorded program is made by voice input.

[0002]

2. Description of the Related Art FIG. 6 shows the configuration of a conventional voice recognition device.

The voice analysis unit 101 generates a voice power signal of an input voice and a voice spectrum for the input voice. The voice power signal of the input voice is the voice section detection unit 10
Sent to 2. The voice spectrum for the input voice is sent to the voice pattern creating unit 103.

The voice section detection unit 102 includes a voice detection unit 11
1 and a voice section cutout unit 112. As shown in FIG. 7, the voice detection unit 111 detects the voice portion in the voice power signal using the voice detection threshold value α.

As shown in FIG. 7, the voice section cutout unit 112 obtains a voice section L effective for voice recognition using a cutout threshold value β. The cut-out threshold value β is determined based on the noise power of a predetermined time before the voice portion detected by the voice detection unit 111.

The voice pattern creating section 103 creates a voice pattern based on the voice spectrum for the voice section L obtained by the voice section cutting section 112. The created voice pattern is input to the learned neural network 104.

The learning of the neural network 104 is performed as follows. First, a standard voice pattern for each recognition target voice is obtained by using voices collected in advance. Then, the neural network 104 is trained by using each standard voice pattern as an input pattern and voice identification data representing a voice corresponding to each input pattern as teacher data.

The learned neural network 104
By inputting the voice pattern to the input device, an output pattern corresponding to the input voice pattern is obtained. This output pattern is sent to the recognition result determination unit 105. The recognition result determination unit 105 recognizes the voice of the voice detection portion based on the sent output pattern and outputs the recognition result.

[0009]

In such a voice recognition device, since there is one cut-out threshold β for setting a voice section effective for voice recognition, noise is included in the voice section. There is a high possibility that erroneous recognition may occur, or a false ending may occur due to a word ending or the like having low voice power dropping from the voice section.

Therefore, the present applicant has developed the following voice recognition method. That is, as shown in FIG. 5, a plurality of voice sections L1, L2, L3, and L4 are set using a plurality of threshold values β1, β2, β3, and β4. A voice pattern is created for each of the voice sections L1 to L4. Each voice pattern is input to the neural network, and an output pattern is obtained for each voice pattern. And, based on these multiple output patterns obtained,
Recognize voice.

The voice identification data representing each recognition target voice is
It is assumed that it is composed of a number of data corresponding to each unit in the output layer of the neural network. And
It is assumed that only one of them is composed of data of "1" and the other is composed of "0", and the position of the data "1" is different for each voice identification data.

In such a voice recognition method, the recognition result of each voice section L1 to L2 in FIG. 5 may be as follows, for example. That is, the recognition result in the voice section L1 is "shichi", and the maximum output value (the maximum value of the outputs of the units in the output layer of the neural network) is 0.90. The recognition result in the voice section L2 is "ni", and the maximum output value is 0.85. The recognition result in the voice section L3 is "
The output maximum value is 0.91 in the case of "ni". The recognition result in the voice section L4 is "ni" and the maximum output value is 0.88.

In such a case, as the final recognition result, the recognition result "in" in the voice section L3 whose output maximum value is closest to "1" is selected as the recognition result of the input voice,
Where it should have been recognized as "shichi", it is incorrectly recognized as "ni".

It is an object of the present invention to provide a voice recognition device capable of improving recognition accuracy.

[0015]

A first voice recognition apparatus according to the present invention creates a voice pattern of a voice section based on a feature of the voice section and a voice section setting means for setting a voice section for an input voice. And a voice recognition means for recognizing the input voice based on the output of the neural network, and a suitable voice section for each recognition target voice. And a standard speech pattern for additional learning based on a speech section different from a suitable speech section are created, and the standard speech pattern for initial learning is used as an input pattern, and a speech corresponding to each input pattern is generated. The neural network is initially trained using the voice identification data that represents When the speech recognition is performed by inputting it to the neural network that has already been trained, the input pattern is the one in which misrecognition occurs, and the neural network is additionally learned using the anti-teaching data. It is characterized by being The features of the voice section include, for example, a voice spectrum.

The anti-teacher data is composed of a number of data in which each voice identification data corresponds to each unit of the output layer of the neural network, and only one of them is "1" and the others are "0". When the position of the data "1" is different for each voice identification data, it is composed of all the data "0".

Each voice identification data is composed of a number of data corresponding to each unit of the output layer of the neural network, only one of which is "0" and the other is all "1". When the position of "0" is different for each voice identification data, the counter teacher data is composed of all "1" data.

The second voice recognition apparatus according to the present invention creates a voice pattern for each voice section on the basis of the voice section setting means for setting a plurality of voice sections for the input voice and the characteristics of each voice section. And a voice recognition means for recognizing the input voice based on the output of the neural network for the voice pattern for each voice section. For each recognition target speech, a standard speech pattern for initial learning based on a suitable speech section and a standard speech pattern for additional learning based on a speech section different from the suitable speech section are created, and a standard speech pattern for initial learning is created. The voice pattern is used as an input pattern, and the voice identification data representing the voice corresponding to each input pattern is used as teacher data. Then, the neural network is initially learned, and among the standard speech patterns for additional learning, when it is input to the already learned neural network and speech recognition is performed, the one that causes misrecognition is taken as the input pattern. , The neural network is additionally learned by using the anti-teacher data. The features of the voice section include, for example, a voice spectrum.

[0019]

In the first voice recognition apparatus according to the present invention, the voice section is set for the input voice. A voice pattern of the voice section is created based on the characteristics of the voice section. The voice pattern is input to the neural network. Then, the input voice is recognized based on the output of the neural network.

In the second voice recognition device according to the present invention,
A plurality of voice sections are set for the input voice. A voice pattern for each voice section is created based on the characteristics of each voice section. The voice pattern for each voice section is input to the neural network. The input voice is recognized based on the output of the neural network for the voice pattern for each voice section.

Learning of the neural network of the first or second speech recognition apparatus according to the present invention is performed as follows.

That is, a standard speech pattern for initial learning based on a suitable speech section and a standard speech pattern for additional learning based on a speech section different from the preferred speech section are created for each speech to be recognized, and initial learning is performed. The neural network is initially learned using the standard voice pattern for input as an input pattern and the voice identification data representing the voice corresponding to each input pattern as the teacher data.

Further, among the standard voice patterns for additional learning, the one that is erroneously recognized when the voice is recognized by inputting it to the neural network which has been initially learned is used as the input pattern, and the anti-teaching data is set. The neural network is additionally trained using this.

[0024]

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

FIG. 1 shows the configuration of a voice recognition device.

The voice recognition device comprises a voice analysis unit 1, a voice section detection unit 2, a voice pattern creation unit 3, a neural network operation unit 4, a recognition result storage unit 5 and a recognition result determination unit 6. The voice section detection unit 2 is a voice detection unit 2
1. Voice segment cutout unit 22 and cutout position storage unit 23
It has.

FIG. 2 shows a neural network operation unit 4
3 shows an example of the structure of the neural network provided in the.

This neural network has an input layer 4
1, an intermediate layer 42 and an output layer 43. Input layer 41
Is composed of, for example, 128 (16 channel × 8 frame) input units. The intermediate layer 42 is composed of, for example, 50 intermediate units that are mutually coupled to the input units of the input layer 41. The output layer 43 is composed of, for example, 20 output units that are mutually coupled to the respective intermediate units of the intermediate layer 42.

Here, it is assumed that there are 20 speeches to be recognized. It is assumed that the voice identification data representing each recognition target voice is composed of 20 pieces of data corresponding to the output unit, only one of which is "1" and the other is all "0". The position of the data "1" is different for each voice identification data.

FIG. 3 shows a learning method of the neural network. A standard speech pattern for initial learning and a standard speech pattern for additional learning are created for each recognition target speech (step 1).

That is, for example, as shown in FIG. 4, a sound power signal for a predetermined sound, for example, a standard sound signal of "shichi" is generated. And the preferred threshold δ1
Is used to set the voice section R1. Further, another one or a plurality of threshold values δ2, δ3 ... δn (in this example, δ
2, δ3, δ4), the voice sections R2, R3 ... Rn
(R2, R3, R4 in this example) are set.

Then, a standard voice pattern is created for each voice section R1 to Rn. The standard voice pattern for the voice section R1 is the standard voice pattern for initial learning,
The standard voice pattern for the voice sections R2 to Rn is the standard voice pattern for additional learning. As each standard speech pattern, an average spectrum of each section obtained by dividing the corresponding speech section into eight equal parts is used. The voice spectrum of each section is composed of voice spectra for 16 predetermined frequency bands.

When the standard voice pattern for initial learning and the standard voice pattern for additional learning have been created in this way for all recognition target voices, initial learning is performed (step 2).

That is, the neural network is trained by the back propagation method using the initial learning standard voice pattern for each recognition target voice as the input pattern and the voice identification data representing the voice corresponding to each input pattern as the teacher data.

Next, a process of selecting an input pattern for additional learning is performed (step 3).

That is, the standard voice patterns for additional learning for the respective voices to be recognized are sequentially input to the neural network which has been initially learned, and the voice recognition result is obtained based on the output. Of the standard voice patterns for additional learning, the one in which erroneous recognition has occurred is selected as the input pattern for additional learning.

For example, the voice sections R2 and R3 shown in FIG.
When voice recognition is performed by sequentially inputting additional learning standard speech patterns for R4 and R4 to a neural network that has been initially learned, there is a place where each additional learning standard speech pattern should be recognized as "shichi". , "To"
It is assumed that it was mistakenly recognized as. In such a case, the standard voice pattern for additional learning for the voice sections R2, R3, and R4 is selected as the input pattern for additional learning.

Next, additional learning is performed (step 4).

That is, the initial learned standard speech pattern selected in step 3 as the additional learning input standard pattern and the initial learning standard speech pattern created in step 1 are used as the input patterns. Train additional networks. At this time, all 0 data is used as teacher data for each standard voice pattern for additional learning. Also, as the teacher data for the standard voice pattern for initial learning, voice identification data representing the voice corresponding to each standard voice pattern for initial learning is used.

Taking FIG. 4 as an example, the voice sections R2, R3,
The standard voice pattern for additional learning for R4 is used as an input pattern, and additional learning is performed using the teacher data of all 0s.

The operation of the voice recognition apparatus of FIG. 1 will be described.

The voice analysis unit 1 generates a voice power signal of the input voice and a voice spectrum for the input voice.
The voice power signal of the input voice is sent to the voice section detection unit 2. The voice spectrum for the input voice is sent to the voice pattern creating unit 3.

As shown in FIG. 5, the voice detecting section 21 detects the voice portion in the input voice power signal by using the voice detecting threshold value α.

As shown in FIG. 5, the voice section cutout unit 22 sets a plurality of voice sections using a plurality of cutout thresholds β1, β2, β3, β4. In this example, the first
To the fourth voice section L1, L2, L3, L4. Then, the set start points and end points of the respective voice sections L1 to L4 are stored in the cutout position storage unit 23 in association with the respective voice sections L1 to L4.

Threshold values β1, β2, β3, β for cutting out
4 is set as follows, for example. First, the minimum clipping threshold β1 is determined based on the noise power detected by the voice detection unit 21 a predetermined time before the start position of the voice portion. Then, a threshold β2 is obtained by adding a constant γ to the determined minimum clipping threshold β1, and a threshold β3 is obtained by adding a constant γ to the threshold β2. Then, the threshold value β4 is obtained by adding the constant γ to the threshold value β3.

The voice pattern creating section 3 creates a voice pattern for each of the voice sections L1 to L4 based on the voice spectrum for each of the voice sections L1 to L4 obtained by the voice section cutout section 22, and calculates the neural network. Input to the section 4.

That is, the voice pattern (P1) for the voice section L1 is created based on the start point and the end point of the first voice section L1 stored in the cut-out position storage section 23. As the voice pattern, the average spectrum of each of the eight voice segments is used. And the speech spectrum pattern of each section is
It is composed of speech spectra for 16 predetermined frequency bands. The created first voice pattern (P1) is input to the learned neural network.

In the learned neural network, the first
By inputting the voice pattern (P1) of
An output pattern corresponding to the voice pattern (P1) is obtained. Then, based on the obtained output pattern, the recognition result and the output maximum value (the maximum value of the 20 outputs) are stored in the recognition result storage unit 5 as the first recognition result.

Next, a voice pattern (P2) for the voice section L2 is created and created based on the start point and end point of the second voice section L2 stored in the cut-out position storage section 13. The second voice pattern (P2) is input to the learned neural network. As a result, an output pattern corresponding to the second voice pattern (P2) is obtained. The recognition result and the maximum output value are stored in the recognition result storage unit 5 as the second recognition result based on the obtained output pattern.

Next, a voice pattern (P3) for the voice section L3 is created based on the start point and end point of the third voice section L3, and is input to the learned neural network. As a result, an output pattern corresponding to the third voice pattern (P3) is obtained. Based on the obtained output pattern, the recognition result and the maximum output value are stored in the recognition result storage unit 5 as the third recognition result.

Next, a voice pattern (P4) for the voice section L4 is created based on the start point and the end point of the fourth voice section L4 and is input to the learned neural network. As a result, an output pattern corresponding to the fourth voice pattern (P4) is obtained. The recognition result and the maximum output value are stored in the recognition result storage unit 5 as the fourth recognition result based on the obtained output pattern.

In this way, when the first to fourth recognition results for the first to fourth voice patterns (P1 to P4) are obtained, the recognition result determination section 6 is stored in the output pattern storage section 5. Among the first to fourth recognition results, the voice recognition result whose output maximum value is closest to "1" is selected and output as the voice recognition result of the detected voice portion. That is, the input voice is recognized based on the output pattern having the highest degree of similarity to the voice identification data (teacher data).

In the above embodiment, a plurality of voice sections L1 to L4 obtained by a plurality of clipping thresholds β1 to β4 are set for one voice detection portion. Then, since the voice of the voice detection portion is recognized based on the voice pattern for each voice section, erroneous recognition occurs due to noise being included in the voice section, and the ending of a voice with a small voice power is voiced. It is possible to prevent erroneous recognition from occurring due to dropping from the section.
As a result, the voice recognition accuracy is improved.

In the above embodiment, a standard voice pattern is created for each voice to be recognized with a plurality of threshold values, and there is a possibility that the standard voice pattern may be erroneously recognized as another voice. For some, the initial learned neural network is additionally learned by using them as input patterns and using teacher data of all zeros. Therefore, a highly sensitive output pattern can be obtained from the neural network only when the voice pattern is close to the standard voice pattern for initial learning. As a result, the recognition accuracy is improved.

In the above embodiment, a plurality of voice sections are set for the input voice by a plurality of thresholds β1 to β4, but one threshold is set for the input voice by one threshold.
It is also possible to set only the voice section.

In the above embodiment, the voice section is set based on the voice power of the input voice and the cut-out threshold value. However, the voice section determination parameters other than the voice power and the parameters are set according to the parameters. The voice section may be set based on the threshold value. Besides the voice power, the parameters for the voice section determination include a power slope, a wide range power, a low range power, and the like.

Further, as the characteristics of the voice section for creating the voice pattern for each voice section, the inclination of the voice spectrum, the voice power, etc. may be used in addition to the voice spectrum.

[0058]

According to the present invention, the recognition accuracy can be improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a voice recognition device.

FIG. 2 is a schematic diagram showing a structure of a neural network provided in a neural network operation unit of FIG.

FIG. 3 is a flowchart for explaining a learning method of a neural network.

FIG. 4 is a time chart for explaining a method of creating a standard speech pattern for initial learning and a standard speech pattern for additional learning of a neural network.

5 is a time chart showing that a plurality of voice sections are set on the basis of a plurality of clipping thresholds in the voice recognition device of FIG. 1. FIG.

FIG. 6 is a block diagram showing a configuration of a conventional voice recognition device.

7 is a time chart showing that one voice section is set based on one clipping threshold value in the voice recognition device of FIG. 6. FIG.

[Explanation of symbols]

 1 voice analysis unit 2 voice section detection unit 3 voice pattern creation unit 4 neural network operation unit 5 recognition result storage unit 6 recognition result determination unit 21 voice detection unit 22 voice section cutout unit 23 cutout position storage unit

Claims (3)

[Claims] 1. A voice section setting means for setting a voice section for an input voice, a voice pattern creating means for producing a voice pattern of a voice section based on the characteristics of the voice section, and a neural network to which the voice pattern is input. And a voice recognition means for recognizing an input voice based on the output of the neural network. For each recognition target voice, a standard voice pattern for initial learning based on a suitable voice segment, and a suitable voice segment Is created as a standard speech pattern for additional learning based on different speech sections, and the standard speech pattern for initial learning is used as an input pattern,
The neural network is initially learned by using the voice identification data representing the voice corresponding to each input pattern as the teacher data, and it is input to the already learned neural network among the additional learning standard voice patterns for voice recognition. A speech recognition device in which a neural network is additionally learned by using anti-teaching data by using an input pattern that is erroneously recognized. 2. A voice section setting means for setting a plurality of voice sections for an input voice, based on characteristics of each voice section,
Based on the output of the neural network having a voice pattern creating means for creating a voice pattern for each voice section and a neural network to which the voice pattern for each voice section is respectively input, and for the voice pattern for each voice section A voice recognition means for recognizing an input voice is provided, and for each recognition target voice, a standard voice pattern for initial learning based on a suitable voice section and a standard voice for additional learning based on a voice section different from the suitable voice section. The pattern and is created, the standard voice pattern for initial learning as the input pattern,
The neural network is initially learned by using the voice identification data representing the voice corresponding to each input pattern as the teacher data, and is input to the initially learned neural network of the additional learning standard voice patterns for voice recognition. A speech recognition device in which a neural network is additionally learned by using anti-teaching data by using an input pattern that is erroneously recognized. 3. The voice recognition device according to claim 1, wherein the feature of the voice section is a voice spectrum.