KR20190099988A - Device for voice recognition using end point detection and method thereof - Google Patents

Abstract

The present invention relates to a method for recognizing a voice using a standard speaker model and a method for recognizing a voice. According to one embodiment of the present invention, the method for recognizing a voice using a standard speaker model comprises the steps of: receiving first trigger voice data from a first speaker; generating a standard speaker model from the first trigger voice data by using a speaker recognition algorithm; receiving command voice data from the first speaker; generating a first command voice segment having a start point and an end point from the command voice data by using a voice area detection algorithm; generating a second command voice segment by correcting the start point and/or the end point of the first command voice segment based on the standard speaker model; and performing a voice recognition with respect to the second command voice segment or providing a voice recognition result, wherein the first command voice segment is corrected based on the first trigger voice data so that it is possible to accurately detect a voice section, thereby increasing performance of the device for recognizing a voice.

Description

Speech recognition device using reference speaker model and speech recognition method using same {DEVICE FOR VOICE RECOGNITION USING END POINT DETECTION AND METHOD THEREOF}

The present invention relates to a speech recognition apparatus using a reference speaker model and a speech recognition method using the same, and more particularly, to generate a reference speaker model from trigger speech data to increase the speech recognition performance in a noise environment. The present invention relates to a speech recognition apparatus and a speech recognition method using the same.

Voice is the most common and convenient means of information transfer that humans use. Words expressed by voice play an important role not only as a means of communication between human beings but also as a means of operating machines and devices using human voices. Recently, voice recognition technology is being developed due to the development of computer performance, the development of various media, the development of signal and information processing technology.

Speech recognition technology is a technology that analyzes or understands human speech by converting the spoken voice into an electrical signal using the human voice having a specific frequency by changing the shape of the mouth and the position of the tongue according to the pronunciation. This technology recognizes pronunciation by extracting frequency characteristics of speech signals.

When the voice signal is input, only the voice part spoken by the actual speaker should be detected, and this voice detection part has a great influence on the voice recognition performance. The actual speech recognition environment is very poor due to ambient noise, and thus the noise is often included in the detected section in the environment where most of the speech recognition is performed.

The problem to be solved by the present invention is to detect the exact starting point and end point of the speech spoken by the speaker, and improve the speech recognition rate.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

In a speech recognition method using a reference speaker model according to an embodiment of the present invention for solving the above problems, the method includes: receiving first trigger voice data from a first speaker, using a speaker recognition algorithm, Generating a reference speaker model from the trigger voice data, receiving command voice data from the first speaker, and generating a first command voice segment having a start point and an end point from the command voice data using a voice region detection algorithm Correcting a start point and / or an end point of the first command voice segment based on the reference speaker model to generate a second command voice segment and performing a voice recognition or providing a voice recognition result for the second command voice segment. Include.

Speech recognition apparatus using a reference speaker model according to an embodiment of the present invention for solving the above problems, an input unit for receiving the trigger voice data and command voice data, the trigger voice data is analyzed to generate a reference speaker model A speaker model generator to analyze the command voice data to generate a first command voice segment having a start point and an end point, and correct the start point and / or the end point of the first command voice segment based on the reference speaker model And a voice segment corrector for generating a second command voice segment.

Specific details of other embodiments are included in the detailed description and drawings.

The present invention can detect the correct speech section and improve the speech recognition rate.

The effects according to the present invention are not limited by the contents exemplified above, and more various effects are included in the present specification.

1A is a block diagram illustrating a configuration of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.
1B is a block diagram schematically illustrating a configuration of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention driven by a processor.
2 is a flowchart illustrating a speech recognition method using a reference speaker model according to an embodiment of the present invention.
3A to 3G are exemplary diagrams for describing a speech recognition method using a reference speaker model according to an embodiment of the present invention.
4 is a block diagram schematically illustrating a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.
5A and 5B are exemplary diagrams for describing a process of storing a second reference speaker model in a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.
6A and 6B are exemplary diagrams for describing a process of storing a second reference speaker model in a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

The following techniques are wired or code division multiple access (CDMA), frequency division multiple access (FDMA), time division multiple access (TDMA), orthogonal frequency division multiple access (OFDMA), and single carrier-frequency division multiple access (SC-FDMA) May be used in various multiple access schemes, such as CDMA may be implemented with a radio technology such as Universal Terrestrial Radio Access (UTRA) or CDMA2000.

Advantages and features of the present invention and methods for achieving them will be apparent with reference to the embodiments described below in detail with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in various forms, and only the present embodiments are intended to complete the disclosure of the present invention, and the general knowledge in the art to which the present invention pertains. It is provided to fully convey the scope of the invention to those skilled in the art, and the present invention is defined only by the scope of the claims.

Also, the first, second, etc. are used to describe various components, but these components are not limited by these terms. These terms are only used to distinguish one component from another. Therefore, of course, the first component mentioned below may be a second component within the technical spirit of the present invention.

Like reference numerals refer to like elements throughout.

Each of the features of the various embodiments of the present invention may be combined or combined with each other in part or in whole, various technically interlocking and driving as can be understood by those skilled in the art, each of the embodiments may be implemented independently of each other It may be possible to carry out together in an association.

Hereinafter, various embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1A is a block diagram illustrating a configuration of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

Hereinafter, the speech recognition apparatus 100 includes any device having various kinds of speech recognition functions not described above in the concept of including a mobile phone, a tablet, a PC, a wearable device, a laptop, a stick PC, a PMP, and an MP3 player. .

The speech recognition apparatus 100 includes an input unit 110, a processor 120, a memory unit 130, and an output unit 140.

The input unit 110 receives voice data spoken by an arbitrary speaker. The input unit 110 may be, for example, a microphone included in the voice recognition apparatus 100, and the present invention is not limited thereto.

The input unit 110 may further include a module such as a filter to remove noise included in the input voice data.

The processor 120 processes various data for performing a speech recognition method using the speech recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention.

In detail, the processor 120 generates a speaker model by analyzing voice data received from the input unit 110 using a speaker recognition algorithm. The processor 120 generates a voice segment by analyzing voice data received from the input unit 110 using a voice region detection algorithm.

In addition, the processor 120 compares the speaker model, corrects the start point (SP) and / or end point (EP) of the speech segment to generate a new speech segment, and performs a speech recognition or a speech recognition result for the new speech segment. to provide.

The voice recognition apparatus 100 may correspond to at least one or more processors 120 or include at least one or more processors 120. Therefore, the speech recognition apparatus 100 may be driven in a form included in a hardware device such as a microprocessor or a general purpose computer system, or may be driven by a separate device.

The memory unit 130 may store data for the operation of the speech recognition apparatus 100, and may temporarily store input / output data. For example, the memory unit 130 may store software code that may be executed by at least one module of the input unit 110, the processor 120, and the output unit 140. In addition, the memory unit 130 may store all kinds of software codes for operating the speech recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention.

For example, the memory unit 130 may be executed by one or more processors 120 and may be stored in a computer readable medium including instructions for causing the one or more processors 120 to perform the following operations. Operations for a computer program to improve speech recognition performance may be stored. The operations include: receiving trigger voice data and command voice data; Generating a reference speaker model from trigger speech data using a speaker recognition algorithm; Generating a speech segment from the speech data using the speech region detection algorithm; Analyzing the similarity between the reference speaker model and the speech segment; Ensuring a start point (SP) and / or an end point (EP) of the speech segment based on the similarity; And performing a speech recognition or providing a speech recognition result for the speech segment whose start point SP and / or the end point EP are corrected.

The memory unit 130 may store various information for executing a program according to an embodiment of the present invention. For example, the memory unit 130 may store a voice region detection algorithm and a speaker recognition algorithm. In addition, the memory unit 130 may store at least one of a speaker model and a language model. Information that the memory unit 130 can store is not limited to the above descriptions. In addition, the information as described above may be stored for each voice segment and the scope of the present invention is not limited thereto.

Meanwhile, the memory unit 130 may store various data necessary for providing the speech recognition apparatus 100 and may provide requested data according to requests of other components.

The memory unit 130 may include a flash memory type, a hard disk type, a multimedia card micro type, and a card type memory unit (eg, SD or XD memory). Etc.), RAM (Random Access Memory, RAM), Static Random Access Memory (SRAM), Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Programmable Read-Only Memory (PROM) It may include a storage medium of at least one type of a magnetic memory unit, a magnetic disk, an optical disk. The text input program for touch input may operate in connection with a web storage which performs a storage function of the memory unit 130 on the Internet.

To this end, the speech recognition apparatus 100 may include a communication unit. The communication unit may include a wired / wireless communication module for network connection.

Wireless Internet technologies may include Wireless LAN (Wi-Fi), Wireless Broadband (Wibro), World Interoperability for Microwave Access (Wimax), High Speed Downlink Packet Access (HSDPA), and the like.

Wired Internet technologies may include Digital Subscriber Line (XDSL), Fibers to the home (FTTH), Power Line Communication (PLC), and the like.

In addition, the communication unit may include a short range communication module to transmit and receive data with an electronic device including the short range communication module. As a short range communication technology, Bluetooth, Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra-Wide band (UWB), and ZigBee may be used. The aforementioned communication techniques are merely examples, and the present invention is not limited thereto.

In this case, the data transmitted and / or received through the communication unit may be stored in the memory unit 130 or transmitted to other devices in a short distance through the short range communication module.

Meanwhile, the speech recognition apparatus 100 may further include components not shown. For example, the voice recognition apparatus 100 may further include a display unit. The display unit may display all information that may be displayed on the voice recognition apparatus 100. Such displays include liquid crystal displays (LCDs), thin film transistor-liquid crystal displays (TFT LCDs), organic light-emitting diodes (OLEDs), flexible displays, It may include at least one of a 3D display.

The various embodiments described herein can be implemented in a recording medium or storage medium that can be read by a computer or similar device using, for example, software, hardware or a combination thereof.

For example, according to a hardware implementation, embodiments described herein may include application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), FPGAs ( It may be implemented using at least one of field programmable gate arrays, processors, controllers, micro-controllers, microprocessors, or electrical units for performing other functions. In this case, the embodiments described herein may be implemented by the controller itself.

As another example, in a software implementation, embodiments such as the procedures and functions described herein may be implemented in separate software modules. Each of the software modules may perform one or more functions and operations described herein. Software code may be implemented in software applications written in the appropriate programming languages. The software code may be stored in the memory unit 130 and executed by the processor 120.

Meanwhile, the output unit 140 outputs the voice recognition result performed by the processor 120 to the output device. In detail, the output unit 140 may output a voice recognition result through the image output device and / or the voice output device. For example, the image output device of the output unit 140 may be a display unit included in the voice recognition device 100, and may output a voice recognition result as an image through the image output device. In addition, the voice output device of the output unit 140 may be a speaker included in the voice recognition device 100, and may output a voice recognition result as a voice through the voice output device.

Hereinafter, for convenience of description, the voice recognition device 100 is a home appliance, and the voice recognition device 100 is stored in a storage medium of the home appliance in the form of an application and driven by the processor 120 of the home appliance. It is assumed that the description will be made.

1B is a block diagram schematically illustrating a configuration of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention driven by a processor.

Referring to FIG. 1B, the speech recognition apparatus 100 using the reference speaker model according to an exemplary embodiment of the present invention may include a speaker model generator 121, a voice segment generator 122, a similarity analyzer 123, and a voice. And a segment correcting unit 124.

The speaker model generator 121 generates a reference speaker model from trigger voice data using a speaker recognition algorithm. In detail, the speaker model generator 121 extracts a feature of the trigger voice data received from the input unit 110 and generates a reference speaker model including the feature of the trigger voice data.

The trigger voice data is a command for operating the voice recognition apparatus 100. For example, when an arbitrary user speaks an arbitrary command to the speech recognition apparatus 100, the speech recognition apparatus 100 may be turned on. The trigger voice data may be variously modified according to the voice recognition apparatus 100.

The voice segment generator 122 analyzes the command voice data using a voice region detection algorithm and generates a first command voice segment SM1 having a start point SP and an end point EP.

In this case, the command voice data is voice data input to the voice recognition device 100 that is in a state of receiving trigger voice data. Specifically, the command voice data contains content for which a user needs voice recognition. For example, when a user inputs command voice data to the voice recognition apparatus 100, the voice recognition apparatus 100 may receive the command voice data and perform a voice recognition operation.

The first command voice segment SM1 divides a section having a start point SP and an end point EP from the command voice data. Specifically, the command voice data may include only contents spoken by an arbitrary user, but may further include noise such as noise or speech contents of another user depending on the surrounding environment. In this case, the voice segment generator 122 extracts a section in which any user actually speaks, except for noise, from the command voice data to generate the first command voice segment SM1.

Meanwhile, the speaker model generator 121 generates a plurality of command speaker models from the first command voice segment SM1 in addition to the reference speaker model. In detail, the first command voice segment SM1 is divided into a plurality of sections, and the speaker model generator 121 extracts a feature from each of the plurality of sections, and generates a command speaker model including the features of each section. . In this case, the plurality of sections are an external area or an internal area of the first command voice segment SM1 based on the start point SP and / or the end point EP of the first command voice segment SM1.

The similarity analyzer 123 analyzes the similarity between the reference speaker model generated from the trigger voice data and the command speaker model generated from the command voice data. In detail, the similarity analyzer 123 analyzes the similarity between the reference speaker model and the command speaker model by comparing the characteristics of the reference speaker model and the feature of the command speaker model.

In this case, the similarity analyzer 123 may determine the identity of the reference speaker model and the command speaker model from the similarity. For example, when the similarity exceeds the threshold, the similarity analyzer 123 may determine the speaker of the reference speaker model and the speaker of the command speaker model as the same speaker. On the other hand, when the similarity is less than or equal to the threshold, the similarity analyzer 123 may determine the speaker of the reference speaker model and the speaker of the command speaker model as different speakers. The threshold may be a predetermined value, and the threshold may be variously set.

The voice segment correcting unit 124 may start or stop an end point (SP) and / or an end point (EP) of the first command voice segment SM1 according to the similarity and the sameness between the reference speaker model and the command speaker model determined by the similarity analyzer 123. ) Is generated to generate a second command voice segment SM2. In detail, when the similarity analyzer 123 determines that the command speaker model and the reference speaker model are not the same in one of a plurality of sections of the first command voice segment SM1, the voice segment corrector 124 may determine whether the command speaker model is the same. The start point SP and / or the end point EP of the first command voice segment SM1 may be corrected so that the one section is excluded from the first command voice segment SM1. Therefore, according to the result of the similarity analyzer 123, the start point SP 'and / or the end point EP' of the second command voice segment SM2 is defined as the start point SP and / or the first command voice segment SM1. Or it may be the same as or different from the end point EP.

On the other hand, each of the components of the speech recognition apparatus 100 is only shown as a separate configuration for convenience of description, it may be implemented in one configuration or one configuration may be divided into two or more configurations according to the implementation method.

Hereinafter, a speech recognition method using the speech recognition apparatus 100 based on the description of the speech recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention described above, and referring to FIGS. 2 to 3G. This will be described.

2 is a flowchart illustrating a speech recognition method using a reference speaker model according to an embodiment of the present invention. 3A is a schematic diagram illustrating a process of receiving first trigger voice data from an input unit in a voice recognition apparatus using a reference speaker model according to an embodiment of the present invention. 3B is a schematic diagram illustrating a process of generating a reference speaker model from first trigger speech data using a speaker recognition algorithm in a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention. 3C is a schematic diagram illustrating a process of receiving command voice data from an input unit in a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention. 3D is a schematic diagram illustrating a process of generating a first command voice segment from command voice data using a voice region detection algorithm in a voice recognition apparatus using a reference speaker model according to an embodiment of the present invention. 3E is a schematic diagram illustrating a process of dividing a first command speech segment into a plurality of sections in a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention. FIG. 3F is a schematic diagram illustrating a process of generating a second command voice segment by correcting a first command voice segment based on the reference speaker model in the voice recognition apparatus using the reference speaker model according to an embodiment of the present invention. 3G is a schematic diagram illustrating a process of providing a speech recognition result for a second command speech segment in a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

First, the voice recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention receives the first trigger voice data TV1 from the first speaker 150 through the input unit 110 (S110).

Referring to FIG. 3A, the voice recognition apparatus 100 receives first trigger voice data TV1 from the first speaker 150. In this case, although the first trigger voice data TV1 is shown as “hello” in FIG. 3A, the first trigger voice data TV1 may be variously modified, but is not limited thereto.

Subsequently, a reference speaker model is generated from the first trigger voice data TV1 using a speaker recognition algorithm (S120).

Referring to FIG. 3B, the speech recognition apparatus 100 extracts a feature of only the first speaker 150 from the first triggered speech data TV1 received from the first speaker 150 using a speaker recognition algorithm. For example, the speech recognition apparatus 100 recognizes the pronunciation, intonation, stress, etc. of the first speaker 150 from the first triggered voice data TV1, and includes a reference including the characteristics of the first speaker 150. Create a speaker model.

In this case, the reference speaker model is a speaker model serving as a reference for correcting the start point SP and the end point EP of the first command voice segment SM1 generated from the command voice data OV.

Meanwhile, the speech recognition apparatus 100 compares the reference speaker model generated from the first triggered speech data TV1 with a plurality of speaker models MO1, MO2,..., MOn stored in the memory unit 130. The model may be updated or the reference speaker model may be newly stored in the memory unit 130. Details of updating and storing the speaker model will be described later with reference to FIGS. 4 to 6B.

Subsequently, command voice data OV is received from the first speaker 150 (S130).

Referring to FIG. 3C, the voice recognition apparatus 100 receives the first trigger voice data TV1 spoken from the first speaker 150 and then receives the command voice data OV. In this case, although the command voice data OV is shown as “tell the weather tomorrow” in FIG. 3C, the command voice data OV may be variously modified, but is not limited thereto.

Subsequently, a first command voice segment SM1 having a start point SP and an end point EP is generated from the command voice data OV using the voice region detection algorithm (S140).

Referring to FIG. 3D, the speech segment generator 122 of the speech recognition apparatus 100 detects a section expected from the command speech data OV as the actual speech section of the first speaker 150 using a speech domain algorithm. do. The voice segment generator 122 generates a first command voice segment SM1 including information about a start point SP and an end point EP of a section that is expected to be an actual speech section of the first speaker 150. .

When the voice recognition apparatus 100 receives the command voice data OV from the first speaker 150, the voice recognition device 100 may receive noise in addition to the voice of the first speaker 150. The voice segment generator 122 detects only a section of the command voice data OV including noise that is expected to be the voice of the first speaker 150 and generates the first command voice segment SM1.

The remaining section of the command voice data OV not included in the first command voice segment SM1 may be a section unrelated to the voice of the first speaker 150.

Next, the second command voice segment SM2 is generated by correcting the start point SP and / or the end point EP of the first command voice segment SM1 based on the reference speaker model (S150).

Referring to FIG. 3E, the speaker model generator 121 generates a plurality of command speaker models from the first command voice segment SM1.

First, the internal area or the external area of the first command voice segment SM1 is divided into a plurality of sections S1, S2, S3, based on the start point SP and / or the end point EP of the first command voice segment SM1. S4) can be divided.

For example, the plurality of sections S1, S2, S3, and S4 may include a first section S1 between a starting point SP of the first command voice segment SM1 and an outer region of the first command voice segment SM1. , The second section S2 between the starting point SP of the first command voice segment SM1 and the internal region of the first command voice segment SM1, and the second point S2 from the end point EP of the first command voice segment SM1. The third section S3 between the inner regions of the first command speech segment SM1, and the fourth section between the outer region of the first command speech segment SM1 from the end point EP of the first command speech segment SM1 ( S4) can be divided.

The speaker model generator 121 extracts a feature of the command voice data OV in each of the first to fourth sections S1, S1, and S1. A second command speaker model is generated in the second section S2, a third command speaker model in the third section S3, and a fourth command speaker model in the fourth section S4.

Meanwhile, in FIG. 3E, although the lengths of the first section S1 to the fourth section S4 are the same, the lengths of the first section S1 to the fourth section S4 may be variously modified. It is not limited.

In addition, although FIG. 3E illustrates that the first command voice segment SM1 is divided into a first section S1 to a fourth section S4, the plurality of sections S1, S2, S3, and S4 may include a first command. The voice segment SM1 may be divided in various ways based on the start point SP or the end point EP, but is not limited thereto.

Subsequently, when generation of the first command speaker model to the fourth command speaker model of each of the first to fourth sections S4 to S4 is completed in the speaker model generator 121, the similarity analysis unit 123 generates the first instruction speaker model. The similarity between each of the first to fourth command speaker models and the reference speaker model is analyzed. The similarity analysis unit 123 determines whether each of the first to fourth command speaker models is identical to the reference speaker model based on the similarity.

For example, the similarity analyzer 123 may analyze the similarity between the first command speaker model and the reference speaker model in the first section S1. As a result of the analysis, when the similarity between the first command speaker model and the reference speaker model exceeds a threshold, the similarity analyzer 123 may finally determine that the speaker of the first command speaker model and the reference speaker model is the same.

Similarly, the similarity analyzer 123 analyzes the similarity between the second command speaker model and the reference speaker model in the second section S2 to finally determine that the speaker of the second command speaker model and the reference speaker model are the same. Can be.

The similarity analyzer 123 may finally determine that the speaker of the third command speaker model and the reference speaker model are the same by analyzing the similarity between the third command speaker model and the reference speaker model in the third section S3. .

The similarity analyzer 123 may analyze the similarity between the fourth command speaker model and the reference speaker model in the fourth section S4. As a result of the analysis, since the similarity between the fourth command speaker model and the reference speaker model is less than or equal to the threshold value, the similarity analyzer 123 may determine that the speaker of the fourth command speaker model and the reference speaker model are not the same.

Next, referring to FIG. 3F, the voice segment corrector 124 corrects the start point SP and / or the end point EP of the first command voice segment SM1 according to the result of the similarity analyzer 123.

In detail, the similarity analyzer 123 may reference the first command speaker model of the first section S1, the second command speaker model of the second section S2, and the third command speaker model of the third section S3. It is determined that it is the same as the speaker model, and it is determined that the fourth command speaker model in the fourth section S4 is not the same as the reference speaker model.

Accordingly, the voice segment correcting unit 124 includes the first section S1 to the third section S3 that are determined to be the same speaker model as the reference speaker model in the first command voice segment SM1, and includes the reference speaker model. The fourth section S4 determined as not being the same speaker model is excluded from being included in the first command voice segment SM1.

That is, when the command speaker model and the reference speaker model are the same, the voice segment corrector 124 may start the first command voice segment SM1 such that a section corresponding to the command speaker model is included in the first command voice segment SM1. Calibrate (SP) and / or end point (EP).

In addition, when the command speaker model and the reference speaker model are not the same, the voice segment corrector 124 may include the first command voice segment SM1 such that a section corresponding to the command speaker model is not included in the first command voice segment SM1. Correct the start point SP and / or the end point EP of).

Therefore, the start point SP and the end point EP of the first command voice segment SM1 are corrected by the voice segment corrector 124 to generate the second command voice segment SM2.

Finally, the voice recognition performed on the second command voice segment SM2 or the voice recognition result is provided (S160).

Referring to FIG. 3G, the speech recognition apparatus 100 performs speech recognition on the second command speech segment SM2 whose start point SP 'and the end point EP' are corrected based on the reference speaker model.

The second command voice segment SM2 includes command voice data OV for “tell me the weather for tomorrow” of the first speaker 150, and the voice recognition apparatus 100 includes the second command voice segment ( SM2) is analyzed.

The voice recognition apparatus 100 analyzes the second command voice segment SM2 and informs the first speaker 150 of the weather for tomorrow. For example, the speech recognition apparatus 100 may output a speech recognition result as ““ the weather is cloudy tomorrow ”.” However, the voice recognition apparatus 100 may output a voice recognition result as an image through a display unit or the like as well as the voice, but is not limited thereto.

The voice recognition method and the voice recognition apparatus 100 using the reference speaker model according to an exemplary embodiment of the present invention use the first trigger voice data TV1 to start point SP and end point of the first command voice segment SM1. Correct (EP). In detail, the input unit 110 of the speech recognition apparatus 100 receives the first trigger voice data TV1 from the first speaker 150, and the speaker model generator 121 receives the first trigger voice data TV1. Create a reference speaker model from. Subsequently, the input unit 110 of the speech recognition apparatus 100 receives the command voice data OV from the first speaker 150, and the voice segment generator 122 receives the first speaker (from the command voice data OV). The first command voice segment SM1 is generated by setting the start point SP and the end point EP of the section expected as the actual utterance section 150. Subsequently, the similarity analyzer 123 of the speech recognition apparatus 100 compares the command speaker model and the reference speaker model in a plurality of sections of the first command voice segment SM1 to determine similarity and identity. Lastly, according to the result of the similarity analyzer 123, the voice segment correcting unit 124 of the speech recognition apparatus 100 determines that the section corresponding to the command speaker model determined to be the same as the reference speaker model is the first command voice segment. The section corresponding to the command speaker model determined to be not the same as the reference speaker model is not included in the first command voice segment SM1. Therefore, the start point SP and / or the end point EP of the first command voice segment SM1 is corrected by the voice segment corrector 124, and the second command voice segment SM2 is generated.

In this case, the command voice data OV may include not only the voice of the first speaker 150 but also ambient noise, voices of other speakers, and the like. Therefore, the first command voice segment SM1 may or may not include a section in which the first speaker 150 actually uttered. Accordingly, the speech recognition method and the speech recognition apparatus 100 using the reference speaker model according to the embodiment of the present invention may increase the accuracy of detecting the start point SP and the end point EP of the command speech segment. The first command voice segment SM1 may be corrected using the first trigger voice data TV1 which is a command for turning on 100. The speech recognition apparatus 100 generates a reference speaker model obtained by extracting the feature of the first speaker 150 from the first trigger voice data TV1 and compares each section of the first command voice segment SM1 with the reference speaker model. Analyze the similarity. At this time, the start point SP and / or the end point EP are corrected such that only a section corresponding to the command speaker model having characteristics similar to the reference speaker model is included in the first command voice segment SM1. Therefore, the voice recognition method and the voice recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention can accurately detect the start point and the end point of the command voice segment using a criterion called trigger voice data, and recognize the voice. Performance can be improved.

4 is a block diagram schematically illustrating a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

Referring to FIG. 4, as described above with reference to FIG. 1A, the memory unit stores a plurality of speaker models.

The plurality of speaker models MO1, MO2, ..., MOn each include a feature for one speaker. For example, the first speaker model MO1 is a speaker model that analyzes voice data received from the first speaker 150 and stores characteristics of the first speaker 150, and the second speaker model MO2. Is a speaker model that analyzes voice data received from the second speaker and stores the characteristics of the second speaker. Therefore, the characteristics of different speakers are stored in each of the speaker models MO1, MO2, ..., MOn.

On the other hand, the speech recognition apparatus 100 compares the speaker model generated from the trigger voice data of a specific speaker with the speaker models MO1, MO2, ..., MOn stored in the memory unit 130, and compares the speaker model. The reference speaker model may be updated or newly stored in the memory unit 130.

5A and 5B are exemplary diagrams for describing a process of storing a second reference speaker model in a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

Referring to FIG. 5A, the voice recognition apparatus 100 receives second trigger voice data from a second speaker, analyzes the second trigger voice data, and generates a second reference speaker model TMO2.

In addition, the speech recognition apparatus 100 compares each of the plurality of speaker models MO1, MO2,..., MOn stored in the memory unit 130 with the second reference speaker model TMO2 to analyze the similarity, and determine whether or not there is an equality. To judge.

When there is a speaker model that is determined to be the same as the second reference speaker model TMO2 among the plurality of speaker models MO1, MO2,..., MOn, the speech recognition apparatus 100 determines that the second reference speaker model TMO2 is the same. The speaker model can be updated to be included in the speaker model.

For example, referring to FIG. 5B, when it is determined that the second reference speaker model TMO2 is the same as the first speaker model MO1, the speech recognition apparatus 100 may determine that the first speaker model MO1 is the second speaker. The first speaker model MO1 may be updated to include information about the reference speaker model TMO2. Therefore, the updated first speaker model MO1 ′ is stored with information about the second reference speaker model TMO2 together with information about the first speaker 150.

On the other hand, when the second reference speaker model TMO2 is not the same as the plurality of speaker models MO1, MO2,..., MOn stored in the memory unit 130, the second reference speaker model TMO2 may be newly stored. have. Detailed description thereof will be described with reference to FIGS. 6A and 6B.

6A and 6B are exemplary diagrams for describing a process of storing a second reference speaker model in a memory unit of a speech recognition apparatus using a reference speaker model according to an embodiment of the present invention.

Referring to FIG. 6A, the voice recognition apparatus 100 receives second trigger voice data from a second speaker, analyzes the second trigger voice data, and generates a second reference speaker model TMO2.

In addition, the speech recognition apparatus 100 compares each of the plurality of speaker models MO1, MO2,..., MOn stored in the memory unit 130 with the second reference speaker model TMO2 to analyze the similarity, and determine whether or not there is an equality. To judge.

At this time, when the similarity with the second reference speaker model TMO2 is high and there is no speaker model determined in the memory unit 130, the second reference speaker model TMO2 is used as the new speaker model in the memory unit 130. Save it.

For example, referring to FIG. 6B, the speech recognition apparatus 100 stores the second reference speaker model TMO2 as the n + 1 speaker model MOn + 1 in the memory unit 130. Accordingly, the memory unit 130 includes a plurality of speaker models MO1, MO2,..., MOn stored previously and an n + 1 speaker model MOn + 1 newly stored from the second reference speaker model TMO2. do.

Accordingly, the voice recognition method and the voice recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention store the reference speaker model generated from the trigger voice data in the memory unit 130. In detail, the plurality of speaker models MO1, MO2,..., MOn are stored in the memory unit 130. The speech recognition apparatus 100 analyzes the similarity between each of the speaker models MO1, MO2,..., MOn stored in the memory unit 130 and the reference speaker model, and determines the sameness. When the speaker model determined to be the same as the reference speaker model is detected, the speaker model is updated by adding information of the reference speaker model to the speaker model stored in the memory unit 130. On the other hand, if the speaker model determined to be the same as the reference speaker model is not detected, the reference speaker model is newly stored in the memory unit 130.

Accordingly, the voice recognition method and the voice recognition apparatus 100 using the reference speaker model according to an embodiment of the present invention store information about the reference speaker model in the memory unit 130. The speaker model newly stored in the memory unit 130 may be updated when speech recognition of the speaker related to the speaker model is performed later, and the speech recognition capability of the speech recognition apparatus 100 may be improved. In addition, the continuously updated speaker model may include more detailed information and features of a specific speaker, and then the voice recognition capability of the voice recognition apparatus 100 may also be improved.

In this specification, each block or each step may represent a portion of a module, segment or code containing one or more executable instructions for executing a specified logical function (s). It should also be noted that in some alternative embodiments the functions noted in the blocks or steps may occur out of order. For example, the two blocks or steps shown in succession may in fact be executed substantially concurrently or the blocks or steps may sometimes be performed in the reverse order, depending on the functionality involved.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be implemented directly in hardware, a software module or a combination of the two executed by a processor. The software module may reside in a RAM memory section, flash memory section, ROM memory section, EPROM memory section, EEPROM memory section, register, hard disk, removable disk, CD-ROM or any other form of storage medium known in the art. have. An exemplary storage medium is coupled to the processor, which can read information from and write information to the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an application specific integrated circuit. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.

Although the embodiments of the present invention have been described in more detail with reference to the accompanying drawings, the present invention is not necessarily limited to these embodiments, and various modifications can be made without departing from the spirit of the present invention. . Therefore, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention but to describe the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. Therefore, it should be understood that the embodiments described above are exemplary in all respects and not restrictive. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the equivalent scope should be interpreted as being included in the scope of the present invention.

100: speech recognition device
110: input unit
120: processor
121: speaker model generator
122: voice segment generator
123: Similarity Analysis
124: voice segment correction unit
130: memory unit
140: output unit
150: first speaker
TV1: first trigger voice data
OV: command voice data
SM1: first command voice segment
SM2: second command voice segment
SP, SP ': starting point
EP, EP ': Endpoint
S1: first section
S2: second section
S3: third section
S4: fourth section
MO1, MO1 ': first speaker model
MO2: Second Speaker Model
MOn: Control Speaker Model
MOn + 1: n + 1 speaker model
TMO2: Second Reference Speaker Model

Claims (14)

Receiving first trigger voice data from a first speaker;
Generating a reference speaker model from the first triggered speech data using a speaker recognition algorithm;
Receiving command voice data from the first speaker;
Generating a first command voice segment having a start point and an end point from the command voice data using a voice region detection algorithm;
Generating a second command speech segment by correcting the starting point and / or the end point of the first command speech segment based on the reference speaker model; And
And performing a voice recognition or providing a voice recognition result for the second command voice segment. The method of claim 1,
Generating the second command voice segment may include:
Dividing the first command voice segment into a plurality of sections;
Generating a command speaker model in each of the plurality of sections using the speaker recognition algorithm; And
And determining the identity of the command speaker model and the reference speaker model. The method of claim 2,
If the command speaker model and the reference speaker model are the same,
The generating of the second command voice segment may include correcting the start point and / or the end point of the first command voice segment such that the section corresponding to the command speaker model is included in the first command voice segment. , Speech recognition method using reference speaker model. The method of claim 2,
If the command speaker model and the reference speaker model are not the same,
The generating of the second command voice segment may include correcting the start point and / or the end point of the first command voice segment such that the section corresponding to the command speaker model is excluded from the first command voice segment. , Speech recognition method using reference speaker model. The method of claim 2,
The plurality of sections,
A first section between the starting point of the first command speech segment and an outer region of the first command speech segment;
A second section between the starting point of the first command speech segment and an interior region of the first command speech segment;
A third section between the end point of the first command speech segment and an interior region of the first command speech segment; And
And at least one of a fourth section between the end point of the first command speech segment and an outer area of the first command speech segment. The method of claim 1,
Storing the reference speaker model in a memory unit;
Receiving second trigger voice data from a second speaker; And
And generating a second reference speaker model from the second trigger speech data using the speaker recognition algorithm. The method of claim 6,
If the second reference speaker model and the reference speaker model are the same,
And updating the reference speaker model based on the second reference speaker model. The method of claim 6,
If the second reference speaker model and the reference speaker model are not the same,
And storing the second reference speaker model in a memory unit. An input unit to receive trigger voice data and command voice data;
A speaker model generator configured to generate the reference speaker model by analyzing the trigger voice data;
A voice segment generator configured to analyze the command voice data to generate a first command voice segment having a start point and an end point; And
And a speech segment corrector configured to correct the starting point and / or the end point of the first command speech segment to generate a second command speech segment based on the reference speaker model. The method of claim 9,
The first command voice segment includes a plurality of sections,
And the speaker model generator generates a command speaker model in each of the plurality of sections. The method of claim 10,
Further comprising a similarity analysis unit for analyzing the similarity between the reference speaker model and the command speaker model,
When the similarity of the reference speaker model and the command speaker model exceeds a threshold,
The speech segment correcting unit corrects the start point and / or the end point of the first command speech segment such that the section corresponding to the command speaker model is included in the first command speech segment. Device. The method of claim 11,
When the similarity between the reference speaker model and the command speaker model is less than or equal to the threshold,
The speech segment corrector corrects the start point and / or the end point of the first command speech segment such that the section corresponding to the command speaker model is excluded from the first command speech segment. Device. The method of claim 10,
The plurality of sections,
And an outer region or an inner region of the first command speech segment based on the start point or the end point of the first command speech segment. The method of claim 9,
Further comprising a memory unit for storing a plurality of speaker models,
The speaker model generating unit analyzes the trigger voice data, and when the trigger voice data corresponds to one speaker model of the plurality of speaker models stored in the memory unit, analyzes the trigger voice data to generate the one speaker model. And update the updated one speaker model as the reference speaker model.

Images