KR102429891B1 - Voice recognition device and method of operating the same - Google Patents

Abstract

A speech recognition apparatus includes: a storage medium for storing weights for an arrangement of predetermined phonemes; a voice pre-processing unit for generating a voice signal according to a voice sensed by a user; a first speech recognition unit configured to generate likelihood values corresponding to the candidate phonemes while discriminating the candidate phonemes from the speech signal; a second speech recognition unit configured to adjust at least a portion of likelihood values corresponding to candidate phonemes according to weights for an arrangement of predetermined phonemes; and a speech estimator configured to estimate an arrangement of phonemes corresponding to the user's speech from the candidate phonemes based on the adjusted likelihood values.

Description

Voice recognition device and method of operation thereof

The present invention relates to an electronic device, and more particularly, to a voice recognition device capable of recognizing a user's voice and an operating method thereof.

Recently, voice recognition technology has been applied and commercialized in real life beyond the level of laboratory demonstrations. The current speech recognition system shows relatively good performance under a set environment, but shows a degraded recognition performance under various environments. This is because real environments in which voice recognition is performed involve factors that degrade recognition performance, such as ambient noise, utterance distance, microphone characteristics, and speaker differences.

These elements may pollute the sensed speech signal to cause changes and thus change the corresponding feature vector. This means that the statistical properties of the feature vector are changed. For example, white noise may reduce the dynamic range (or variance) of a feature vector, such as a cepstrum, that represents the envelope information of a spectrum.

The above description is only for helping the understanding of the background of the technical spirit of the present invention, and therefore it cannot be understood as the content corresponding to the prior art known to those skilled in the art.

Currently, speech recognition for a short language such as a single word depends on an acoustic model based on a feature vector, and additional correction for a word recognized based on the acoustic model is not considered. For example, in a device activated using a command, since the command is made up of a short language such as “high”, additional correction using a language model that has learned the arrangement of a plurality of words is not considered. Accordingly, the recognition rate for the corresponding command is low.

An object of the present invention is to provide a voice recognition apparatus capable of recognizing a voice with improved reliability. For example, the voice recognition apparatus may recognize phonemes from the user's voice and provide weights for adjusting likelihood values for the recognized phonemes. Accordingly, the voice recognition apparatus may perform voice recognition with relatively high reliability even for a language of a relatively short length, such as a command.

A voice recognition apparatus according to an embodiment of the present invention includes: a storage medium for storing weights for an arrangement of predetermined phonemes; a voice pre-processing unit for generating a voice signal according to a voice sensed by a user; a first speech recognition unit configured to generate likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal; a second speech recognition unit configured to adjust at least a portion of the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the predetermined phonemes; and a speech estimator configured to estimate an arrangement of phonemes corresponding to the user's speech from the candidate phonemes based on the adjusted likelihood values.

The second speech recognition unit selects from among the candidate phonemes that match at least a part of the arrangement of the predetermined phonemes, and assigns the selected candidate phonemes to the selected candidate phonemes according to weights corresponding to the matched arrangement of the predetermined phonemes. may be configured to adjust the corresponding likelihood values.

The voice signal may include feature vector values obtained by converting the sensed voice into a frequency domain.

The first speech recognition unit may be configured to generate the likelihood values corresponding to the candidate phonemes based on a hidden Markov model.

The speech estimator may be configured to generate a trigger signal when the arrangement of the estimated phonemes matches a predetermined command.

The voice recognition apparatus may further include a function block configured to activate a predetermined operation in response to the trigger signal.

A voice recognition apparatus according to another embodiment of the present invention includes an acoustic sensor configured to detect a voice from a user; a processor configured to determine whether the sensed voice corresponds to a command; and a storage medium for storing weights for an arrangement of predetermined phonemes, wherein the processor is configured to: generate a voice signal according to the sensed voice; generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal; adjust at least some of the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the predetermined phonemes; and estimate an arrangement of phonemes corresponding to the voice of the user from the candidate phonemes based on the adjusted likelihood values.

The processor selects among the candidate phonemes that match at least a portion of the arrangement of the given phonemes, and a likelihood corresponding to the selected candidate phonemes according to weights corresponding to the matched arrangement of the given phonemes. may be configured to adjust the values.

The voice signal may include feature vector values obtained by converting the sensed voice into a frequency domain.

The processor may be configured to generate the likelihood values corresponding to the candidate phonemes based on a hidden Markov model.

The processor may be configured to generate a trigger signal when the arrangement of the estimated phonemes matches the instruction.

The voice recognition apparatus may further include a speaker operating in response to the control of the processor, wherein the processor may be configured to control the speaker to generate a predetermined voice in response to the trigger signal.

Another aspect of the present invention relates to a method for recognizing a user's voice. The method includes the steps of: storing weights for an arrangement of predetermined phonemes in a storage medium; detecting the user's voice; generating a voice signal according to the sensed voice; generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal; adjusting at least some of the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the predetermined phonemes; and estimating an arrangement of phonemes corresponding to the voice from the candidate phonemes based on the adjusted likelihood values.

The method may further include generating a trigger signal when the arrangement of the estimated phonemes matches a predetermined instruction.

The method may further include activating a predetermined operation in response to the trigger signal.

The adjusting may include: selecting ones of the candidate phonemes that match at least a part of the arrangement of the predetermined phonemes; and adjusting likelihood values corresponding to the selected candidate phonemes according to weights corresponding to the matching arrangement of the predetermined phonemes.

The voice signal may include feature vector values obtained by converting the sensed voice into a frequency domain.

The generating of the likelihood values may include generating the likelihood values corresponding to the candidate phonemes based on a hidden Markov model.

Another aspect of the invention relates to a client server in communication with a computer device. The client server may include: a communicator; a database storing weights for an arrangement of predetermined phonemes, and program codes; and a processor configured to communicate with a computer device via the communicator to provide the weights for the arrangement of the predetermined phonemes and the program codes. The program codes, when executed by the computer device, include: first instructions for generating a voice signal according to a voice sensed from a user; second instructions for generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal; third instructions for adjusting at least a portion of the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the predetermined phonemes; and fourth instructions for estimating an arrangement of phonemes corresponding to the user's voice from the candidate phonemes based on the adjusted likelihood values.

the third instructions, when executed by the computer device, select ones of the candidate phonemes that match at least a portion of the arrangement of the given phonemes, and apply to weights corresponding to the matched arrangement of the given phonemes. Accordingly, fifth instructions for adjusting likelihood values of the selected candidate phonemes may be included.

According to the present invention, there is provided a voice recognition apparatus capable of recognizing a voice with improved reliability.

1 is a block diagram illustrating a voice recognition apparatus according to an embodiment of the present invention.
2 is a diagram conceptually illustrating candidate phonemes generated from a user's voice.
3 is a diagram conceptually showing candidate phonemes and likelihood values corresponding to them.
FIG. 4 is a diagram conceptually showing weights included in the weight data set of FIG. 1 .
5 is a diagram conceptually illustrating likelihood values adjusted by applying weights.
6 is a flowchart illustrating a method of recognizing a user's voice according to an embodiment of the present invention.
FIG. 7 is a block diagram illustrating a computer device suitable for implementing the voice recognition apparatus of FIG. 1 .
8 is a block diagram illustrating an embodiment of a client server configured to provide the voice recognition module of FIG. 7 .

Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings. It should be noted that in the following description, only parts necessary for understanding the operation according to the present invention are described, and descriptions of other parts will be omitted so as not to obscure the gist of the present invention. Also, the present invention is not limited to the embodiments described herein and may be embodied in other forms. However, the embodiments described herein are provided to explain in detail enough to be able to easily implement the technical spirit of the present invention to those of ordinary skill in the art to which the present invention pertains.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "indirectly connected" with another element interposed therebetween. . The terminology used herein is for the purpose of describing particular embodiments and not for limiting the present invention. Throughout the specification, when a part "includes" a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated. “At least any one of X, Y, and Z” and “at least any one selected from the group consisting of X, Y, and Z” means one X, one Y, one Z, or two of X, Y, and Z or It can be interpreted as any combination of more (eg, XYZ, XYY, YZ, ZZ). Herein, “and/or” includes any combination of one or more of the components.

1 is a block diagram illustrating a voice recognition apparatus according to an embodiment of the present invention. 2 is a diagram conceptually illustrating candidate phonemes generated from a user's voice. 3 is a diagram conceptually showing candidate phonemes and likelihood values corresponding to them. FIG. 4 is a diagram conceptually showing weights included in the weight data set of FIG. 1 . 5 is a diagram conceptually illustrating likelihood values adjusted by applying weights.

Referring to FIG. 1 , the voice recognition apparatus 100 includes a microphone 110 , a first interface (I/F) 120 , a voice recognition processor 130 , a second interface 140 , a storage medium 150 , It may include a third interface 160 , and a function block 170 .

An acoustic sensor such as a microphone 110 is connected to the voice recognition processor 130 through the first interface 120 . The microphone 110 may detect the user's voice, and may provide digital and/or analog data corresponding to the sensed voice to the voice recognition processor 130 through the first interface 120 .

The first interface 120 may provide an interface between the microphone 110 and the voice recognition processor 130 so that the voice recognition processor 130 can communicate with the microphone 110 .

The voice recognition processor 130 is connected to the microphone 110 through the first interface 120 , is connected to the storage medium 150 through the second interface 140 , and is a functional block through the third interface 160 . 170 may be connected. The voice recognition processor 130 may receive data corresponding to the user's voice from the microphone 110, and may be configured to determine whether the voice data corresponds to a predetermined command (eg, a word or text). have. The voice recognition processor 130 may include a voice preprocessor 131 , a first voice recognizer 132 , a second voice recognizer 133 , and a voice estimator 134 .

The voice preprocessor 131 is configured to generate a voice signal in a format suitable for processing in the first voice recognition unit 132, for example, feature vector values according to voice data from the microphone 110 .

The first voice recognition unit 132 is configured to generate likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the voice signal generated by the voice preprocessor 131 . Referring to FIG. 2 , the voice preprocessor 131 provides a voice signal based on voice data SND from the microphone 110 , and the voice signal converts the voice data SND into a plurality of frames FR1 to FR3 . ) may include feature vector values converted into signals (or values) of each frequency domain. In embodiments, the voice preprocessor 131 divides the voice data SND into data of a plurality of frames FR1, FR2, FR3, and converts each divided data into signals of a frequency domain to obtain a feature vector value. can create The first speech recognition unit 132 may generate candidate phonemes CPHN1 , CPHN2 , and CPHN3 corresponding to the plurality of frames FR1 , FR2 , and FR3 from the feature vector values based on the acoustic model. 2 , first candidate phonemes CPHN1 are generated from feature vector values of a first frame FR1, and second candidate phonemes CPHN2 are generated from feature vector values of a second frame FR2, Third candidate phonemes CPHN3 are generated from feature vector values of the third frame FR3.

Next, referring to FIG. 3 , the first speech recognition unit 132 obtains likelihood values LH1 and LH2 corresponding to the first to third candidate phonemes CPHN1 to CPHN3 from the feature vector values based on the acoustic model. create For neighboring candidate phonemes (or neighboring frames), a likelihood value corresponding to a candidate phone arranged after one of them may be generated. For example, the first voice recognition unit 132 may determine the first candidate phonemes CPHN1 of the first frame FR1 as “h” and “p”, and the second frame FR2 Candidate phonemes CPHN2 may be determined as “a”, “ㅔ”, and “ㅣ”. The first voice recognition unit 132 sets the likelihood value (or probability) in which “h” is followed by “h” as 0.4, the likelihood value in which “ㅔ” is arranged after “h” as 0.4, and “heh” as Next, the likelihood value at which “ㅐ” will be arranged may be determined to be 0.2. The first voice recognition unit 132 may determine likelihood values in which “a”, “ㅔ”, and “ㅐ” are to be arranged after “p” as 0.4, 0.3, and 0.3, respectively. The first voice recognition unit 132 may determine the third candidate phonemes CPHN3 of the third frame FR3 as “a”, “ㅣ”, and “ㅐ”. The first voice recognition unit 132 may determine likelihood values in which “a”, “ㅣ”, and “ㅐ” are to be arranged as 0.1, 0.6, and 0.3, respectively. Likelihood values in which “a”, “ㅣ”, and “ㅐ” are arranged after “ㅔ” may be determined to be 0.2, 0.6, and 0.2, respectively. Likelihood values in which “a”, “ㅣ”, and “ㅐ” are arranged after “ㅐ” may be determined to be 0.2, 0.7, and 0.1, respectively.

In embodiments, the first voice recognition unit 132 may generate candidate phonemes CPHN1 to CPHN3 and likelihood values LH1 and LH2 including an algorithm based on a Hidden Markov Model. In embodiments, the acoustic model may include a Gaussian Mixture Model (GMM) or a Deep Neural Network (DNN).

Referring back to FIG. 1 , the second speech recognition unit 133 may select at least some of the likelihood values (LH1, LH2, see FIG. 3 ) corresponding to the candidate phonemes (CPHN1, CPHN2, CPHN3, see FIG. 3). It is configured to adjust according to a weight data set (WTS). Referring to FIG. 4 , the weight data set WTS includes weights WT1 and WT2 for the arrangement of phonemes (CMDPHN1, CMDPHN2, CMDPHN3, hereinafter, target phonemes) of a target instruction (hereinafter, target instruction). may include In FIG. 4 , target phonemes CMDPHN1 to CMDPHN3 are “heh”, “a”, and “ㅣ”, and the weight data set WTS has a first weight ( WT1), and a second weight (WT2) for “ㅣ” arranged after “A”. The first and second weights WT1 and WT2 may be greater than one. According to the first to third target phonemes CMDPHN1 to CMDPHN3 “h”, “h”, and “l”, the target command may be “high”.

The command “high” may not be detected based on the candidate phonemes CPHN1 to CPHN3 generated by the first voice recognition unit 132 and their likelihood values LH1 and LH2. Referring back to FIG. 3 , the likelihood value in which “a” is arranged after “h” is 0.4, and the likelihood value in which “ㅔ” is arranged after “h” is 0.4 or “ㅐ” is arranged after “h”. It may not be greater than the likelihood value 0.2 by at least a threshold. In this case, even if the user pronounces “high”, “high” may not be detected, for example, “hey” may be detected, or voice recognition may fail. As such, in particular, the speech recognition for a short-length target command such as a single word may affect the recognition performance such as ambient noise, utterance distance, microphone characteristics, and speaker difference, if only relying on an acoustic model based on feature vector values. The detection of the target command may fail even if the user pronounces the target command due to various factors causing the change.

Referring back to FIG. 1 together with FIGS. 3 and 4 , the second speech recognition unit 133 arranges target phonemes CMDPHN1 to CMDPHN3 of the weight data set WTS among the candidate phonemes CPHN1 , CPHN2 , and CPHN3 . Candidate phonemes that match , may be selected, and likelihood values of the selected candidate phonemes may be adjusted according to weights WT1 and WT2 corresponding to the arrangement of matched target phonemes CMDPHN1 to CMDPHN3. The voice estimator 134 is configured to estimate the arrangement of phonemes corresponding to the user's voice from the candidate phonemes CPHN1 , CPHN2 , and CPHN3 based on the adjusted likelihood values. In embodiments, the speech estimator 134 may perform the above estimation based on a Viterbi decoding algorithm.

As described above, the weight data set (WTS) includes a first weight (WT1) for “a” arranged after “h” and a second weight (WT2) for “l” arranged after “h”. contains The second voice recognition unit 133 selects “a” arranged after “h” and “ㅣ” arranged after “h” among the candidate phonemes CPHN1 to CPHN3, and arranges them after “h” It is adjusted by applying the first weight (WT1) to the likelihood value 0.4 corresponding to “a” being You can control it. For example, the second voice recognition unit 133 may calculate the adjusted likelihood value by multiplying the likelihood value by a corresponding weight. Accordingly, as shown in FIG. 5 , “a” arranged after “h” has an adjusted likelihood value (MLH1) of 0.8, and “ㅣ” arranged after “h” has an adjusted likelihood value (MLH2). ), which is 0.9. The adjusted likelihood value (MLH1) is a likelihood value of 0.4 where “h” and/or other likelihood values associated with “a”, for example “h” followed by “ㅔ”, and “h” followed by “ㅐ” The likelihood value to be arrayed 0.2, “p” followed by “a” may be greater than the likelihood value to be arrayed 0.4 by at least a threshold. In addition, the adjusted likelihood value (MLH2) is the likelihood values associated with “a” and/or “ㅣ”, for example, a likelihood value of 0.1 in which “a” is followed by “a”, and “a” followed by “ㅐ” may be greater than a likelihood value of 0.3 to be arranged, a likelihood value of which “I” will be arranged followed by a likelihood value of 0.6, and a likelihood value of “ㅐ” followed by “l” of 0.7 to be arranged at least by a threshold value. Based on the adjusted likelihood values MLH1 and MLH2, the voice estimator 134, for example, based on the Viterbi decoding algorithm, sets the arrangement of phonemes corresponding to the user's voice as indicated by a thick line in FIG. ”, “a”, and “l” may be estimated and/or determined.

As such, the voice recognition processor 130 may detect the target command with a relatively high probability even if the user pronounces a target command having a short length such as “high”. Accordingly, the voice recognition apparatus 100 capable of recognizing a voice with improved reliability may be provided.

The voice estimator 134 may generate a trigger signal when the arrangement of the estimated phonemes matches the target command.

The second interface 140 provides an interface between the voice recognition processor 130 and the storage medium 150 . The storage medium 150 is configured to store a weight data set (WTS). The target command may be predetermined to trigger an arbitrary operation of the voice recognition apparatus 100 , and a weight data set (WTS) corresponding to the target command is also predetermined and stored and/or updated in the storage medium 150 . can The weight data set WTS may be determined in consideration of various environments in which the voice recognition processor 130 may be used.

In embodiments, the storage medium 150 may include various types of nonvolatile storage media that retain stored data even when power is cut off, for example, a flash memory, a hard disk, and the like. .

The third interface 160 provides an interface between the voice recognition processor 130 and the function block 170 . The function block 170 may include a processor for performing at least one of various functions of the voice recognition processor 130 . The function block 170 is configured to perform or activate any operation in response to a trigger signal generated from the voice recognition processor 130 . For example, the function block 170 may output a voice signal such as “yes” or activate an arbitrary operation in response to a trigger signal. In embodiments, the functional block 170 may be configured through hardware, software, and/or a combination thereof.

In embodiments, the voice recognition device 100 is an artificial intelligence (AI) speaker, a mobile phone, a smart phone, an Ultra Mobile PC (UMPC), a workstation, a net-book, a PDA. Various computer devices capable of sensing sound and processing the detected sound as described above, such as (Personal Digital Assistants), portable computers, web tablets, portable multimedia players (PMPs), portable game consoles, etc. can be implemented as

6 is a flowchart illustrating a method of recognizing a user's voice according to an embodiment of the present invention.

Referring to FIG. 6 , in step S110, the user's voice is detected. In step S120, a voice signal is generated according to the sensed voice. The voice signal may have a data format suitable for performing step S130. In embodiments, the voice signal may include feature vector values obtained by converting the voice data sensed in step S110 into signals of a frequency domain.

In step S130, candidate phonemes are determined from the speech signal, and likelihood values corresponding to the candidate phonemes and/or an arrangement therebetween are generated.

In operation S140 , at least some of likelihood values corresponding to candidate phonemes are adjusted according to weights for an arrangement of target phonemes stored in a database, for example, the storage medium 150 of FIG. 1 . In embodiments, ones matching at least a part of an arrangement of target phonemes may be selected from among candidate phonemes, and likelihood values corresponding to the selected candidate phonemes may be adjusted according to weights corresponding to the matching arrangement of target phonemes.

In operation S150, an arrangement of phonemes corresponding to the user's voice is estimated from the candidate phonemes based on the likelihood values. Due to the likelihood values adjusted according to the weights, even if the user pronounces a short-length target command such as “high”, the estimation of the arrangement of target phonemes may be successful with a relatively high probability. Accordingly, a voice recognition method capable of recognizing a voice with improved reliability may be provided.

In step S160, when the arrangement of the estimated phonemes matches the target instruction, step S170 is performed. In step S170, a trigger signal is generated. In response to the trigger signal, the voice processing device may perform an arbitrary operation or activate it.

FIG. 7 is a block diagram illustrating a computer device suitable for implementing the voice recognition apparatus of FIG. 1 .

Referring to FIG. 7 , the computer device 1000 includes a communicator 1100 , an acoustic sensor 1200 , a user interface 1300 , a nonvolatile storage medium 1400 , a speaker 1500 , a processor 1600 , and a system memory. (1700).

The communicator 1100 is configured to transmit a wired and/or wireless signal to an external server through a network on a mobile communication network. Furthermore, the communicator 1100 may be configured to communicate with the user terminal through short-range communication, and may include Bluetooth, Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra Wideband (UWB), and Zigbee. Short-range communication technologies such as (ZigBee) may be used.

The acoustic sensor 1200 is for input of a user's voice, and is configured to detect the user's voice and provide voice data to the processor 1600 . The acoustic sensor 1200 may include the microphone 110 of FIG. 1 .

The user interface 1300 may receive a user input for controlling operations of the computer device 1000 or the processor 1600 . The user interface 1300 may include a key pad, a dome switch, a jog wheel, a jog switch, a finger mouse, and the like.

The nonvolatile storage medium 1400 may be at least one of a flash memory type, a hard disk type, and a multimedia card. The nonvolatile storage medium 1400 is configured to write and read data in response to the control of the processor 1600 . The nonvolatile storage medium 1400 may be provided as the storage medium 150 of FIG. 1 . In embodiments, the processor 1600 may receive the weight data set WTS from an external server through the communicator 1100 , and store the received weight data set WTS in the storage medium 150 .

In embodiments, a system bus configured to connect the communicator 1100 , the acoustic sensor 1200 , the user interface 1300 , the nonvolatile storage medium 1400 , the processor 1600 , and the system memory 1700 to each other is further can be provided.

The processor 1600 may include either a general-purpose or dedicated processor, and may include a communicator 1100 , an acoustic sensor 1200 , a user interface 1300 , a nonvolatile storage medium 1400 , a speaker 1500 , and a system. Controls operations of the memory 1700 .

The processor 1600 is configured to load program codes including instructions providing various functions when executed into the system memory 1700 from the nonvolatile storage medium 1400 and execute the loaded program codes. The processor 1600 loads the voice recognition module 1710 that performs functions of the voice recognition processor 130 of FIG. 1 when executed by the processor 1600 into the system memory 1700, and the loaded voice recognition module 1710 ) can be executed. For example, when the voice recognition module 1710 is executed by the processor 1600 , instructions and/or program codes for performing functions of the voice preprocessor 131 , the first voice recognition unit 132 . Commands and/or program codes for performing functions of , commands and/or program codes for performing functions of the second voice recognition unit 133 , and commands for functions of the voice estimation unit 134 and/or program codes. The processor 1600 loads an operating system 1720 configured to provide an environment suitable for the voice recognition module 1710 to be executed when executed by the processor 1600 into the system memory 1700, and the loaded operating system ( 1720) can be executed. The operating system 1720 includes components such as the communicator 1100 , the acoustic sensor 1200 , the user interface 1300 , the nonvolatile storage medium 1400 , and the system memory 1700 of the computer device 1000 and voice An interface between the recognition modules 1710 may be provided. For example, the operating system 1720 may perform the functions of the first to third interfaces 120 , 140 , and 160 described with reference to FIG. 1 when executed by the processor 1600 . The operating system 1720 may control overall operations of the computer device 1000 .

The computer device 1000 may include hardware, software, and/or a combination thereof for performing at least one of various functions according to embodiments, which may be provided as the function block 170 of FIG. 1 . . In embodiments, the processor 1600 loads instructions and/or program codes for performing at least one of the above various functions from the nonvolatile storage medium 1400 into the system memory 1700 , and includes the loaded instructions and The functional block 170 of FIG. 1 may be provided by/or executing program codes. Such a function block 170 may activate any operation in response to a trigger signal. In embodiments, the function block 170 may control the speaker 1500 to output a voice signal such as “yes” in response to a trigger signal.

In embodiments, the system memory 1700 may include at least one of random access memory (RAM), read only memory (ROM), and other types of computer-readable storage media. Although the system memory 1700 is illustrated as a separate component from the processor 1600 , this is exemplary and at least a portion of the system memory 1700 may be integrated into the processor 1600 . In embodiments, the system memory 1700 may be provided as a buffer memory.

8 is a block diagram illustrating an embodiment of a client server configured to provide the voice recognition module of FIG. 7 .

According to an embodiment of the present invention, program codes and/or instructions executed by the computer device 1000 of FIG. 7 may be provided from the client server 2000 . Referring to FIG. 8 , the client server 2000 may include a communicator 2100 , a processor 2200 , and a database 2300 . The communicator 2100 may communicate with the computer device 1000 through a network. The database 2300 includes program codes and/or instructions that may be executed by the computer device 1000 and/or the processor 1600 of FIG. 7 , for example, an application including the voice recognition module 1710 of FIG. 7 . You can save the application. The processor 2200 may provide the application application stored in the database 2300 to the computer device 1000 through the communicator 2100 in response to a request from the computer device 1000 . The application application may be installed and executed in the computer device 1000 .

Although specific embodiments and application examples have been described herein, these are provided only to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments, and those of ordinary skill in the art to which the present invention pertains Various modifications and variations are possible from this substrate as it grows.

Therefore, the spirit of the present invention should not be limited to the described embodiments, and not only the claims described below, but also all those with equivalent or equivalent modifications to the claims will be said to belong to the scope of the spirit of the present invention. .

110: microphone
130: speech processing processor
150: storage medium
170: function block

Claims (20)

In the speech recognition device:
a storage medium storing weights for the arrangement of phonemes of a predetermined voice command for triggering a predetermined operation of the voice recognition apparatus;
a voice pre-processing unit for generating a voice signal according to a voice sensed by a user;
a first speech recognition unit configured to generate likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal;
a second speech recognition unit configured to selectively adjust the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the phonemes of the speech command; and
and a voice estimator configured to estimate an arrangement of phonemes corresponding to the user's voice from the candidate phonemes based on the adjusted likelihood values. The method of claim 1,
The second speech recognition unit selects among the candidate phonemes that match the arrangement of the phonemes of the speech command, and corresponds to the selected candidate phonemes according to the weights for the arrangement of the phonemes of the speech command. Speech recognition device configured to adjust likelihood values. The method of claim 1,
The voice signal includes feature vector values obtained by converting the sensed voice into a frequency domain. The method of claim 1,
The first speech recognition unit is configured to generate the likelihood values corresponding to the candidate phonemes based on a Hidden Markov Model. The method of claim 1,
and the speech estimator is configured to generate a trigger signal when the arrangement of the estimated phonemes matches the arrangement of the phonemes of the voice command. 6. The method of claim 5,
and a function block configured to activate the predetermined operation in response to the trigger signal. In the speech recognition device:
an acoustic sensor configured to detect a voice from a user;
a processor configured to determine whether the sensed voice corresponds to a command; and
a storage medium for storing weights for the arrangement of phonemes of a predetermined voice command for triggering a predetermined operation of the voice recognition device,
The processor is
generate a voice signal according to the sensed voice;
generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal;
selectively adjusting the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of phonemes of the voice command;
and estimating an arrangement of phonemes corresponding to the user's voice from the candidate phonemes based on the adjusted likelihood values. 8. The method of claim 7,
the processor selects ones of the candidate phonemes that match the arrangement of the phonemes of the speech command, and a likelihood value corresponding to the selected candidate phonemes according to the weights for the arrangement of the phonemes of the speech command A voice recognition device configured to control the sounds. 8. The method of claim 7,
The voice signal includes feature vector values obtained by converting the sensed voice into a frequency domain. 8. The method of claim 7,
and the processor is configured to generate the likelihood values corresponding to the candidate phonemes based on a hidden Markov model. 8. The method of claim 7,
and the processor is configured to generate a trigger signal when the arrangement of the estimated phonemes matches the arrangement of the phonemes of the speech command. 12. The method of claim 11,
Further comprising a speaker operating in response to the control of the processor,
and the processor is configured to control the speaker to generate a predetermined voice as the predetermined operation in response to the trigger signal. In a method for recognizing a user's voice:
storing in a storage medium weights for arrangement of phonemes of a predetermined voice command for triggering a predetermined operation;
detecting the user's voice;
generating a voice signal according to the sensed voice;
generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal;
selectively adjusting the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of the phonemes of the voice command; and
and estimating an arrangement of phonemes corresponding to the voice from the candidate phonemes based on the adjusted likelihood values. 14. The method of claim 13,
generating a trigger signal when the estimated arrangement of phonemes matches the arrangement of the phonemes of the predetermined spoken command. 15. The method of claim 14,
activating the predetermined action in response to the trigger signal. 14. The method of claim 13,
The adjusting step is
selecting ones of the candidate phonemes that match the arrangement of the phonemes of the voice command; and
and adjusting likelihood values corresponding to the selected candidate phonemes according to the weights for the arrangement of phonemes of the speech command. 14. The method of claim 13,
The voice signal includes feature vector values obtained by converting the sensed voice into a frequency domain. 14. The method of claim 13,
The generating the likelihood values includes generating the likelihood values corresponding to the candidate phonemes based on a hidden Markov model. a database storing program codes and weights for the arrangement of phonemes of a predetermined voice command for triggering a predetermined operation; and
a processor configured to communicate with a computer device via a communicator to provide the weights for the arrangement of the phonemes of the spoken command and the program codes;
When the program codes are executed by the computer device,
first instructions for generating a voice signal according to a voice sensed by the user;
second instructions for generating likelihood values corresponding to the candidate phonemes while discriminating candidate phonemes from the speech signal;
third instructions for selectively adjusting the likelihood values corresponding to the candidate phonemes according to the weights for the arrangement of phonemes of the voice command; and
and fourth instructions for estimating an arrangement of phonemes corresponding to the user's voice from the candidate phonemes based on the adjusted likelihood values. 20. The method of claim 19,
the third instructions, when executed by the computer device, select ones of the candidate phonemes that match the arrangement of the phonemes of the voice command, according to the weights for the arrangement of the phonemes of the voice command and fifth instructions for adjusting likelihood values of the selected candidate phonemes.

Images