KR102213177B1 - Apparatus and method for recognizing speech in robot - Google Patents

Abstract

An apparatus and method for speech recognition in a robot are presented. According to an exemplary embodiment, a method for recognizing a voice in a robot includes: extracting input feature vectors from voice data including motion information of the robot and noise input through a microphone; Concatenating each of the extracted input feature vectors; And learning the combined input feature vector through a feature mapping model and an acoustic model, and learning the combined input feature vector through a feature mapping model and an acoustic model in a noise environment generated during operation of the robot. Voice recognition can be performed.

Description

Voice recognition device and method in robot {APPARATUS AND METHOD FOR RECOGNIZING SPEECH IN ROBOT}

The following embodiments relate to a speech recognition apparatus and method in a robot, and more particularly, to a speech recognition apparatus and method capable of robust speech recognition even in a noisy environment by using the motor operation state inside the robot as auxiliary information. .

With the recent growth of the robot industry, voice recognition technology is being used as the fastest and most powerful means of enabling communication between robots and humans. The speech recognition technology uses a method of pre-learning the probability of each phoneme from previously collected speech data, and then determining which phoneme the input speech data is closest to, and then estimating a phoneme sequence therefrom.

At this time, the probability model for each phoneme used is called an acoustic model, and the acoustic model is one of the important factors that influence the performance of speech recognition technology. The noise (noise) robustness of the acoustic model, an important factor influencing the performance of the speech recognition system, is an important issue, because the performance of the speech recognition system is seriously degraded due to the difference between the training environment and the actual environment.

Conventionally, as the collection of big data becomes easier, deep learning is used as a preprocessing for robust voice recognition. A deep neural network (DNN) model is trained to estimate and obtain a clean speech signal from an input mixed with noise, and map it from a distorted signal to a clean signal.

However, the deep neural network (DNN)-based feature enhancement models have been developed only for robust voice recognition in external environments by training the model using only voice information, which is information obtained from a microphone input. However, in an environment in which a robot actually recognizes voice, noise generated when the robot moves the body according to a user's command, or noise from the motor and fan is present, which degrades the performance of voice recognition.

Korea Publication No. 10-2016-0032536 relates to a deep neural network-based speech recognition device in which such a signal processing algorithm is integrated and a learning method thereof. It describes the related technology.

Korean Publication No. 10-2016-0032536

The embodiments describe a voice recognition apparatus and method in a robot, and more specifically, a voice recognition technology capable of robust voice recognition even in a noisy environment is provided by using the motor operation state inside the robot as auxiliary information.

Embodiments provide an apparatus and method for speech recognition in a robot that minimizes discrepancy between a training environment and an actual environment by using motion state information that can be obtained by the robot itself as auxiliary information when learning a voice recognition model.

According to an exemplary embodiment, a method for recognizing a voice in a robot includes: extracting input feature vectors from voice data including motion information of the robot and noise input through a microphone; Concatenating each of the extracted input feature vectors; And learning the combined input feature vector through a feature mapping model and an acoustic model, and learning the combined input feature vector through a feature mapping model and an acoustic model in a noise environment generated during operation of the robot. Voice recognition can be performed.

Extracting an input feature vector from voice data including motion information of the robot and noise input through a microphone may include: extracting a feature vector from voice data including noise input through the microphone; And receiving the motion information of the robot and performing one-hot encoding.

The step of combining the extracted respective input feature vectors may include: a sequence of voice feature vectors extracted from the noise-containing voice data and a sequence of motion information vectors one-hot encoded from motion information of the robot into one input vector sequence. Can be combined.

The training of the combined input feature vector through a feature mapping model and an acoustic model includes the combined input feature vector being a deep neural network (DNN) estimating a feature vector of speech that does not contain noise. Learning through a feature mapping model; And extracting an improved input feature vector by combining a result learned through the feature mapping model with a one-hot-encoded motion information vector sequence from motion information of the robot.

The training of the combined input feature vector through a feature mapping model and an acoustic model may further include training the extracted enhanced input feature vector through the acoustic model, which is a deep neural network for estimating a phonetic sequence. .

The method may further include recognizing a speech by estimating a word sequence through a language model from the phoneme sequence estimated through the feature mapping model and the acoustic model.

Here, the motion information of the robot may be at least one of an on/off state of an internal motor of the robot, an on/off state of an internal fan, and motion information of the robot body. I can.

The speech recognition apparatus in a robot according to another embodiment extracts each input feature vector from voice data including motion information of the robot and noise input through a microphone, and concatenates the extracted input feature vectors. And, by learning the combined input feature vector through a feature mapping model and an acoustic model, speech recognition may be performed in a noise environment generated when the robot is operated.

A feature extraction unit for extracting a feature vector from speech data including noise input to the microphone; And a one-hot encoding unit receiving the motion information of the robot and performing one-hot encoding.

A first feature that combines a speech feature vector sequence extracted from the noise-containing voice data by the feature extraction unit and a motion information vector sequence that is one-hot encoded from the motion information of the robot in the one-hot encoding unit into one input vector sequence It may further include a vector combining unit.

A feature mapping unit that trains the combined input feature vector by the first feature vector combiner through the feature mapping model, which is a deep neural network (DNN) for estimating a feature vector of speech without noise; And a second feature vector combiner for extracting an improved input feature vector by combining a result learned through the feature mapping model with a motion information vector sequence one-hot encoded from motion information of the robot in the one-hot encoding unit. .

It may further include a speech recognition unit for recognizing speech by estimating a word sequence from the improved input feature vector.

The speech recognition unit may include an acoustic model unit for estimating a phoneme sequence by learning the enhanced input feature vector through the acoustic model, which is a deep neural network.

A language modeling unit for estimating the word sequence from the phoneme sequence estimated from the acoustic model unit may be further included.

Here, the motion information of the robot may be at least one of an on/off state of an internal motor of the robot, an on/off state of an internal fan, and motion information of the robot body. I can.

According to embodiments, it is possible to provide a speech recognition apparatus and method in a robot capable of robust speech recognition even in a noisy environment by using the motor operation state inside the robot as auxiliary information.

According to embodiments, it is possible to provide a speech recognition apparatus and method in a robot that minimizes discrepancy between a training environment and an actual environment by using motion state information that can be obtained by the robot itself as auxiliary information when learning a voice recognition model. .

1 is a diagram illustrating an example in which a robot according to an exemplary embodiment operates according to a user's command and generates noise.
2 is a diagram schematically showing the configuration of a voice recognition method in a robot according to an embodiment.
3 is a block diagram of a speech recognition apparatus in a robot according to an exemplary embodiment.
4 is a diagram for explaining a combination of a feature vector of voice data and motion information of a robot according to an embodiment.
FIG. 5 is a diagram for describing a combination of a feature vector of improved voice data and motion information of a robot according to an exemplary embodiment.
6 is a diagram illustrating generation of a word sequence according to an exemplary embodiment.
7 is a diagram illustrating the performance of a speech recognition device in a robot according to an exemplary embodiment.

Hereinafter, embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in various forms, and the scope of the present invention is not limited by the embodiments described below. In addition, various embodiments are provided to more completely explain the present invention to those of ordinary skill in the art. In the drawings, the shapes and sizes of elements may be exaggerated for clearer explanation.

1 is a diagram illustrating an example in which a robot according to an exemplary embodiment operates according to a user's command and generates noise.

An object of the present embodiment is to provide a speech recognition apparatus and method for minimizing inconsistency between a training environment and an actual environment by using motion state information obtained by the robot itself as auxiliary information during speech recognition model learning.

1, a robot according to an embodiment shows an example in which a robot operates according to a user's command and generates noise.In the case of a robot that interacts according to a human command, it is generated when the robot moves together with motor and fan noise. There is an internal noise, such as the noise that makes. Unlike external background noise, where it is difficult to estimate the noise type, the internal noise type of such a robot is information that can be sufficiently obtained through the motion information of the robot, so it can be used for training and testing of a deep neural network (DNN). It is possible.

2 is a diagram schematically showing the configuration of a voice recognition method in a robot according to an embodiment.

Referring to FIG. 2, a configuration of a voice recognition method in a robot according to an embodiment is schematically shown, and a voice recognition method in a robot will be described with reference to this.

The speech recognition method in a robot according to an embodiment includes the steps of extracting input feature vectors from the motion information 202 of the robot and voice data 201 including noise input through a microphone, respectively, and the extracted input features Concatenating the vectors, and learning the combined input feature vector through the feature mapping model 231 and the acoustic model 232, and the combined input feature vector with the feature mapping model 231 By learning through the acoustic model 232, speech recognition may be performed in a noisy environment generated during operation of the robot. Here, the feature mapping model 231 and the acoustic model 232 may be a deep neural network (DNN) learning (sub) 230.

Here, the motion information 202 of the robot is at least one of an on/off state of an internal motor of the robot, an on/off state of an internal fan, and motion information of the robot body. I can.

According to embodiments, it is possible to provide a speech recognition apparatus and method in a robot that minimizes discrepancy between the training environment and the actual environment by using motion state information obtained by the robot itself as auxiliary information when learning a speech recognition model. have.

Hereinafter, a voice recognition method in a robot according to an exemplary embodiment will be described in more detail with an example.

First, a voice recognition method in a robot according to an embodiment may be described in detail using a voice recognition apparatus in a robot according to an embodiment to be described later. Here, the speech recognition apparatus in the robot according to an embodiment may include a feature extraction unit, a one-hot encoding unit, a first feature vector combining unit, a feature mapping unit, a second feature vector combining unit, and a speech recognition unit. Here, the speech recognition unit may include an acoustic model 232 unit and a language model unit.

First, each input feature vector may be extracted from the motion information 202 of the robot and the voice data 201 including noise input through the microphone.

More specifically, the feature vector extractor may extract a feature vector from the speech data 201 including noise input through the microphone. In addition, the one-hot encoding unit may receive the motion information 202 of the robot and perform one-hot encoding.

Next, each of the extracted input feature vectors can be combined.

More specifically, the first feature vector combiner combines the voice feature vector sequence extracted from the noise-containing voice data 201 and the motion information vector sequence one-hot encoded from the motion information 202 of the robot into one input vector sequence. can do.

Then, the combined input feature vectors may be trained through the feature mapping model 231 and the acoustic model 232. Here, the feature mapping model 231 and the acoustic model 232 may be formed of a deep neural network, for example, a stacked deep neural network to estimate a clean speech feature. By learning the combined input feature vectors as described above through the feature mapping model 231 and the acoustic model 232, speech recognition can be performed in a noisy environment generated during the operation of the robot, and it is stacked rather than using a single deep neural network. Voice recognition performance can be improved by using a deep neural network.

More specifically, the feature mapping unit may train the combined input feature vector through the feature mapping model 231, which is a deep neural network that estimates feature vectors of speech without noise. In addition, the second feature vector combiner may extract an improved input feature vector through the feature mapping model 231. In this case, the second feature vector combiner may combine the result of learning through the feature mapping model 231 and the motion information vector sequence one-hot encoded from the motion information 202 of the robot into one input vector sequence.

In addition, the acoustic model 232 of the speech recognition unit may train the extracted enhanced input feature vector through the acoustic model 232 which is a deep neural network that estimates the phoneme sequence 203. In addition, the language model unit of the speech recognition unit may recognize a speech by estimating a word sequence through the language model from the phoneme sequence 203 estimated through the feature mapping model 231 and the acoustic model 232.

3 is a block diagram of a speech recognition apparatus in a robot according to an exemplary embodiment.

3, the speech recognition apparatus in the robot according to an embodiment extracts input feature vectors from the motion information 302 of the robot and voice data 301 including noise input through a microphone, respectively, and the extracted By concatenating each input feature vector and learning the combined input feature vector 303 through a feature mapping model and an acoustic model, speech recognition can be performed in a noisy environment generated during the operation of the robot.

Here, the motion information 302 of the robot is at least one or more of an on/off state of the robot's internal motor, an on/off state of the internal fan, and motion information of the robot body. I can. In addition, the motion information 302 of the robot may include a motion speed since noise may vary by a motor or the like according to the motion speed.

The speech recognition apparatus in the robot according to this embodiment includes a feature extraction unit 310, a one-hot encoding unit 320, a first feature vector combining unit 330, a feature mapping unit 340, and a second feature vector combining unit. It may include 350 and a voice recognition unit 360. Here, the speech recognition unit 360 may include an acoustic model unit 361 and a language model unit 362.

The feature extraction unit 310 extracts features from speech containing noise, and may extract a feature vector from speech data 301 containing noise input through a microphone.

The one-hot encoding unit 320 performs one-hot encoding on the motion information 302 of the robot, and may receive the motion information 302 of the robot and perform one-hot encoding.

The first feature vector combiner 330 may combine a feature vector containing voice information and a vector containing robot information. The first feature vector combining unit 330 includes a speech feature vector sequence extracted from the speech data 301 containing noise in the feature extraction unit 310 and the motion information 302 of the robot in the one-hot encoding unit 320. The one-hot-encoded motion information vector sequence can be combined into one input vector sequence.

The feature mapping unit 340 may train the input feature vector 303 combined by the first feature vector combiner 330 through a feature mapping model, which is a deep neural network that estimates a feature vector of a speech without noise.

In addition, the second feature vector combiner 350 combines the result of learning through the feature mapping model and the motion information vector sequence one-hot encoded from the motion information 302 of the robot in the one-hot encoding unit 320 to improve the input feature vector. (304) can be extracted.

Further, the speech recognition unit 360 is for estimating a clean speech from noise data, and may recognize the speech by estimating the word sequence 305 from the improved input feature vector 304.

Here, the speech recognition unit 360 may include an acoustic model unit 361 and a language model unit 362.

The acoustic model unit 361 may estimate the phoneme sequence by learning the improved input feature vector 304 through an acoustic model that is a deep neural network.

The language model unit 362 may estimate the word sequence 305 from the phoneme sequence estimated from the acoustic model unit 361.

FIG. 4 is a diagram for explaining a combination of a feature vector of voice data and motion information 302 of a robot according to an exemplary embodiment.

Referring to FIG. 4, the extracted voice feature vector and a one-hot vector containing robot motion information may be combined.

The feature extraction unit 310 is a part that performs a process of converting input voice data into a signal of another type. The input voice waveform can be converted into a signal of another abbreviated format having features of the voice waveform, and the signal can be abbreviated by extracting only the necessary feature signal and excluding unnecessary information. This abbreviated signal is called a feature vector. That is, the feature extraction unit 310 extracts a feature vector from voice data that is input from the microphone.

In addition, the one-hot encoding unit 320 may receive the motion information 302 of the robot and perform one-hot encoding, and through the first feature vector combiner 330, a feature vector containing voice information and a vector containing robot information By combining, the speech feature vector sequence extracted from the speech data 301 containing noise and the motion information vector sequence one-hot encoded from the motion information 302 of the robot in the one-hot encoding unit 320 are converted into one input vector sequence. Can be combined.

5 is a diagram for explaining a combination of a feature vector of improved voice data and motion information 302 of a robot according to an exemplary embodiment.

Referring to FIG. 5, an improved feature vector is obtained from the feature mapping unit 340 by inputting the feature vector combined in FIG. 4, and the improved feature vector and motion information 302 of the robot in the second feature vector combiner 350. Can be combined.

The feature mapping unit 340 learns using a deep neural network. The feature mapping unit 340 connects and inputs a feature vector extracted from voice data mixed with noise inside the robot and a vector including motion information 302 of the robot. It is designed to estimate the features of clean speech from the connected input feature vectors.

As shown in Equation 1 below, the deep neural network is trained by minimizing the mean square error between the output of the feature mapping deep neural network and the desired target clean speech.

[Equation 1]

는 f 번째 프레임에 대해 심화 신경망으로 추정한 잡음이 제거된 특징이고,

는 기준이 되는 깨끗한 특징이다.

와

는 모두

차원의 크기를 갖는 특징 벡터다. 그리고 W는 가중치 값 k는 균일화 가중치 계수, F는 mini-batch 프레임 크기를 나타낸다here,

Is the feature from which the noise estimated by the deep neural network is removed for the f- th frame,

Is a standard clean feature.

Wow

Is all

It is a feature vector of dimension size. And W is the weight value k is the equalization weight coefficient, F is the mini-batch frame size.

6 is a diagram illustrating generation of a word sequence 305 according to an exemplary embodiment.

Referring to FIG. 6, a word sequence 305 may be generated through a speech recognition model by inputting the combined feature vector of FIG. 5.

The speech recognition unit 360 is composed of an acoustic model unit 361 and a language model unit 362, and can identify the word sequence 305 for the speech data through analysis of the similarity of the feature vector data. have.

The acoustic model unit 361 may be trained using a deep neural network. The acoustic model unit 361 is a portion for generating an acoustic model necessary for the speech recognition unit 360 to recognize a word, and is a portion for recognizing and modeling a word pronounced by a user in phoneme units.

In this case, the acoustic model unit 361 connects the enhanced feature vector output of the feature mapping unit 340 and the vector including the robot motion information 302 as an input, and estimates the phoneme sequence from the connected feature vector. can do.

The language modeling unit 362 is a part that finds a relationship between words in a given sentence and reflects the relationship to speech recognition. Particularly, when several words are given in order, we pay attention to the fact that the next word is highly related to the previous word. This language model can use a general statistical model.

According to embodiments, in addition to the voice data information obtained from the microphone during the learning and testing of the feature mapping network and the acoustic model network, the robot motion information 302 is used as auxiliary information to enable robust voice recognition even in the internal motion noise. .

7 is a diagram illustrating the performance of a speech recognition device in a robot according to an exemplary embodiment.

Referring to FIG. 7, spectrum graphs are compared to describe the performance of a speech recognition device in a robot according to an exemplary embodiment. Here, (a) represents the original noise-free clear speech signal, (b) represents the microphone's input speech signal (noise state at 5 dB), and (c) represents the single deep neural network (DNN)-based algorithm. It is the result, and (d) it is the result of reinforcement with the stacked deep neural network (DNN) based algorithm

Here, the stacked deep neural network (DNN)-based algorithm is the result of estimating clean speech features through the feature mapping model and acoustic model consisting of the deep neural network (DNN) described above. When a stacked deep neural network (DNN)-based algorithm is used, speech recognition performance can be improved than when a single deep neural network (DNN)-based algorithm is used.

Table 1 shows an example of the performance of a speech recognition device in a robot according to an embodiment.

[Table 1]

Referring to Table 1, it shows the phonetic error rate (PER) performance of the TIMIT dataset, and the model is divided into cases without the features of auxiliary information (robot motion information) and cases with the features of auxiliary information, respectively. Three models: (1) model without feature mapping (X), (2) model with single deep neural network (DNN)-based feature mapping, (3) stacked model model with deep neural network (DNN)-based feature mapping And compared the performance.

As a result, the performance of the model using auxiliary information (robot motion information) is superior to all models of the existing model. In addition, when using auxiliary information, the shape mapping model using a stacked deep neural network (DNN) showed the best performance. Therefore, robot motion information such as a motor may be a related feature.

As described above, by using the speech recognition method according to the embodiments, the operation state of the robot can be utilized when learning the speech recognition model, and robust speech recognition is possible even in the presence of noise generated inside the robot as well as background noise. can do.

Such a speech recognition device and method in a robot can be applied to a speech recognition technology in a social robot and a speech recognition technology in an artificial intelligence speaker.

The apparatus described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices and components described in the embodiments include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It can be implemented using one or more general purpose computers or special purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to behave as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. Can be embodyed. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -A hardware device specially configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described by the limited embodiments and drawings, various modifications and variations are possible from the above description by those of ordinary skill in the art. For example, the described techniques are performed in a different order from the described method, and/or components such as a system, structure, device, circuit, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and claims and equivalents fall within the scope of the claims to be described later.

Claims (15)

Extracting input feature vectors from speech data including motion information of the robot and noise input through the microphone;
Concatenating each of the extracted input feature vectors; And
Learning the combined input feature vector through a feature mapping model and an acoustic model
Including,
Extracting an input feature vector from voice data including motion information of the robot and noise input through a microphone,
Extracting a feature vector from speech data including noise input to the microphone; And
Receiving the motion information of the robot and performing one-hot encoding
Including,
Combining the extracted respective input feature vectors,
Combine the voice feature vector sequence extracted from the noise-containing voice data and the motion information vector sequence one-hot encoded from the motion information of the robot into one input vector sequence,
The step of learning the combined input feature vector through a feature mapping model and an acoustic model,
Training the combined input feature vector through the feature mapping model, which is a deep neural network (DNN) that estimates a feature vector of a speech without noise; And
Extracting an improved input feature vector by combining the result learned through the feature mapping model with a one-hot-encoded motion information vector sequence from motion information of the robot
Including,
A speech recognition method in a robot, wherein the combined input feature vector is learned through a feature mapping model and an acoustic model to perform speech recognition in a noisy environment generated during operation of the robot. delete delete delete The method of claim 1,
The step of learning the combined input feature vector through a feature mapping model and an acoustic model,
Learning the extracted enhanced input feature vector through the acoustic model, which is a deep neural network for estimating a phonetic sequence
A method for speech recognition in a robot further comprising a. The method according to claim 1 or 5,
Recognizing a speech by estimating a word sequence through a language model from the phoneme sequence estimated through the feature mapping model and the acoustic model
A method for speech recognition in a robot further comprising a. The method of claim 1,
The motion information of the robot,
At least one of at least one of an on/off state of the robot's internal motor, an on/off state of an internal fan, and motion information of the robot body
Characterized in, the voice recognition method in the robot. A feature extracting unit for extracting a feature vector from speech data including noise input through a microphone configured in the robot; And
One-hot encoding unit that receives the motion information of the robot and performs one-hot encoding
Including,
A first feature that combines a speech feature vector sequence extracted from the noise-containing voice data by the feature extraction unit and a motion information vector sequence that is one-hot encoded from the motion information of the robot in the one-hot encoding unit into one input vector sequence Vector joiner
A feature mapping unit that trains the combined input feature vector by the first feature vector combiner through the feature mapping model, which is a deep neural network (DNN) for estimating a feature vector of speech without noise; And
A second feature vector combiner for extracting an improved input feature vector by combining the result learned through the feature mapping model with a motion information vector sequence one-hot encoded from motion information of the robot in the one-hot encoding section
Including more,
Each input feature vector is extracted from the motion information of the robot and the voice data including noise input to the microphone, and the extracted input feature vectors are concatenated, and the combined input feature vector is used as a feature mapping model. A speech recognition device implemented in a robot that performs speech recognition in a noisy environment generated during operation of the robot by learning through an acoustic model. delete delete delete The method of claim 8,
Speech recognition unit for recognizing speech by estimating a word sequence from the enhanced input feature vector
A voice recognition device implemented in the robot further comprising a. The method of claim 12,
The voice recognition unit,
An acoustic model unit for estimating a phoneme sequence by learning the enhanced input feature vector through the acoustic model, which is a deep neural network
Including a voice recognition device implemented in the robot. The method of claim 13,
Language model unit for estimating the word sequence from the phoneme sequence estimated from the acoustic model unit
A voice recognition device implemented in the robot further comprising a. The method of claim 8,
The motion information of the robot,
At least one of at least one of an on/off state of the robot's internal motor, an on/off state of an internal fan, and motion information of the robot body
Characterized in, the voice recognition device implemented in the robot.

Images