JP2022069766A - Bone conduction microphone, method for enhancing voice of bone conduction microphone, and voice enhancement program - Google Patents

Abstract

To provide a bone conduction microphone, a voice enhancement method for a bone conduction microphone, and a voice enhancement program that can obtain clearer voice with a simple process.SOLUTION: A bone conduction microphone comprises: a vibration detecting element that contacts a human body and detects vocal fold vibration; a transducer 23 that converts data of vocal fold vibration detected by the vibration detecting element into a first utterance code and a first sound quality code; a storage device 21 that stores the second sound quality code when a second transducer that converts voice data of a human voice caused by air vibration due to the vocal fold vibration into a second utterance code and a second sound quality code converts the voice data into the second utterance code and the second sound quality code; and a generator 24 that generates voice-enhanced data in which the voice is enhanced, based on the first utterance code converted by the converter 23 and the second sound quality code stored by the storage device 21.SELECTED DRAWING: Figure 2

Description

The present invention relates to a bone conduction microphone, a speech enhancement method for the bone conduction microphone, and a speech enhancement program.

A bone conduction microphone is a device that detects vocal cord vibration that conducts bone. However, since vocal cord vibration attenuates high-frequency components when it conducts bone, the sound detected by the bone conduction microphone is usually not clear. Therefore, in order to obtain a clearer sound, a speech enhancement device that emphasizes the speech component in the speech data detected by the bone conduction microphone has been developed.

For example, Patent Document 1 describes a discriminating means for discriminating between voiced sound and unvoiced sound by analyzing voice data detected by a bone conduction microphone, and first air conduction voice data by correcting voice data discriminated as voiced sound. The first correction means for generating the second air conduction voice data, the second correction means for correcting the voice data determined to be an asexual sound, and the generated first air conduction voice data and the second air. A sound enhancement device including an output generation means for generating output data by combining guide sound data is disclosed.

Japanese Unexamined Patent Publication No. 2012-208177

In the speech enhancement device described in Patent Document 1, since the speech of the bone conduction microphone is analyzed and divided into voiced sound and unvoiced sound, the processing of voice data is complicated.

The present invention has been made to solve the above problems, and to provide a bone conduction microphone, a speech enhancement method for a bone conduction microphone, and a speech enhancement program capable of obtaining clearer speech by simple processing. The purpose.

In order to achieve the above object, the bone conduction microphone according to the first aspect of the present invention is
A vibration detection element that detects vocal cord vibration in contact with the human body,
A first converter that converts the vocal cord vibration data detected by the vibration detection element into a first utterance code and a first sound quality code.
The second converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second utterance code and the second sound quality code, converts the voice data into the second utterance code and the second sound quality. A storage device that stores the second sound quality code when converted to a code, and
A generator that generates speech enhancement data in which speech is emphasized based on the first utterance code converted by the first converter and the second sound quality code stored in the storage device.
To prepare for.

Even if the first transducer is a first encoder that forms a first autoencoder when combined with a first decoder that restores vocal cord vibration data from the first utterance code and the first sound quality code. good.

The second converter may be a second encoder that forms a second autoencoder when combined with the second utterance code and the second decoder that restores the voice data from the second sound quality code.

The speech enhancement method of the bone conduction microphone according to the second aspect of the present invention is
A conversion step for converting the vocal cord vibration data acquired from the vibration detection element of the bone conduction microphone that comes into contact with the human body to detect vocal cord vibration into the first speech code and the first sound quality code.
The converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second speech code and the second sound quality code converts the voice data into the second speech code and the second sound quality code. The voice data is read from the storage device that stores the second sound quality code at the time of conversion, and the voice data is based on the read second sound quality code and the first speech code converted in the conversion step. A generation step to generate voice-enhanced data that emphasizes the voice of
To prepare for.

The speech enhancement program according to the third aspect of the present invention is
A vibration detection element that detects vocal cord vibration in contact with the human body,
The converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second utterance code and the second sound quality code converts the voice data into the second utterance code and the second sound quality code. A storage device that stores the second sound quality code at the time of conversion,
A speech enhancement program for bone conduction microphones
On the computer
A conversion step for converting the vocal cord vibration data acquired from the vibration detection element into a first utterance code and a first sound quality code, and
A generation step of reading the second sound quality code from the storage device and generating voice-enhanced data based on the read second sound quality code and the first utterance code converted in the conversion step.
Is to execute.

According to the configuration of the present invention, the first converter converts the voice band vibration data detected by the vibration detection element into the first speech code and the first sound quality code, and the generator stores the first speech code and the storage. The speech enhancement data that emphasizes the voice is generated based on the second sound quality code when the second converter converts the speech data of the human voice stored in the unit. As a result, clear voice data can be obtained from the vocal cord vibration data. In addition, it is not necessary to perform complicated preprocessing on the vocal cord vibration data, and the processing is simple.

It is a component block diagram of the bone conduction microphone which concerns on embodiment of this invention. It is a block diagram of the speech enhancement device provided in the bone conduction microphone. It is a data structure diagram of the sound quality table stored in the storage device provided in the speech enhancement device. It is a block diagram of the learning device which trains an encoder and a decoder included in a speech enhancement device. It is a block diagram of the learning model provided in the learning device. It is a flowchart of the learning process carried out by a learning apparatus.

Hereinafter, the bone conduction microphone, the speech enhancement method of the bone conduction microphone, and the speech enhancement program according to the embodiment of the present invention will be described in detail with reference to the drawings. In the figure, the same or equivalent parts are designated by the same reference numerals.

The bone conduction microphone according to the embodiment uses a trained neural network to generate speech enhancement data from vocal cord vibration data detected by a vibration detection element in order to obtain clear speech. First, the configuration of the bone conduction microphone will be described with reference to FIGS. 1 to 3.

FIG. 1 is a component configuration diagram of a bone conduction microphone 100 according to an embodiment of the present invention. FIG. 2 is a block diagram of the speech enhancement device 20 included in the bone conduction microphone 100. FIG. 3 is a data configuration diagram of the sound quality table 212 stored in the storage device 21 included in the speech enhancement device 20. Note that FIG. 1 also shows a speaker 200 output by the bone conduction microphone 100 for easy understanding.

As shown in FIG. 1, the bone conduction microphone 100 includes a vibration detection element 10 that detects vocal cord vibration, a speech enhancement device 20 that generates speech enhancement data based on vocal cord vibration data detected by the vibration detection element 10, and a speech enhancement device 20. To be equipped.

Although not shown, the vibration detection element 10 includes a case and a piezoelectric element housed in the case. In the vibration detection element 10, when the vibration propagates to the case, the piezoelectric element bends due to the vibration, and a potential is generated in the piezoelectric element. The vibration detection element 10 detects vibration from its potential.

The case of the vibration detection element 10 has a shape capable of contacting a part of the human body, for example, the skin such as the crown, the temporal region, the pharynx, and the nasal cavity. As a result, the vibration detection element 10 is mounted in a state where the case is in contact with a part of the human body. In this state, the vibration detection element 10 detects vocal cord vibration propagating in a part of the human body such as the skull. The vibration detection element 10 transmits the detected vocal cord vibration data to the speech enhancement device 20.

The speech enhancement device 20 receives the data of vocal cord vibration. The speech enhancement device 20 includes a storage device 21 and a controller 22 in order to process the data of the vocal cord vibration.

The storage device 21 has an EEPROM (Electrical Erasable Programmable Read-Only Memory), a flash memory, or the like. Then, the storage device 21 stores the speech enhancement program 211 that generates speech enhancement data from the vocal cord vibration data. Further, the storage device 21 stores a model database 213 that stores the parameters of the speech enhancement program 211.

The controller 22 includes a microcomputer including a CPU (Central Processing Unit) for performing arithmetic processing, and a memory including a ROM (Read Only Memory) and a RAM (Random Access Memory). The CPU performs various processes by reading the program stored in the ROM or the storage device 21 into the RAM and executing the program. For example, in the controller 22, the CPU executes the speech enhancement program 211. Then, the model database 213 is read out. As a result, speech enhancement processing is performed. In order to perform this speech enhancement process, the controller 22 includes a processing block of a converter 23 and a generator 24 configured as software, as shown in FIG.

The converter 23 receives the data of vocal cord vibration detected by the vibration detection element 10 shown in FIG. As shown in FIG. 2, the converter 23 includes an encoder E1 and converts vocal cord vibration data into an utterance code and a microphone sound quality code (hereinafter referred to as utterance code C1 and sound quality code C2) by the encoder E1. The converter 23 transmits the converted utterance code C1 and the sound quality code C2 to the generator 24.

On the other hand, the sound quality table 212 is stored in the storage device 21. In the sound quality table 212, the voice data recorded by using a normal microphone is stored in the encoder E2, which will be described later, separately from the encoder E1, and is referred to as a speech code and a microphone sound quality code (hereinafter referred to as speech code C3 and sound quality code C4). ), The sound quality code C4 is associated with the sound quality code C2 of the encoder E1 as shown in FIG.

As shown in FIG. 2, the generator 24 receives the utterance code C1 and the sound quality code C2 from the above-mentioned converter 23. Further, the generator 24 reads the sound quality code C4 corresponding to the received sound quality code C2 from the sound quality table 212 of the storage device 21. The generator 24 includes a decoder D4, and speech enhancement data that emphasizes the voice corresponding to the vocal cord vibration detected by the vibration detection element 10 based on the received utterance code C1 and the sound quality code C4 read by the decoder D4. To generate. Specifically, the decoder D4 generates speech enhancement data by decoding the utterance code C1 and the sound quality code C4.

The generator 24 outputs the generated speech enhancement data to an external device. For example, the generator 24 outputs speech enhancement data to the speaker 200 illustrated in FIG. Speech is emphasized in the speech enhancement data. Therefore, the sound acquired by the bone conduction microphone 100 is clear and easy to hear.

The encoder E1 of the converter 23 and the decoder D4 of the generator 24 described above perform conversion processing and generation processing by using the trained neural network. Subsequently, the learning device 300 for learning the neural network of the encoder E1 and the decoder D4 will be described with reference to FIGS. 4 and 5.

FIG. 4 is a block diagram of a learning device 300 for learning the encoder E1 and the decoder D4 included in the speech enhancement device 20. FIG. 5 is a block diagram of the learning model 330 included in the learning device 300.

In the learning device 300, a CPU (not shown) reads the learning program 311 stored in the storage device 310 shown in FIG. 4 into the RAM and executes it. As a result, the learning device 300 performs the learning process. As a result, the learning device 300 includes a learning unit 320 and a learning model 330 configured as software.

The learning data 312 is stored in the storage device 310. The learning unit 320 reads the learning data 312 from the storage device 310, and inputs the read learning data 312 into the learning model 330.

As shown in FIG. 5, the learning model 330 is a model in which the encoders E1 and E2 and the decoders D1-D4 are combined. In the learning model 330, the encoder E1 and the decoder D1 form a set of networks. Further, the encoder E2 and the decoder D2 form another set of networks. Further, the decoders D3 and D4 are connected to the encoder E1 and the encoder E2, respectively, and each forms a different network.

Although not shown, the encoder E1 and the decoder D1 are constructed by a neural network model having an input layer, a hidden layer, and an output layer. The input layer and the output layer of the neural network model have the same number of dimensions, and the hidden layer has a smaller number of dimensions than the input layer and the output layer. The encoder E1 is constructed by the portion from the input layer to the hidden layer of the neural network model, and the decoder D1 is constructed by the portion from the hidden layer to the output layer of the neural network model.

As shown in FIG. 4, the learning data 312 described above includes vocal cord vibration data A pre-recorded using the vibration detection element 10 and voice data B pre-recorded using a normal microphone. , Is stored.

Here, the voice data B is data obtained by recording the voice emitted by the vocal cord vibration when the vocal cord vibration data A is recorded. That is, the voice data B records the voice corresponding to the vocal cord vibration data A. In addition, in this specification, a normal microphone means a microphone which converts the sound generated by the vibration of air by vocal cord vibration into an electric signal, and is also referred to as an air conduction sound microphone.

The learning unit 320 inputs the vocal cord vibration data A of the learning data 312 to the encoder E1 as shown in FIG. Then, the vocal cord vibration data A and the output of the decoder D1 are compared to adjust the weights between the nodes in the neural network model. As a result, the learning unit 320 reduces the error between the output of the vocal cord vibration data A and the decoder D1. As a result, the encoder E1 and the decoder D1 learn the autoencoder. That is, it learns self-coding.

Further, the encoder E2 and the decoder D2 are constructed by a neural network model different from the neural network model described in the encoder E1 and the decoder D1. Note that this other neural network model has the same layer structure as the neural network model described in the encoder E1 and the decoder D1.

The learning unit 320 inputs the voice data B of the learning data 312 to the encoder E2. Then, the learning unit 320 compares the output of the voice data B and the output of the decoder D2, and adjusts the weight between the nodes in the other neural network model. As a result, the error between the output of the audio data B and the output of the decoder D2 is reduced. As a result, the encoder E2 and the decoder D2 learn self-coding.

When the encoder E1 and the decoder D1 and the encoder E2 and the decoder D2 each learn self-coding, the encoder E1 outputs the utterance code C1 and the sound quality code C2 in which the vocal cord vibration data A is encoded. Further, the encoder E2 outputs the utterance code C3 and the sound quality code C4 in which the voice data B is encoded.

On the other hand, the decoder D3 is constructed by a neural network unit having the same layer structure as the decoders D1 and D2. The learning unit 320 inputs the sound quality code C2 output by the encoder E1 and the utterance code C3 output by the encoder E2 to the decoder D3. Here, the sound quality code C2 is a microphone sound quality code of the vibration detection element 10 that records the vocal cord vibration data A. The learning unit 320 compares the output of the vocal cord vibration data A with the output of the decoder D3 in order to obtain an output corresponding to the code of the microphone sound quality. The learning unit 320 adjusts the weights between the nodes of the neural network unit based on the comparison result to reduce the error between the outputs of the vocal cord vibration data A and the decoder D3. As a result, the learning unit 320 trains the decoder D3 in a state of outputting the vocal cord vibration data A.

The decoder D4 is constructed by another neural network unit having the same layer structure as the neural network unit of the decoder D3. The learning unit 320 inputs the utterance code C1 output by the encoder E1 and the sound quality code C4 output by the encoder E2 to the decoder D4. The sound quality code C4 is a microphone sound quality code of a normal microphone that records voice data B. In order to obtain an output corresponding to the code of the microphone sound quality, the learning unit 320 compares the output of the voice data B and the output of the decoder D4 and adjusts the weight between the nodes of the neural network unit. As a result, the learning unit 320 reduces the error between the audio data B and the output of the decoder D4. As a result, the learning unit 320 trains the decoder D4 in a state of outputting the voice data B.

When the learning unit 320 learns the encoders E1 and E2 and the decoders D1-D4, that is, when the learning model 330 is trained, the learning unit 320 displays parameters such as the weighting coefficients of the encoders E1 and the decoder D4 of the learned learning model 330. It is stored in the storage device 21 as the model database 213 shown in 2. As a result, the learning unit 320 supplies the speech enhancement device 20 with the database necessary for the operation of the converter 23 and the generator 24.

Further, the learning unit 320 inputs the learning data 312 again into the learned learning model 330. The learning unit 320 creates the sound quality table 212 shown in FIG. 3 by using the sound quality code C2 output by the encoder E1 and the sound quality code C4 output by the encoder E2 at that time. The learning unit 320 stores the created sound quality table 212 in the storage device 21. As a result, the learning unit 320 supplies the speech enhancement device 20 with the table necessary for the operation of the generator 24. As a result, as described above, the speech enhancement device 20 generates speech enhancement data, and the vibration detected by the vibration detection element 10 is converted into a clear and easy-to-hear voice.

Next, the learning method of the learning device 300 will be described in more detail with reference to FIG. Although not shown in the following description, it is assumed that the learning device 300 is composed of a personal computer or a server (hereinafter, referred to as a server or the like). Then, it is assumed that the learning program 311 and the learning data 312 are stored in the storage device provided in the server or the like, and the icon of the learning program 311 is displayed on the display device. Further, it is assumed that these servers and the like are connected to the controller included in the speech enhancement device 20 of the bone conduction microphone 100 via the Internet.

FIG. 6 is a flowchart of the learning process performed by the learning device 300.

First, the user of the learning device 300 presses the above icon to activate the learning program 311. As a result, the learning program is executed by the CPU of the server or the personal computer, and the flow of the learning process is started.

When the flow of the learning process is started, the learning unit 320 first reads the learning data 312 from the storage device 310. As a result, the learning data 312 is acquired (step S1).

It is desirable that the learning data 312 stores vocal cord vibration data A and voice data B when a person makes various pronunciations in order to increase the types of voices that can be emphasized by the voice enhancement device 20. For example, it is desirable that the learning data 312 stores vocal cord vibration data A and voice data B for each character when reading those characters for most characters in a specific language.

Subsequently, the learning unit 320 trains the neural network of the encoder E1 and the decoder D1 of the learning model 330 and the neural network of the encoder E2 and the decoder D2 using the acquired learning data 312 (step S2).

Specifically, the vocal cord vibration data A of the learning data 312 is input to the encoder E1, and the voice data B of the learning data 312 is input to the encoder E2. When the output of the decoder D1 is A ^* and the output of the decoder D1 is B ^* , the weights between the nodes in the network until the cost function L _all represented by the formula 1-formula 3 converges within a certain value. To adjust. As a result, the network of the encoder E1 and the decoder D1 and the network of the encoder E2 and the decoder D2 are learned.

Next, the learning unit 320 trains the entire learning model 330 using the learning data 312 (step S3).

Specifically, as in step S2, the vocal cord vibration data A of the learning data 312 is input to the encoder E1, and the voice data B of the learning data 312 is input to the encoder E2. Then, when the output of the decoder D3 is A ^** and the output of the decoder D4 is B ^** , the nodes in the network until the cost function L _all represented by the formula 4-formula 6 converges within a certain value. Adjust the weight between. As a result, the entire learning model 330 including the decoders D3 and D4 is trained.

When the learning of the entire learning model 330 is completed, the learning unit 320 stores the parameters of the learned learning model 330 in the storage device 21 (step S4). Specifically, parameters such as the number of network layers of the encoder E1 and the decoder D4, the number of nodes, and the weighting coefficient between the nodes are stored in the model database 213 of the storage device 21.

Further, the learning unit 320 creates a sound quality table 212 using the learned learning model 330, and stores the sound quality table 212 in the storage device 21 (step S5).

Specifically, the learning unit 320 inputs the learning data 312 into the learned learning model 330, and from the output data of the encoders E1 and E2 at that time, the sound quality code C2 which is the output of the encoder E1 is input to the encoder E2. The output sound quality code C4 is associated with it. At this time, for example, when vocal cord vibration data A and voice data B are stored in the learning data 312 for most of the characters in a specific language, the sound quality code C2 is associated with the sound quality code C4 for each of these characters. As a result, the learning unit 320 creates the sound quality table 212. Then, the created sound quality table 212 is stored in the storage device 21.

By the above steps, the learning of the learning device 300 is completed.

After the learning of the learning device 300 is completed, the power button (not shown) of the bone conduction microphone 100 is pressed to activate the bone conduction microphone 100, and the controller 22 stores the model database 213 stored in the storage device 21 in step S4. Based on the read and read model database 213, a neural network model of the encoder E1 and the decoder D4 is constructed. As a result, the learning of the learning device 300 is reflected in the operation of the bone conduction microphone 100.

Subsequently, when the vibration detection element 10 of the bone conduction microphone 100 detects the voice band vibration, the controller 22 acquires the voice band vibration data from the vibration detection element 10, and the acquired voice band vibration data is a neural structure constructed by the model database 213. It is converted into the speech code C1 and the sound quality code C2 by the encoder E1 of the network model (this step is also called a conversion step).

The controller 22 reads the sound quality code C4 corresponding to the converted sound quality code C2 from the sound quality table 212 stored in the storage device 21 in step S5, and decodes the utterance code C1 converted by the encoder E1 and the read sound quality code C4. By inputting to D4, the utterance code C1 and the sound quality code C4 are decoded. As a result, the controller 22 generates speech enhancement data (this step is also referred to as a generation step). As a result, in the bone conduction microphone 100, the voice is emphasized and it is easy to hear.

The above-mentioned encoders E1 and E2 are examples of the first converter or the first encoder, the second converter or the second encoder as defined in the present specification and the claims. The decoders D1 and D2 are examples of the first decoder and the second decoder as defined in the present specification and claims. Further, the autoencoder composed of the encoder E1 and the decoder D1 and the autoencoder composed of the encoder E2 and the decoder D2 are examples of the first autoencoder and the second autoencoder as defined in the present specification and claims. Is. Further, the utterance code C1, the sound quality code C2, the utterance code C3 and the sound quality code C4 converted by the encoders E1 and E2 are the first utterance code, the first sound quality code, and the second sound quality code as defined in the present specification and claims. It is an example of an utterance code and a second sound quality code.

Further, in the above embodiment, the bone conduction microphone 100 is connected to the learning unit 320 and the normal microphone is not connected, but the learning unit 320 is connected to the normal microphone in addition to the bone conduction microphone 100. It may be connected. In this case, when the user reads the text data aloud, the vibration detection element 10 of the bone conduction microphone 100 detects the vocal cord vibration of the user at that time, and the normal microphone detects the voice at that time. good. Then, the learning unit 320 may use the detected vocal cord vibration and voice data as the learning data 312 in step S1. In this case, the learning unit 320 may store the detected vocal cord vibration and voice data in the storage device 310 as learning data 312.

As described above, in the bone conduction microphone 100 according to the embodiment, the encoder E1 included in the converter 23 converts the voice band vibration data detected by the vibration detection element 10 into the speech code C1 and the sound quality code C2, and the generator. The decoder D4 included in the 24 is stored in the speech code C1 converted by the encoder E1 and the sound quality code C4 corresponding to the sound quality code C2 stored in the sound quality table 212 of the storage device 21 and converted by the encoder E1. Generate speech enhancement data. Therefore, with the bone conduction microphone 100, it is possible to obtain clear and easy-to-hear voice. In addition, it is not necessary to perform complicated preprocessing on the vocal cord vibration data, and the processing is simple.

Further, the encoder E1 included in the converter 23 and the decoder D4 included in the generator 24 are learned by the learning device 300. Therefore, by training the encoder E1 and the decoder D4 using the learning data 312 that stores the vocal cord vibration data and voice data of the user of the bone conduction microphone 100, the voice enhancement data corresponding to the user's vocal cord vibration and voice can be obtained. Can be generated.

Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments. For example, in the embodiment, the vibration detection element 10 includes a piezoelectric element, but the present invention is not limited thereto. In the present invention, it is sufficient that the vibration detection element 10 can detect the vibration of the vocal cords, and the element is arbitrary as long as it can detect the vibration of the vocal cords. For example, instead of the piezoelectric element, an electromagnetic element or an electrostatic element may be used.

Further, in the above embodiment, the speaker 200 is exemplified as an external device to which the bone conduction microphone 100 is connected, but the present invention is not limited thereto. In the present invention, the bone conduction microphone 100 may generate speech enhancement data in which speech is emphasized, and the connection destination thereof is not limited. For example, the bone conduction microphone 100 may be connected to a controller of a cabin such as a crane device or an aerial work platform. Then, it may be connected to the speaker 200 arranged in the cabin via the controller. It may also be connected to a bone conduction earphone. With such a form, it is possible to make it easier to hear the worker's voice even in an environment where a loud work sound is generated and the worker's voice is difficult to hear.

In the above embodiment, the learning device 300 is a device different from the bone conduction microphone 100. However, the present invention is not limited to this. In the present invention, the bone conduction microphone 100 may include the learning device 300. For example, the controller 22 of the speech enhancement device 20 may include a learning device 300, that is, a learning unit 320 and a learning model 330. In this case, it is preferable that the controller 22 is connected to a normal microphone. Then, in the speech enhancement device 20, the learning mode and the operation mode can be switched, and the learning model 330 is based on the vocal cord voice data detected by the vibration detection element 10 and the voice data detected by the normal microphone in the learning mode. You should learn. With such a form, the bone conduction microphone 100 can be adjusted according to the vocal cord vibration and voice of the user.

10 ... Vibration detection element, 20 ... Speech enhancement device, 21 ... Storage device, 22 ... Controller, 23 ... Converter, 24 ... Generator, 100 ... Bone conduction microphone, 200 ... Speaker, 211 ... Speech enhancement program, 212 ... Sound quality Table, 213 ... model database, 300 ... learning device, 310 ... storage device, 311 ... learning program, 312 ... learning data, 320 ... learning unit, 330 ... learning model, C1, C3 ... speech code, C2, C4 ... sound quality code , D1-D4 ... Decoder, E1, E2 ... Encoder

Claims (5)

A vibration detection element that detects vocal cord vibration in contact with the human body,
A first converter that converts the vocal cord vibration data detected by the vibration detection element into a first utterance code and a first sound quality code.
The second converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second utterance code and the second sound quality code, converts the voice data into the second utterance code and the second sound quality. A storage device that stores the second sound quality code when converted to a code, and
A generator that generates speech enhancement data in which speech is emphasized based on the first utterance code converted by the first converter and the second sound quality code stored in the storage device.
Bone conduction microphone with. The first transducer is a first encoder that forms a first autoencoder when combined with a first utterance code and a first decoder that restores vocal cord vibration data from the first sound quality code.
The bone conduction microphone according to claim 1. The second converter is a second encoder that forms a second autoencoder when combined with the second utterance code and the second decoder that restores the voice data from the second sound quality code.
The bone conduction microphone according to claim 1 or 2. A conversion step for converting the vocal cord vibration data acquired from the vibration detection element of the bone conduction microphone that comes into contact with the human body to detect vocal cord vibration into the first speech code and the first sound quality code.
The converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second speech code and the second sound quality code converts the voice data into the second speech code and the second sound quality code. The voice data is read from the storage device that stores the second sound quality code at the time of conversion, and the voice data is based on the read second sound quality code and the first speech code converted in the conversion step. A generation step to generate voice-enhanced data that emphasizes the voice of
Speech enhancement method for bone conduction microphones. A vibration detection element that detects vocal cord vibration in contact with the human body,
The converter that converts the voice data of the human voice generated by the vibration of the air due to the vocal band vibration into the second utterance code and the second sound quality code converts the voice data into the second utterance code and the second sound quality code. A storage device that stores the second sound quality code at the time of conversion,
A speech enhancement program for bone conduction microphones
On the computer
A conversion step for converting the vocal cord vibration data acquired from the vibration detection element into a first utterance code and a first sound quality code, and
A generation step of reading the second sound quality code from the storage device and generating voice-enhanced data based on the read second sound quality code and the first utterance code converted in the conversion step.
A speech enhancement program to execute.

Images