WO2022177063A1

WO2022177063A1 - Electronic device and control method therefor

Info

Publication number: WO2022177063A1
Application number: PCT/KR2021/003944
Authority: WO
Inventors: 마흐무드 토마아나스; 술레이만 무하마드 알자파리아야; 무스타파 마흐무드 야센압둘라만; 마사이드알리 아흐메드 아델 알; 칼레드 사데모하메드; 후세인 에드웨인모드
Original assignee: 삼성전자주식회사
Priority date: 2021-02-18
Filing date: 2021-03-30
Publication date: 2022-08-25
Also published as: KR20220118242A

Abstract

Disclosed are an electronic device and a control method therefor. The electronic device comprises a microphone, a memory and a processor, which obtains a plurality of voice segments by inputting a first user voice, input through the microphone, into a first model trained to segment voices on the basis of word boundaries, obtains the language type of each of the plurality of voice segments by inputting the plurality of voice segments into a second model trained to identify language types of voices, obtains a plurality of text segments corresponding to the plurality of voice segments, respectively, by inputting each of the plurality of voice segments into an ASR model, corresponding to the language type of each of the plurality of voice segments, from among a plurality of ASR models, and obtains text corresponding to the first user voice, on the basis of the plurality of text segments.

Description

Electronic device and its control method

The present disclosure relates to an electronic device and a method for controlling the same, and more particularly, to an electronic device for obtaining text corresponding to a user's voice using an artificial intelligence model, and a method for controlling the same.

Recently, artificial intelligence systems have been used in various fields. Unlike the existing rule-based smart system, the artificial intelligence system is a system in which the machine learns, judges, and becomes smarter by itself. The more the artificial intelligence system is used, the better the recognition rate and the more accurate understanding of user preferences, and the existing rule-based smart systems are gradually being replaced by deep learning-based artificial intelligence systems.

Recently, Auto Speech Recognition (ASR) technology for recognizing a user's voice using a deep learning-based artificial intelligence system and converting the recognized voice into text has been widely used. On the other hand, it is common that the ASR model performing the ASR technique is learned based on one specific language. Therefore, when a Korean voice is input to the English-based ASR model, there is a high possibility that the ASR model cannot accurately output the text corresponding to the voice.

Accordingly, conventionally, a method for learning an ASR model based on a plurality of languages has been devised. For example, by training an ASR model based on learning data composed of English and Korean, a method for more accurately outputting text corresponding to a mixed voice of English and Korean was devised. However, there are limitations in that learning one ASR model based on a plurality of languages is disadvantageous in terms of learning time and accuracy as well as high learning and maintenance costs.

Also, conventionally, when a user's voice is input, a method of recognizing a language type corresponding to the input user's voice and inputting the user's voice into an ASR model corresponding to the recognized language type has been devised.

However, when a plurality of types of languages are mixed with the voice input to the ASR model, there is a high possibility that the ASR model learned based on one language cannot output the text corresponding to the correct input voice. The above problem can be further raised in that the number of speakers who can use more than two languages (bilingual) is increasing worldwide in recent years.

In addition, in the case of the above method, since the language type of the user's voice is recognized based on the user's voice input for a short period of time (eg, 50 ms), there is a problem that the entire language type included in the user's voice cannot be accurately determined. .

The present disclosure has been devised to solve the above problems, and an object of the present disclosure is to segment an input user voice, and input the segmented voice segment into an ASR model corresponding to the language type of each voice segment to respond to the user voice An object of the present invention is to provide an electronic device for obtaining a text to be used and a method for controlling the same.

In an embodiment of the present disclosure, the electronic device acquires a plurality of voice segments by inputting a first user voice input through the microphone into a first model trained to segment a voice based on a microphone, a memory, and a word boundary, and , input the plurality of speech segments into a second model trained to identify a language type of speech to obtain a language type of each of the plurality of speech segments, and apply each of the plurality of speech segments to the plurality of speech segments among a plurality of ASR models. A plurality of text segments corresponding to each of the plurality of speech segments are obtained by inputting into an ASR model corresponding to the language type of each speech segment, and a text corresponding to the first user's speech is obtained based on the plurality of text segments. It may include a processor that

As another embodiment of the present disclosure, a method for controlling an electronic device acquires a plurality of voice segments by inputting a first user voice input to the electronic device into a first model trained to segment a voice based on a word boundary. obtaining a language type of each of the plurality of speech segments by inputting the plurality of speech segments into a second model trained to identify a language type of speech; obtaining a plurality of text segments corresponding to each of the plurality of speech segments by inputting them into an ASR model corresponding to the language type of each of the plurality of speech segments; and corresponding to the first user's speech based on the plurality of text segments It may include the step of obtaining the text to be

According to various embodiments as described above, there is an effect that the electronic device may more accurately obtain a text corresponding to a user's voice in which a plurality of languages are mixed.

1 is a block diagram schematically illustrating a configuration of an electronic device according to an embodiment of the present disclosure;

2 is a view for explaining a process in which an electronic device acquires text corresponding to a user's voice according to an embodiment of the present disclosure;

3 is a view for explaining a process in which an electronic device divides a user's voice according to an embodiment of the present disclosure;

4 is a view for explaining the configuration and operation of the first model and the second model, according to an embodiment of the present disclosure;

5 is a flowchart illustrating a method of controlling an electronic device according to an embodiment of the present disclosure;

6 is a sequence diagram illustrating operations of an electronic device and a server according to an embodiment of the present disclosure;

7 is a view for explaining a process in which an electronic device acquires text corresponding to a user's voice according to an embodiment of the present disclosure;

8 and 9 are flowcharts for explaining a process in which an electronic device acquires text corresponding to a user's voice based on voice reliability according to an embodiment of the present disclosure;

10 is a detailed block diagram illustrating the configuration of an electronic device according to an embodiment of the present disclosure.

Prior to describing the present disclosure in detail, a description will be given of the description of the present specification and drawings.

First, terms used in the present specification and claims have been selected in consideration of functions in various embodiments of the present disclosure. However, these terms may vary depending on the intention or legal or technical interpretation of a person skilled in the art, and the emergence of new technology. Also, some terms are arbitrarily selected by the applicant. These terms may be interpreted in the meaning defined in the present specification, and if there is no specific term definition, it may be interpreted based on the general content of the present specification and common technical knowledge in the art.

Also, the same reference numerals or reference numerals in each drawing attached to this specification indicate parts or components that perform substantially the same functions. For convenience of description and understanding, the same reference numerals or reference numerals are used in different embodiments. That is, even though all components having the same reference number are illustrated in a plurality of drawings, the plurality of drawings do not mean one embodiment.

Various elements and regions in the drawings are schematically drawn. Accordingly, the technical spirit of the present invention is not limited by the relative size or spacing drawn in the accompanying drawings.

In addition, in this specification and claims, terms including an ordinal number, such as "first" and "second", may be used to distinguish between elements. This ordinal number is used to distinguish the same or similar elements from each other, and the meaning of the term should not be construed as limited due to the use of the ordinal number. For example, the components combined with such an ordinal number should not be limited in the order of use or arrangement by the number. If necessary, each ordinal number may be used interchangeably.

In this specification, the singular expression includes the plural expression unless the context clearly dictates otherwise. In the present application, terms such as "comprises" or "consisting of" are intended to designate that the features, numbers, steps, operations, components, parts, or combinations thereof described in the specification exist, and are intended to indicate that one or more other It is to be understood that this does not preclude the possibility of addition or presence of features or numbers, steps, operations, components, parts, or combinations thereof.

The expression "configured to (or configured to)" as used in this disclosure depends on the context, for example, "suitable for," "having the capacity to" ," "designed to," "adapted to," "made to," or "capable of." The term “configured (or configured to)” may not necessarily mean only “specifically designed to” in hardware.

Instead, in some circumstances, the expression “a device configured to” may mean that the device is “capable of” with other devices or parts. For example, the phrase “a coprocessor configured (or configured to perform) A, B, and C” may refer to a dedicated processor (eg, an embedded processor), or one or more software programs stored in a memory device, to perform the corresponding operations. By doing so, it may mean a generic-purpose processor (eg, a CPU or an application processor) capable of performing corresponding operations.

In an embodiment of the present disclosure, terms such as “module”, “unit”, “part”, etc. are terms for designating a component that performs at least one function or operation, and such component is hardware or software. It may be implemented or implemented as a combination of hardware and software. In addition, a plurality of "modules", "units", "parts", etc. are integrated into at least one module or chip, except when each needs to be implemented as individual specific hardware, and thus at least one processor. can be implemented as

A component (eg, a first component) is "coupled with/to (operatively or communicatively)" to another component (eg, a second component) When referring to "connected to", it should be understood that the certain element may be directly connected to the other element or may be connected through another element (eg, a third element). On the other hand, when it is said that a component (eg, a first component) is "directly connected" or "directly connected" to another component (eg, a second component), the component and the It may be understood that other components (eg, a third component) do not exist between other components.

Hereinafter, the present disclosure will be described in detail with reference to the drawings.

As an embodiment of the present disclosure, FIG. 1 is a block diagram schematically illustrating a configuration of an electronic device 100 . 1 , the electronic device 100 may include a memory 110 , a microphone 120 , and a processor 130 . However, the configuration shown in FIG. 1 is an exemplary diagram for implementing embodiments of the present disclosure, and appropriate hardware and software configurations at a level obvious to those skilled in the art may be additionally included in the electronic device 100 .

The memory 110 may store data related to at least one other component of the electronic device 100 or at least one instruction. The command means one action statement that can be directly executed by the processor 130 in a programming language, and is a minimum unit for execution or operation of a program. The memory 110 is accessed by the processor 130 , and reading/writing/modification/deletion/update of data by the processor 130 may be performed.

In the present disclosure, the term memory may be implemented as a memory 110 , a ROM (not shown), a RAM (not shown) in the processor 130 , or a memory separate from the processor 130 . In this case, the memory 110 may be implemented in the form of a memory embedded in the electronic device 100 or may be implemented in the form of a memory detachable from the electronic device 100 according to the purpose of data storage. For example, data for driving the electronic device 100 is stored in a memory embedded in the electronic device 100 , and data for an extended function of the electronic device 100 is detachable from the electronic device 100 . It can be stored in any available memory.

The microphone 120 may receive a user's voice. In this case, the microphone 120 may receive a trigger voice (or a wake-up voice) indicating the start of voice recognition, and may receive a user query or command requesting specific information.

The microphone 120 may be provided inside the electronic device 100 , but may be provided outside and electrically connected to the electronic device 100 . As another example, the microphone 120 may be provided outside the electronic device 100 to be connected to the electronic device 100 through a wireless communication module.

The processor 130 may be electrically connected to the memory 110 to control overall operations and functions of the electronic device 100 .

The processor 130 may include one or a plurality of processors. At this time, one or more processors are general-purpose processors such as CPU (Central Processing Unit), AP (Application Processor), DSP (Digital Signal Processor), etc., and dedicated graphics such as GPU (Graphic Processing Unit) and VPU (Vision Processing Unit). It can be a processor or a processor dedicated to artificial intelligence, such as a Neural Processing Unit (NPU). When one or more processors are AI-only processors, the AI-only processor may be designed with a hardware structure specialized for processing a specific AI model.

In addition, the processor 130 may be implemented as a system on chip (SoC), large scale integration (LSI), or a field programmable gate array (FPGA) having a built-in processing algorithm. In addition, the processor 130 may perform various functions by executing computer executable instructions stored in the memory.

An operation performed by the processor 130 will be described in detail with reference to FIGS. 2 to 3 . As shown in FIG. 2 , the processor 130 may receive a voice input from the user 10 through the microphone 120 . The processor 130 may acquire a plurality of voice segments by inputting the voice input from the user into the first model 20 .

The first model 20 refers to an artificial intelligence model trained to divide speech into at least one segment based on a word boundary. In one embodiment, the first model 20 may be trained to identify at least one word boundary based on waveform data of a user's speech. The word boundary may mean a boundary for separating each of at least one word included in the user's voice or a boundary for separating a reading section spaced for each word.

For example, it is assumed that the voice input from the user 10 through the microphone 120 is 'Que pasa, call me'. The first model 20 may divide the input voice into four segments based on word boundaries. The first voice segment (or seg 1) is 'Que', the second voice segment (or seg 2) is 'pasa', the third voice segment (or seg 3) is 'call', and the fourth The voice segment (or seg 4) may be 'me'.

A process in which the first model 20 divides the user's voice into at least one voice segment will be described in more detail with reference to FIG. 3 . In the case of the graph shown in FIG. 3 , the x-axis (horizontal direction) indicates time, and the y-axis (vertical direction) indicates a waveform graph indicating the strength of a voice.

The first model 20 may store a voice signal from the point in time 310 when the user's voice is input until the first word boundary is detected in a buffer. When the first word boundary 320 is detected, the first model 10 generates a voice signal (that is, including 'Que') from the point in time 310 when the user's voice is input to the detected first word boundary 320. voice signal) as the first voice segment. Here, the detection of the word boundary may mean that a preset time (ie, a spaced reading time period) has elapsed from the utterance of a specific word to the utterance of the next word.

In addition, the first model 20 may store the speech signal from the first word boundary 320 until the next word boundary is detected in the buffer. When the second word boundary 330 is detected, the first model 20 generates a voice signal from the first word boundary 320 to the second word boundary 330 (that is, a voice signal including 'pasa'). It can be obtained as a second voice segment. The first model 20 may acquire the third and fourth voice segments based on the

respective word boundaries

340 and 350 by repeating the above-described method. That is, the first model 20 may divide the user's voice into four voice segments in the manner described above.

Referring to FIG. 2 , the processor 130 may obtain a language type of each of the plurality of speech segments by inputting a plurality of speech segments into the second model 30 . Acquiring the language type of the voice segment may mean acquiring ID (Identification) information (identification information) indicating the language type of the voice segment. And, the second model 30 means an artificial intelligence model trained to identify the language type of speech.

In an embodiment, when the language type of the voice segment is identified, the second model 30 may output ID information that may indicate the identified language type. For example, the second model 30 outputs ID information (eg, <ES>) that the language type of the first voice segment and the second voice segment is Spanish, and the third voice segment and the fourth voice segment are English ID information (eg, <EN>) can be output.

The processor 130 inputs each of the plurality of speech segments into an ASR model corresponding to the language type of each of the plurality of speech segments among the plurality of Auto Speech Recognition (ASR)

models

200 , 210 , and 230 to obtain a plurality of speech segments corresponding to each of the plurality of speech segments. A text segment can be obtained.

Meanwhile, as shown in FIG. 2 , each

ASR model

200 , 210 , 230 may be stored in an external server. As another embodiment, each of the

ASR models

200 , 210 , and 230 may be stored in a separate external server instead of one external server. As another embodiment, at least one of the

ASR models

200 , 210 , and 230 may be built in the electronic device 100 .

In addition, each of the plurality of

ASR models

200 , 210 , and 220 may be a model learned based on learning data related to a specific language type. For example, the Spanish (hereinafter, ES) ASR model is a model trained based on Spanish learning data, the English (hereinafter, EN) ASR model is a model trained based on English learning data, and Korean (hereinafter, KR) The ASR model may be a model trained based on Korean learning data. Of course, the ASR model shown in FIG. 2 is only an example and may be implemented as a model learned based on another language (eg, addition/deletion/change, etc.).

In an embodiment, the processor 130 may obtain a text segment by inputting a specific voice segment into an ASR model corresponding to the specific voice segment through the ASR router 40 . The ASR router 40 means a software or hardware module configured to identify an ASR model corresponding to a language type of a voice segment and transmit the voice segment to the identified ASR model.

For example, when the language type ID information of the first and second voice segments is ES, the processor 130 transmits each of the first and second voice segments to the ES ASR model 200 through the ASR router 40 . It is possible to obtain a first text segment (eg, 'Que') and a second text segment (eg, 'pasa') corresponding to the first voice segment by inputting the . When the ES ASR model 200 is stored in an external server, the processor 130 may control the communication unit 140 to transmit the first and second voice segments to the external server. The communication unit 140 will be described in detail with reference to FIG. 10 .

As another example, when the language type ID information of the third and fourth voice segments is EN, the processor 130 inputs each of the third and fourth voice segments to the EN ASR model 210 through the ASR router, A third text segment (eg, 'call') corresponding to the third voice segment and a fourth text segment (eg, 'me') corresponding to the fourth voice segment may be acquired.

The processor 130 may obtain a text corresponding to the user's voice based on the plurality of text segments. In an embodiment, the processor 130 uses the post-processing module 50 to generate a plurality of voice segments corresponding to the plurality of text segments according to the order in which they are inputted through the microphone 120 . A text corresponding to the first user's voice may be obtained by combining the text segments of . Here, the post-processing module 50 refers to a module that combines a plurality of text segments using a language type or an input sequence for each of a plurality of speech segments obtained through the second model 30 .

For example, each of the first to fourth text segments may correspond to each of the first to fourth voice segments. The processor 130 generates text (eg, Que pasa call me) corresponding to the user's voice by combining text segments corresponding to each voice segment according to the order in which each voice segment is input through the post-processing module 50 . can be obtained

According to the above-described embodiment of the present disclosure, the electronic device 100 does not need to receive an input of a language type for recognition before using the ASR model from the user.

In addition, when the voice input from the user includes various types of languages, the electronic device 100 identifies the language types of the divided voice segments, and inputs each voice segment into an ASR model corresponding to the identified language type. By doing so, a more accurate text segment can be obtained.

In addition, the electronic device 100 does not need to additionally learn the existing ASR model by using words (eg, entity names, etc.) of a type of language that the existing ASR model has not learned.

On the other hand, as another embodiment of the present disclosure, when a user command requesting that the second model 30 learn to recognize a third language type that has not been learned is input, the processor 130 is related to the third language type The second model 30 may be trained based on the training data. However, this is only an embodiment, and the processor 130 may train both the first model 20 and the second model 30 based on the learning data related to the third language type.

For example, it is assumed that the second model 30 is an artificial intelligence model learned based on English, Spanish, and Korean. The processor 130 may receive a user command for requesting learning to recognize a user's voice including German. Specifically, the processor 130 may display a setting UI for setting a language that the electronic device 100 can recognize. The processor 130 may receive a user command requesting to learn to recognize German through the setting UI.

When a user command is input, the processor 130 may receive learning data related to German from an external server through the communication unit 140 , and may further train the second model 30 based on the received learning data. However, this is only an embodiment, and the processor 130 may additionally train both the first model 20 and the second model 30 based on the received training data.

For example, the first model 20 may output a plurality of speech segments by detecting a word boundary of a user's speech including German. In addition, the second model 20 may identify a specific voice segment made of German among a plurality of voice segments, and output ID information indicating the language type of the specific voice segment. In addition, the processor 130 may input a German voice segment into the German-based ASR model.

According to the above-described embodiment of the present disclosure, the electronic device 100 does not need to additionally train the ASR model in order to recognize a voice including a new type of language and output a text corresponding thereto.

Meanwhile, the function related to artificial intelligence according to the present disclosure is operated through the processor 130 and the memory 110 .

One or more processors 130 control to process input data according to a predefined operation rule or artificial intelligence model stored in the memory 110 .

The predefined action rule or artificial intelligence model is characterized in that it is created through learning. Here, being made through learning means that a basic artificial intelligence model is learned using a plurality of learning data by a learning algorithm, so that a predefined action rule or artificial intelligence model set to perform a desired characteristic (or purpose) is created means burden. Such learning may be performed in the device itself on which artificial intelligence according to the present disclosure is performed, or may be performed through a separate server and/or system.

Examples of the learning algorithm include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning.

The artificial intelligence model includes a plurality of artificial neural networks, and the artificial neural network may be composed of a plurality of layers. Each of the plurality of neural network layers has a plurality of weight values, and a neural network operation is performed through an operation between an operation result of a previous layer and a plurality of weights. The plurality of weights of the plurality of neural network layers may be optimized by the learning result of the artificial intelligence model. For example, a plurality of weights may be updated so that a loss value or a cost value obtained from the artificial intelligence model during the learning process is reduced or minimized.

4 is a view for explaining the configuration and operation of a first model and a second model according to an embodiment of the present disclosure.

In an embodiment, each of the first model and the second model includes at least one fully connected layer (FC) layer (410-1,410-2,430-1,430-2,430-3) and at least one self-attention layer (self-attention layer). attention layers) 400-1, 400-2, 440-1, and 440-2 may be included. The self-attention layer may be included in each model to consider adjacent frames while processing some frames of the input user's voice.

As shown in FIG. 4 , the first model may extract a mel-spectrum by using an input user voice. The first model may output the result data 420 by inputting the extracted Mel spectrum to the self-attention layer 400 - 1 or the FC layer. For example, in the

result data

420 , 1 may mean a word boundary and 0 may mean a word. That is, the first model may divide the user's voice into three voice segments based on three word boundaries.

In addition, the second model may output an ID indicating the language type of the voice segment by inputting the voice segment to the pre-connection layer 430-1 or another self-attention layer.

Meanwhile, although FIG. 4 illustrates a case in which the first model and the second model are separate models, the present invention is not limited thereto, and the first model and the second model may be implemented as one model. For example, it is assumed that the first model and the second model are implemented as one model (eg, the third model). When a user's voice is input, the third model may output a plurality of voice segments and language type information of each of the plurality of voice segments.

Also, the first model and the second model may be stored in the electronic device 100 , but are not limited thereto. At least one of the first model and the second model may be stored in an external server.

5 is a flowchart illustrating a method of controlling the electronic device 100 according to an embodiment of the present disclosure.

The electronic device 100 may receive the first user voice (S510). The electronic device 100 may obtain a plurality of voice segments by inputting the first user voice into the first model (S520). The first model may detect a word boundary in the user's voice, and divide and identify the user's voice into a plurality of voice segments based on the detected word boundary.

The electronic device 100 may obtain a language type of each of the plurality of speech segments by inputting the plurality of speech segments into the second model ( S530 ). The second model may output ID information indicating the language type of each of the plurality of voice segments.

The electronic device 100 may obtain a plurality of text segments corresponding to each of the plurality of speech segments by inputting each of the plurality of speech segments into an ASR model corresponding to a language type of each of the plurality of speech segments among the plurality of models ( S540).

For example, it is assumed that the language type of each of the first voice segment and the second voice segment among the plurality of voice segments is a first language type (eg, Korean) and a second language type (eg, English). The electronic device 100 may obtain a text segment corresponding to the first voice segment by inputting the first voice segment into an ASR model corresponding to a first language type (eg, Korean) among a plurality of ASR models. In this case, the ASR model corresponding to the first language type means a model learned based on the first language type. In addition, the electronic device 100 may obtain a text segment corresponding to the second voice segment by inputting the second voice segment into an ASR model corresponding to the second language type among the plurality of ASR models.

The electronic device 100 may obtain a text corresponding to the first user's voice based on the plurality of text segments ( S550 ). Specifically, the electronic device 100 obtains the text corresponding to the first user voice by combining the plurality of text segments according to the order in which the voice segments corresponding to the plurality of text segments are input to the electronic device 100 . can do.

6 is a sequence diagram illustrating operations of an electronic device and a server according to an embodiment of the present disclosure. Since each of S610 to S630 of FIG. 6 is the same as each of S510 to S530 of FIG. 5 , a redundant description will be omitted.

The electronic device 100 obtains a first voice segment and a second voice segment, a language type of the first voice segment is a first language type (eg, Korean), and a language type of the second voice segment is a second language type (eg, English) is assumed. When the first ASR model learned based on the first language type is stored in the first server 600 and the second ASR model learned based on the second language type is stored in the second server 605 assume

The electronic device 100 may input each voice segment into an ASR model corresponding to the language type of each voice segment among the plurality of ASR models. For example, the electronic device 100 may transmit the first voice segment to the first server 600 in which the first ASR model is stored ( S640 ). Then, the electronic device 100 may transmit the second voice segment to the second server 605 in which the second ASR model is stored. In this case, the first voice segment may be a voice segment input to the electronic device 100 before the second voice segment.

However, this is only an embodiment, and the electronic device 100 may transmit each of the first voice segment and the second voice segment to each of the first server 600 and the second server 605 at the same time or in any order. have.

The first server 600 may obtain the first text segment by inputting the received first voice segment into the first ASR model ( S660 ). The first server 600 may transmit the first text segment to the electronic device 100 (S665). Then, the second server 605 may obtain the second text segment by inputting the received second voice segment into the second ASR model ( S670 ). The second server 605 may transmit the second text segment to the electronic device 100 (S675). Meanwhile, the order of S660 to S675 may be changed. That is, the second server 605 may first obtain the second text segment and transmit it to the electronic device 100 .

The electronic device 100 may obtain a text corresponding to the user's voice based on the obtained text segment (S680). The electronic device 100 may obtain a text corresponding to the user's voice by combining the first text and the second text segment according to the order in which the first and second voice segments are input.

7 is a diagram for describing a process in which the electronic device 100 acquires text corresponding to a user's voice according to an embodiment of the present disclosure.

The electronic device 100 may divide or convert the voice input from the user into a plurality of frames having a predefined time period (eg, 25 ms). In addition, the electronic device 100 may filter a non-speech frame. Accordingly, the electronic device 100 may acquire only a plurality of voice frames. Here, the voice frame may mean a frame including a voice component, and the non-voice frame may mean a frame (eg, a silence or a silence period) that does not include a voice component.

For example, the electronic device 100 classifies a voice frame and a non-voice frame among the user's voices using Speech Activity Detection (SAD) (or Voice Activity Detection, VAD) technology, and identifies the non-voice frame. can be filtered.

The electronic device 100 may acquire the language types of the plurality of voice frames by inputting the plurality of acquired voice frames into the second model 700 . The language type of the voice frame may mean a language type corresponding to a voice component included in the voice frame. Also, the second model 700 may be the same model as the second model 30 described with reference to FIG. 2 . As another example, the second model 700 may be implemented as an artificial intelligence model trained to acquire a language type for each frame having a pre-defined time (eg, 25 ms) for a language frame.

The electronic device 100 may obtain text segments corresponding to the plurality of voice frames by inputting each of the plurality of voice frames into the ASR model corresponding to the language type of each of the plurality of voice frames.

For example, as shown in FIG. 7 , the electronic device 100 may divide or convert the user's voice into a plurality of frames having a predefined time by using the SVD technology. In addition, the electronic device 100 may obtain a plurality of voice frames by filtering non-voice frames among the plurality of frames.

When the language type of the first to fourth frames among the plurality of voice frames is Korean, the electronic device 100 inputs the first to fourth frames into the Korean-based ASR model and text corresponding to each frame. segment can be obtained. And, when the language type of the fifth to sixth frames is English, the electronic device 100 inputs the fifth to sixth frames into the English-based ASR model to obtain text segments corresponding to each frame. can do.

In this case, each ASR model may be included in the electronic device 100 , but may be stored in an external server. Each ASR model may be stored in one external server or may be stored in separate external servers.

The electronic device 100 may acquire the text corresponding to the user's voice by inputting the acquired text segment into the post-processing module 710 . The electronic device 100 may obtain a text corresponding to the user's voice by combining the plurality of text segments according to the order in which frames corresponding to the plurality of text segments are input to the electronic device 100 .

Meanwhile, according to another embodiment of the present disclosure, the electronic device 100 may recombined the filtered non-voice frame and the plurality of voice frames according to the order they are input to the electronic device 100 . In addition, the electronic device 100 inputs the first voice frame and the first non-voice frame adjacent to the first voice frame among the plurality of voice frames into the ASR model corresponding to the language type of the first voice frame, and inserts the first voice frame into the first voice frame. The corresponding text can be obtained. When a non-voice frame is recombined with a speech frame and the recombined frame is input to the ASR model, robustness of speech recognition by the ASR model may be increased.

8 is a flowchart illustrating a process in which the electronic device 100 acquires text corresponding to a user's voice based on voice reliability, according to an embodiment of the present disclosure.

When a user voice is input, the electronic device 100 may input the input user voice into a first ASR model among a plurality of ASR models ( S810 ). The first ASR model means an ASR model learned based on the type of language set (or selected) by the user, and may be expressed as a basic ASR model (or an original ASR model). For example, when the user selects Korean among a plurality of languages, the electronic device 100 may input the input user voice into an ASR model (ie, a basic ASR model) learned based on Korean.

The electronic device 100 may identify whether a voice confidence output through the first ASR model is less than a threshold value ( S820 ). Speech reliability is a numerical measure of how reliable the ASR model is to recognize a voice and output it as text corresponding to the voice. A high speech reliability may mean that the ASR model recognizes a speech with high reliability and outputs a text corresponding thereto. That the voice reliability is less than the threshold value may mean that the first ASR model needs to be further updated (or retrained). And, it goes without saying that the threshold value may be a value preset by research or experimentation, and may be changed by a user.

On the other hand, when the input user voice includes a word corresponding to a foreign word based on the language type learned by the first ASR model, the first ASR model may not accurately recognize the voice corresponding to the foreign word among the user's voices. this exists Accordingly, there is a possibility that the first ASR model outputs a speech reliability below the threshold.

For example, it is assumed that the first ASR model is a model learned based on Korean, and the user's voice input to the electronic device 100 is 'find an accessory house nearby'. At this time, since 'accessory' is a foreign word introduced from English, the first ASR model may not accurately recognize the voice corresponding to the accessory among the input user voices. Accordingly, the first ASR model may output a voice reliability lower than the threshold value.

When the voice reliability is less than the threshold, the electronic device 100 may additionally update (or relearn) the first ASR model based on the text corresponding to the user's voice and the user's voice ( S830 ).

As an embodiment, when the voice reliability is less than the threshold, the electronic device 100 may receive a text corresponding to the user's voice from the user in order to update the first ASR model. Specifically, the electronic device 100 may display a UI requesting input of text corresponding to the input user's voice. In addition, the electronic device 100 may receive a text corresponding to the user's voice input from the user through the displayed UI.

As another embodiment, the electronic device 100 may transmit the user's voice to the server including the ASR model learned based on the type of language learned by the first ASR model. Then, the electronic device 100 may receive a text corresponding to the user's voice from the server. The ASR model learned in the server is learned based on a larger amount of training data than the ASR stored in the terminal device, so that the recognition rate may be relatively high.

The electronic device 100 may update the first ASR model by using the text corresponding to the user's voice input from the user or the server and the input user's voice as training data. That is, the electronic device 100 may additionally update the first ASR model to recognize the user's voice and output text corresponding to the user's voice. With further updates, performance such as speech recognition rate of the first ASR model may be improved.

Meanwhile, when the voice reliability output through the first ASR model is equal to or greater than the threshold value, the electronic device 100 may obtain a text corresponding to the user's voice through the first ASR model ( S840 ). That is, the electronic device 100 may obtain the output text of the first ASR model as text corresponding to the user's voice.

9 is a flowchart illustrating a process in which the electronic device 100 acquires text corresponding to a user's voice based on voice reliability, according to an embodiment of the present disclosure.

When a user voice is input, the electronic device 100 may input the input user voice into a first ASR model among a plurality of ASR models ( S910 ). As described with reference to FIG. 8 , the first ASR model refers to an ASR model learned based on a type of language set (or selected) by a user.

The electronic device 100 may identify whether the voice reliability output through the first ASR model is less than a threshold value (S920).

When the voice reliability is less than the threshold, the electronic device 100 may obtain a text corresponding to the user's voice through at least one ASR model (S930). Specifically, as described with reference to FIG. 5 , the electronic device 100 divides the user's voice into a plurality of voice segments, and inputs the divided voice segments into an ASR model corresponding to the language type of each voice segment to segment the text. may be obtained, and the text corresponding to the user's voice may be obtained by combining the obtained text segments.

The electronic device 100 may process the text corresponding to the user's voice to be converted into the text of the language type corresponding to the first ASR model (S940). Specifically, the electronic device 100 may identify whether a text of a language that the first ASR model has not learned from among the text corresponding to the user's voice is included. When some of the texts corresponding to the user's voice are configured in a language of a type not learned by the first ASR model, the electronic device 100 processes the partial texts to be converted into texts of a type of language learned by the first ASR model can do.

As an embodiment, the electronic device 100 may process some text to be converted to match the pronunciation of a type of language learned by the first ASR model. For example, it is assumed that the first ASR model is a model learned based on Korean, and the partial text is 'taxi'. The electronic device 100 may process the partial text 'taxi' to be converted into 'taxi' to match the Korean pronunciation. That is, the electronic device 100 may process some texts composed of English to be converted into Korean, which is a type learned by the first ASR model.

The electronic device 100 may train the first ASR model based on the converted text and the user's voice (S950). That is, the electronic device 100 re-learns the first ASR model based on the input to the text converted into the language of the type learned by the first ASR model and the user's voice corresponding to the converted text to obtain the voice of the first ASR model. recognition performance can be improved.

Meanwhile, when the voice reliability output through the first ASR model is equal to or greater than the threshold value, the electronic device 100 may obtain a text corresponding to the user's voice through the first ASR model ( S960 ). That is, the electronic device 100 may obtain the output text of the first ASR model as text corresponding to the user's voice.

10 is a block diagram illustrating in detail the configuration of the electronic device 100 according to an embodiment of the present disclosure. 10 , the electronic device 100 includes a memory 110 , a microphone 120 , a processor 130 , a communication unit 140 , a display 150 , an input unit 160 , a speaker 170 , and a sensor. (180) may be included. Since the memory 110 , the microphone 120 , and the processor 130 have been described in detail with reference to FIG. 1 , redundant descriptions will be omitted.

The communication unit 140 may communicate with an external device (eg, an external server, etc.). In this case, the communication connection of the communication unit 140 with the external device may include communication through a third device (eg, a repeater, a hub, an access point, a server, or a gateway).

The communication unit 140 may include various communication modules to communicate with an external device. For example, the communication unit 140 may include a wireless communication module, for example, LTE, LTE Advance (LTE-A), 5G (5th generation), CDMA (code division multiple access), WCDMA (wideband CDMA) , a cellular communication module using at least one of universal mobile telecommunications system (UMTS), wireless broadband (WiBro), and global system for mobile communications (GSM).

As another example, the wireless communication module includes, for example, at least one of wireless fidelity (WiFi), Bluetooth, Bluetooth low energy (BLE), Zigbee, radio frequency (RF), or body area network (BAN). can do.

The communication unit 140 may transmit the voice segment to an external server storing the ASR model. For example, the communication unit 140 may transmit, under the control of the processor 130 , a voice segment in Korean to an external server in which an ASR model learned based on Korean is stored. The communication unit 140 may receive a text segment corresponding to the voice segment from an external server in which the ASR model is stored.

Meanwhile, when at least one of the first model and the second model is stored in the external server, the communication unit 140 may transmit the user's voice or each of the voice segments to the first model or the second model. In addition, the communication unit 140 may receive data transmitted from the first model or the second model.

The display 150 may display various information under the control of the processor 130 . In particular, the display 150 may display a UI for setting a basic ASR model by selecting one of a plurality of language types. When one of the plurality of language types is selected through the UI, the processor 130 may set the learned ASR model based on the selected language type as the basic ASR model.

In addition, the display 150 may be implemented as a touch screen together with a touch panel, or may be implemented as a flexible display.

The input unit 160 may receive a user input for controlling the electronic device 100 . In particular, the input unit 160 may include a touch panel for receiving a user touch input using a user's hand or a stylus pen, a button for receiving a user manipulation, and the like. In addition, the input unit 160 may be implemented as another input device (eg, a keyboard, a mouse, a motion input unit, etc.). A user command for selecting one of a plurality of languages may be input through the input unit 160 to set the basic ASR model.

The speaker 170 is configured to output not only various audio data on which various processing tasks such as decoding, amplification, and noise filtering have been performed by the audio processing unit, but also various notification sounds or voice messages.

In particular, the speaker 170 may output a response obtained by processing the text output through the ASR model in the form of a voice. For example, the processor 130 may identify the intent of the text by inputting the obtained text into a Natural Language Understanding (NLU) module. In addition, the processor 130 may obtain a response to the input voice based on the identified intention, and output the acquired response in the form of a voice through the speaker 170 .

Meanwhile, a configuration for outputting audio may be implemented as a speaker, but this is only an exemplary embodiment and may be implemented as an output terminal capable of outputting audio data.

The sensor 180 may detect various state information of the electronic device 100 . For example, the sensor 180 may include a motion sensor (eg, a gyro sensor, an acceleration sensor, etc.) capable of detecting motion information of the electronic device 100, a sensor capable of detecting that a user's voice is input; A sensor capable of detecting location information (eg, a global positioning system (GPS) sensor), a sensor capable of detecting environmental information around the electronic device 100 (eg, a temperature sensor, a humidity sensor, a barometric pressure sensor, etc.), a user It may include a sensor (eg, a camera, a UWB sensor, an IR sensor, a proximity sensor, an optical sensor, etc.) capable of detecting the presence of the . In addition, the sensor 180 may further include an image sensor for photographing the outside of the electronic device 100 .

Meanwhile, various embodiments of the present disclosure may be implemented as software including instructions stored in a machine-readable storage medium readable by a machine (eg, a computer). A device that calls an instruction and can operate according to the called instruction, and may include a server cloud according to the disclosed embodiments. When the instruction is executed by a processor, the processor directly or other configuration under the control of the processor A function corresponding to the command can be performed using the elements.

Instructions may include code generated or executed by a compiler or interpreter. The device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory storage medium' does not include a signal and only means that it is tangible and does not distinguish that data is semi-permanently or temporarily stored in the storage medium. For example, the 'non-transitory storage medium' may include a buffer in which data is temporarily stored.

According to an embodiment, the method according to various embodiments disclosed in the present disclosure may be provided by being included in a computer program product. Computer program products may be traded between sellers and buyers as commodities. The computer program product may be distributed in the form of a machine-readable storage medium (eg, compact disc read only memory (CD-ROM)) or online through an application store (eg, Play Store™). In the case of online distribution, at least a portion of the computer program product (eg, a downloadable app) is at least temporarily stored in a storage medium such as a memory of a server of a manufacturer, a server of an application store, or a relay server, or temporarily can be created with

Each of the components (eg, a module or a program) according to various embodiments may be composed of a singular or a plurality of entities, and some sub-components of the aforementioned sub-components may be omitted, or other sub-components may be various. It may be further included in the embodiment. Alternatively or additionally, some components (eg, a module or a program) may be integrated into a single entity to perform the same or similar functions performed by each corresponding component prior to integration. According to various embodiments, operations performed by a module, program, or other component may be sequentially, parallel, repetitively or heuristically executed, or at least some operations may be executed in a different order, omitted, or other operations may be added. can

Claims

In an electronic device,

MIC;

Memory; and

Obtaining a plurality of speech segments by inputting a first user voice input through the microphone into a first model trained to segment speech based on a word boundary,

inputting the plurality of speech segments into a second model trained to identify a language type of speech to obtain a language type of each of the plurality of speech segments;

inputting each of the plurality of speech segments into an ASR model corresponding to the language type of each of the plurality of speech segments among a plurality of Auto Speech Recognition (ASR) models to obtain a plurality of text segments corresponding to each of the plurality of speech segments; ,

and a processor configured to obtain a text corresponding to the first user's voice based on the plurality of text segments.
According to claim 1,

The first model is

wherein the electronic device is trained to identify at least one word boundary based on waveform data of a user's speech.
According to claim 1,

The processor is

When a language type of each of a first voice segment and a second voice segment among the plurality of voice segments is a first language type and a second language type,

The first speech segment is input to an ASR model corresponding to a first language type among the plurality of ASR models to obtain a text segment corresponding to the first speech segment, and the second speech segment is selected from the plurality of ASR models. An electronic device for obtaining a text segment corresponding to a second voice segment by inputting it into an ASR model corresponding to a second language type.
According to claim 1,

The processor is

An electronic device that trains the second model based on learning data related to the third language type when a user command for requesting that the second model learn to recognize a third language type that has not been learned is input.
According to claim 1,

The processor is

The electronic device obtains the text corresponding to the first user's voice by combining the plurality of text segments according to an input order of the voice segments corresponding to the plurality of text segments.
According to claim 1,

The processor is

The second user's voice input through the microphone is converted into a plurality of frames, and a plurality of speech frames is obtained by filtering non-speech frames among the converted plurality of frames. do,

inputting the plurality of obtained speech frames into the second model to obtain a language type of each of the plurality of speech frames;

inputting each of the plurality of speech frames into an ASR model corresponding to the language type of each of the plurality of speech frames to obtain a text segment corresponding to the plurality of speech frames;

The electronic device obtains the text corresponding to the second user's voice based on the text segments corresponding to the plurality of acquired voice frames.
7. The method of claim 6,

The processor is

Recombining the filtered non-voice frame and the plurality of voice frames according to the order inputted to the microphone,

A text segment corresponding to the first voice frame by inputting a first voice frame among the plurality of voice frames and a first non-voice frame adjacent to the first voice frame into an ASR model corresponding to the language type of the first voice frame An electronic device to obtain
According to claim 1,

The processor is

inputting the first user's voice into a first ASR model among the plurality of ASR models to obtain a voice reliability value;

When the voice reliability value is less than a threshold value, the electronic device updates the first ASR model based on a text corresponding to the first user voice and the user voice.
According to claim 1,

The processor is

inputting the first user's voice into a first ASR model among the plurality of ASR models to obtain a voice reliability value;

When the voice reliability value is less than the threshold value, the text corresponding to the first user's voice is obtained by using at least one ASR model among the plurality of ASR models,

converting text corresponding to the obtained first user voice into text of a language type corresponding to the first ASR model;

The electronic device updates the first ASR model based on the converted text and the first user voice.
9. The method of claim 8,

The processor is

When the voice reliability value is equal to or greater than the threshold value, the electronic device acquires a text corresponding to the first user voice through the first ASR model.
A method for controlling an electronic device, comprising:

acquiring a plurality of speech segments by inputting a first user's speech input to the electronic device into a first model trained to segment speech based on word boundaries;

obtaining a language type of each of the plurality of speech segments by inputting the plurality of speech segments into a second model trained to identify a language type of speech;

inputting each of the plurality of speech segments into an ASR model corresponding to a language type of each of the plurality of speech segments among a plurality of ASR models to obtain a plurality of text segments corresponding to each of the plurality of speech segments; and

and obtaining a text corresponding to the first user's voice based on the plurality of text segments.
12. The method of claim 11,

The first model is

A method according to claim 1, wherein the control method is learned to identify at least one word boundary based on waveform data of a user's speech.
12. The method of claim 11,

Obtaining the plurality of text segments comprises:

When the language type of each of the first and second voice segments among the plurality of voice segments is the first language type and the second language type,

The first speech segment is input to an ASR model corresponding to a first language type among the plurality of ASR models to obtain a text segment corresponding to the first speech segment, and the second speech segment is selected from the plurality of ASR models. and obtaining a text segment corresponding to the second speech segment by inputting the input to the ASR model corresponding to the second language type.
12. The method of claim 11,

When a user command for requesting that the second model learn to recognize a third language type that has not been learned is input, learning the second model based on learning data related to the third language type; further comprising control method.
12. The method of claim 11,

and obtaining a text corresponding to the first user's voice by combining the plurality of text segments according to the order in which the voice segments corresponding to the plurality of text segments are input.