KR101889465B1 - voice recognition device and lighting device therewith and lighting system therewith - Google Patents

Abstract

The present invention separates the original signals S1 and S2 and the noise signals N1 and N2 from the acoustic signals H1 and H2 input from the two microphones, (S3) and a noise signal (N3) from an acoustic signal (H3) input from another microphone, a first speech recognition module for detecting the first original signal (X1) And a second voice recognition module for detecting the final original signal X2 by summing the original signal S3 separated after the separation and the primary signal X1 input from the first voice recognition module, It is possible to improve the accuracy and reliability of voice recognition by canceling the phenomenon that the signals are removed together, and the voice can be accurately and precisely detected by repeatedly detecting the original signal by the voice recognition modules, Module and the second speech recognition module use different signal separation algorithms The advantages of each signal separation algorithm are emphasized, but the disadvantages can be offset to improve the accuracy of speech recognition, and each microphone can eliminate the noise recurrence phenomenon from the input signal (AEC) to adaptively remove a dynamic noise source, an illumination apparatus including the speech recognition apparatus, and an illumination system using the illumination apparatus.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a voice recognition device, a lighting recognition device and a lighting system using the same,

BACKGROUND OF THE INVENTION 1. Field of the Invention [0001] The present invention relates to a speech recognition apparatus, an illumination apparatus including the speech recognition apparatus, and an illumination system using the same. More particularly, the present invention relates to an illumination system using a plurality of speech recognition modules The present invention relates to a speech recognition apparatus, a speech recognition apparatus, and a lighting system using the speech recognition apparatus. The speech recognition apparatus is capable of recognizing speech, thereby simplifying the calculation process and significantly improving the accuracy and reliability of speech recognition.

Speech recognition is a technique for extracting a characteristic parameter by removing noise from an acoustic speech signal obtained through an acoustic collection sensor such as a microphone or the like, The field of application and demand are gradually increasing due to the ability to support interfaces of human and devices without being defined and to be physically connected, and to be able to interoperate with various applications such as security, home network, illumination, robots and navigation There is a trend.

Especially in recent years, electronic devices have been developed and various devices have been networked with each other, and the diffusion of the internet of things (IoT) service has been increasing. The most convenient means for recognizing is the human bio-signal, and since the voice is the most basic signal among the human bio-signals, various researches on the voice recognition technology are being carried out.

Speech enhancement is a technique for improving speech quality by removing noise components from a sound including background noise of various types in the speech recognition technology. Channel speech signal, such as a MMSE-STSA, and 2) a Generalized Sideobe Canceller (GSC) and a Delay-Sum Beamforming (DSB)

In this case, since the speech recognition technology using the multi-channel speech signal can adaptively remove the static noise and the dynamic noise simultaneously using the difference between the microphone input signals which can not be obtained with the single microphone signal, And is suitable for a real-life environment in which various kinds of noises such as human conversation are present.

FIG. 1 is a block diagram showing a signal source separation apparatus according to a first embodiment of the present invention disclosed in Korean Patent No. 10-0486736 (entitled " Method and Apparatus for Separating Signals According to Objective Using Two Sensors).

1, a signal source separation unit 100 includes a signal absence probability calculation unit 111, a destination identification unit 113, a signal estimation unit 115, And a signal separation unit 117 for each source.

The signal absence probability calculation unit 111 calculates a wide-area signal absence probability for the m-th frame with respect to one of the first microphone reception signal and the second microphone reception signal received from the pair of microphones and converted into the frequency domain.

The signal estimator 115 estimates a spectrum vector from which a noise signal is removed for each frequency band according to the global signal absence probability calculated by the signal absence probability calculator 111. [

The target circle identification unit 113 receives the local signal absence probability of each of the first microphone reception signal or the second microphone reception signal calculated by the signal absence probability calculation unit 111 and compares it with a predefined first threshold value And determines whether the target signal exists in the corresponding frequency band of each frame according to the comparison result.

Also, if it is determined that the target signal is present, the target identification unit 113 generates a magnitude attenuation and a delay time with respect to the signal value of the frequency band, And the frequency band belonging to each destination circle.

The destination-specific signal separation unit 117 multiplies the estimated vector vector for each frequency band obtained by the signal estimation unit 115 by the target-source label vector obtained by the destination-source identification unit 113, do.

The conventional art 100 configured as described above has an advantage that noise can be removed from a mixed signal received from two microphones in real time, and at the same time, an arbitrary N destination signal can be separated.

However, in the conventional art 100, when the signal estimator 115 estimates a spectrum vector, the noise signal is removed for each frequency band according to the global signal absence probability calculated by the signal absence probability calculator 111, However, this configuration causes a problem that the voice recognition rate is lowered because a part of the original signal is removed together with the noise when the noise is removed.

For example, even when it is assumed that M microphones having three or more microphones are used in the prior art 100, when estimating the spectral vectors of at least one of the first through M microphone reception signals input from the M microphones , And the noise signal is removed for each frequency band by the wide-area signal absence probability. Therefore, the target signal is removed together with the noise reduction, and the voice recognition rate is lowered.

That is, the conventional art 100 estimates the source signals by eliminating the noise signal discriminated as noise from the received signal, generates a label vector according to the destination, multiplies the generated label vectors according to the destination object, Even if three or more microphones are used, it is structurally limited that the original signal is removed at the time of noise removal because it is configured to acquire a separated signal.

In addition, since the prior art (100) comprises only one module (including a signal absence probability calculation number, a source identification unit, a signal estimation unit, and a target signal separation unit) for separating a signal for each source from a received signal, The operation for separating the signal is performed only once and the speech recognition rate is lowered.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems and it is an object of the present invention to provide a sound recognition apparatus and a sound recognition method in which sound signals S1 and S2 are respectively output from sound signals H1 and H2 input from two microphones, A first speech recognition module for separating the signals N1 and N2 and then summing the original signals S1 and S2 to detect the primary signal X1; The original signal S3 separated from the original signal S3 and the noise signal N3 is summed with the primary signal X1 input from the first voice recognition module to obtain the final original signal X2 A voice recognition device capable of enhancing the accuracy and reliability of voice recognition by canceling the phenomenon that the original signals are removed together when the noise is removed by including the second voice recognition module for detecting the voice, Illumination system.

In addition, another object of the present invention is to provide a speech recognition apparatus capable of accurately and precisely detecting speech by repeatedly detecting the original signal by speech recognition modules, an illumination lamp apparatus having a speech recognition apparatus, Illumination system.

Another advantage of the present invention is that the first speech recognition module and the second speech recognition module are configured to separate the original signal and the noise signal from the acoustic signal using different signal separation algorithms, The present invention is to provide a speech recognition apparatus, speech recognition apparatus, speech recognition apparatus, and illumination system using the same, which can improve the accuracy of speech recognition by canceling out the disadvantages.

Another object of the present invention is to provide a speech recognition apparatus capable of adaptively removing a dynamic noise source by applying Acoustic Echo Cancellation (AEC) for eliminating a noise recurrence phenomenon from an input signal of each microphone, And a lighting system using the same.

Another object of the present invention is to increase the voice recognition rate to accurately determine whether or not an unexpected event occurs, and at the same time, when an unexpected event occurs, the light is directed toward the emergency exit, And an illumination system using the illumination device. [0002] The present invention relates to an illumination system and a lighting system using the same.

According to an aspect of the present invention, there is provided an audio signal processing apparatus including first, second, and third microphones that collect acoustic signals and convert the acoustic signals into electric signals; A reference model database unit storing predetermined reference models; An acoustic signal input unit receiving acoustic signals obtained by the microphones; A voice recognition unit for analyzing the acoustic signals input by the acoustic signal input unit and detecting the original signal X2; A feature parameter generation unit for extracting a feature vector of the original signal X2 detected by the speech recognition unit and generating a feature parameter using the extracted feature vector; A comparison and matching unit for analyzing the reference models stored in the reference model database unit and the feature parameters generated by the feature parameter generation unit using a predetermined comparison algorithm to detect a reference model having the highest similarity to the feature parameters; And a word determining unit for searching for a word using a character corresponding to the reference model detected by the comparison and matching unit as a search word and finally outputting the searched word to perform speech recognition, The sound signal of the first microphone is separated into the original signal S1 and the noise signal N1 by using the separation algorithm and the sound signal of the second microphone is separated into the original signal S2 and the noise A first speech recognition module for separating the original signal S1 by a signal N2 and for summing the original signals S1 and S2 to detect a primary signal X1; And separating the sound signal of the third microphone into the original signal (S3) and the noise signal (N3) using a predetermined second signal separation algorithm, and then separating the original signal into a first order signal detected by the first speech recognition module Further comprising a second speech recognition module for summing the signal X1 to detect the final original signal X2, wherein the first signal separation algorithm and the second signal separation algorithm generate an original signal < RTI ID = 0.0 > S) and the noise signal (N).

delete

Also, in the present invention, the first, second and third microphones are applied with a beam-forming technique, the first microphone forms a beam in the straight direction, and the second and third microphones form a beam to the left and right Preferably, the first, second, and third microphones are applied with Acoustic Echo Cancellation (AEC) for removing a dynamic noise source from input acoustic signals.

In the present invention, when the final original signal X2 is detected, the speech recognition unit combines the vowel set in the initials of the detected original signal X2 into syllables whose continuity is eliminated, and the comparison and matching unit inputs Dynamic Time Warping (DTW) for recognizing the speech of the reference model with the highest similarity by nonlinearly matching the input feature parameter and the reference model in order to compensate for the difference in the speaking speed and the length between the speech and the reference speech ) Algorithm to calculate a Squared Euclidean distance of each of the feature parameters and the reference models, and then recognizes the reference model having the smallest distance as the model most similar to the feature parameter, The reference model having the highest degree of similarity is determined as the input voice if there is one reference model within the predetermined threshold value If there are two or more reference models whose similarity with a specific parameter is within a predetermined threshold value, the speech signal is divided into phonemes and then the highest similarity is obtained through a pattern comparison algorithm through a hidden Markov model It is desirable to determine phonemes as input speech.

delete

delete

delete

delete

delete

According to another aspect of the present invention, there is provided an illumination system including a light source device, a housing, a microphone installed outside the housing, and an illumination lamp device including a control unit installed inside the housing, the illumination system comprising: A turn-off management unit; The control data of the actual illumination for each of the 'illumination control related comparison characters' defined by the characters that can be determined as the characters for controlling the illumination of the voice word and the 'illumination control related comparison characters' And a memory in which a matched matching table is stored, wherein the controller analyzes a sound signal input from the microphone to determine a voice word, and the determined voice word is any one of the 'illumination control related comparison characters' The control unit detects the control data by searching the matching table and then controls the lighting according to the control data detected by the lighting control unit. The microphones are composed of first, second and third microphones installed at intervals, The control unit includes a reference model database unit storing predetermined reference models; An acoustic signal input unit receiving acoustic signals obtained by the first, second and third microphones; A voice recognition unit for analyzing the acoustic signals input by the acoustic signal input unit and detecting the original signal X2; A feature parameter generation unit for extracting a feature vector of the original signal X2 detected by the speech recognition unit and generating a feature parameter using the extracted feature vector; A comparison and matching unit for analyzing the reference models stored in the reference model database unit and the feature parameters generated by the feature parameter generation unit using a predetermined comparison algorithm to detect a reference model having the highest similarity to the feature parameters; A word determination unit for searching for a word using a character corresponding to the reference model detected by the comparing and matching unit as a search word and finally outputting the searched word to perform speech recognition; When the detected relation is greater than a threshold value after detecting the association between the voice word determined by the word determination unit and each of the 'illumination control related comparison characters' stored in the memory, the matching table is searched for, Further comprising a determination unit configured to detect the control data corresponding to the control-related comparison object character 'and input the control data to the light-off management unit, wherein the voice recognition unit uses the predetermined first signal separation algorithm to convert the acoustic signal of the first microphone into the original signal And separates the sound signal of the second microphone into the original signal S2 and the noise signal N2 using the first signal separation algorithm and outputs the separated original signal S1 ) And (S2) to detect the primary signal X1; And separating the sound signal of the third microphone into the original signal (S3) and the noise signal (N3) using a predetermined second signal separation algorithm, and then separating the original signal into a first order signal detected by the first speech recognition module Further comprising a second speech recognition module for summing the signal X1 to detect the final original signal X2, wherein the first signal separation algorithm and the second signal separation algorithm generate an original signal < RTI ID = 0.0 > S) and the noise signal (N).

delete

In the present invention, when the final original signal X2 is detected, the speech recognition unit combines the vowel set in the initials of the detected original signal X2 into syllables whose continuity is eliminated, and the comparison and matching unit inputs Dynamic Time Warping (DTW) for recognizing the speech of the reference model with the highest similarity by nonlinearly matching the input feature parameter and the reference model in order to compensate for the difference in the speaking speed and the length between the speech and the reference speech ) Algorithm to calculate a Squared Euclidean distance of each of the feature parameters and the reference models, and then recognizes the reference model having the smallest distance as the model most similar to the feature parameter, The reference model having the highest degree of similarity is determined as the input voice if there is one reference model within the predetermined threshold value When two or more reference models with similarity to a specific parameter are within a predetermined threshold value, the speech signal is divided into phonemes and then the similarity degree is calculated through a pattern comparison algorithm through a hidden Markov model It is desirable to determine a high phoneme input voice.

Further, in the present invention, 'memory-related characters to be compared', which can be determined as an unexpected situation, are further stored in the memory, and the determination unit compares the voice word determined by the word determination unit and the ' And if it is determined by the determination unit that an unexpected situation has occurred, it is determined that an unexpected situation has occurred, and if the detected relation is greater than the threshold value, By controlling the lighting to be turned on in a set manner, it is preferable that the illumination lamp apparatus is used for illumination in a normal situation where an unexpected situation does not occur, and is used as an emergency illumination in an unexpected situation.

Further, in the present invention, the lighting system may further include a controller for managing and controlling the lighting apparatuses, and when the controller determines that an unexpected condition has occurred, the controller determines that an unexpected condition has occurred And the controller stores the location information of each of the luminaire mechanisms and the predetermined emergency evacuation route and when the unexpected situation confirmation data is received from any one of the luminaire mechanisms, It is preferable to arrange the light fixtures in the order corresponding to the emergency evacuation route by using the location information and the emergency evacuation route, and then to perform the blinking according to the ordered sequence.

According to the present invention having the above-described problems and the solution, the speech recognition unit generates the original signals S1 and S2 and the noise signals N1 and N2 from the acoustic signals H1 and H2 input from the two microphones, A first speech recognition module for detecting the first original signal X1 by summing the original signals S1 and S2 after separating the original signals S3 and N2 from the acoustic signals H3 inputted from other microphones; And a second voice recognition unit for detecting a final original signal X2 by summing the original signal S3 separated from the noise signal N3 and the original signal X1 input from the first speech recognition module, By including the module, it is possible to improve the accuracy and reliability of speech recognition by canceling the phenomenon that the original signal is removed together when the noise is removed.

Further, according to the present invention, detection of the original signal is repeatedly performed by the speech recognition modules, so that the speech can be accurately and precisely detected.

According to the present invention, since the first speech recognition module and the second speech recognition module separate the original signal and the noise signal from the acoustic signal using different signal separation algorithms, the advantages of each signal separation algorithm are highlighted, So that the accuracy of speech recognition can be further enhanced.

According to the present invention, a dynamic noise source can be adaptively removed by applying acoustic echo cancellation (AEC) for eliminating noise recurrence from an input signal of each microphone.

In addition, according to the present invention, it is possible to accurately determine whether an unexpected situation occurs by raising the voice recognition rate, and to illuminate in the direction toward the exit when an unexpected situation occurs, thereby promptly coping with the unexpected situation, .

FIG. 1 is a block diagram showing a signal source separation apparatus according to a first embodiment of the present invention disclosed in Korean Patent No. 10-0486736 (entitled " Method and Apparatus for Separating Signals According to Objective Using Two Sensors).
2 is a block diagram showing a speech recognition apparatus according to an embodiment of the present invention.
3 is an exemplary diagram for explaining a pre-processing technique applied to the microphone of FIG.
4 is a block diagram showing the speech recognition unit of FIG.
5 is an exemplary diagram illustrating a first signal separation algorithm applied to the first speech recognition module of FIG.
FIG. 6 is a flowchart for explaining the operation procedure of FIG. 2. FIG.
FIG. 7 is a block diagram illustrating a lighting system to which the speech recognition apparatus of FIG. 2 is applied.
8 is a block diagram showing a control unit installed in the illumination lamp of Fig.
Fig. 9 is an exemplary view of Fig. 7. Fig.
10 is a block diagram showing the controller of Fig.
11 is an exemplary diagram showing a blinking period determined by the lamp control unit of Fig.

Hereinafter, an embodiment of the present invention will be described with reference to the accompanying drawings.

2 is a block diagram showing a speech recognition apparatus according to an embodiment of the present invention.

The speech recognition apparatus 1 according to an embodiment of the present invention receives a sound signal from a plurality of microphones 11-1, 11-2, and 11-3, The module 51 outputs each of the input signals H1 and H2 to the input signals H1 and H2 of the first and second microphones 11-1 and 11-2 using the first signal separation algorithm. S2 + S2 from the original signals S1 and S2 after the original signals S1 and S2 and the noise signals N1 and N2 are separated from the original signals S1 and S2, ).

The second speech recognition module 52 of FIG. 4 to be described later uses the second signal separation algorithm to calculate the original signal S3 and the noise signal N3 from the input signal H3 of the third microphone 11-3 The detected original signal S3 is added to the primary signal X1 input from the first voice recognition module 51 to detect the final original signal X2 = S3 + X1.

That is, in the speech recognition apparatus 1 according to the present invention, since the input signals input to the microphones for the same acoustic signal are different in frequency magnitude of the original signal and noise according to the direction and distance to the acoustic generating position, Through the two speech recognition modules with different signal separation algorithms, it is possible to cancel out the phenomenon that the original signals are removed together with the noise removal for each input signal, and at the same time, the original signal is repeatedly detected, The advantages of each signal separation algorithm are highlighted, but the accuracy of speech recognition can be significantly improved by canceling out the disadvantages.

At this time, the first and second speech recognition modules 51 and 52 may be configured to separate the original signal and the noise signal from the acoustic signal using the same signal separation algorithm, but different signal separation algorithms may be applied , It is more effective in speech recognition because it can offset the disadvantages of each signal separation algorithm.

In the present invention, for example, three microphones and two voice recognition modules have been described for convenience of description. However, the number of microphones may be four or more and the number of voice recognition units may be three or more Of course.

2, the speech recognition apparatus 1 includes an acoustic signal input unit 3, a speech recognition unit 5, a feature parameter detection unit 6, a comparison and matching unit 7, (8), and a word determination unit (9).

The acoustic signal input unit 3 receives acoustic signals input from the three microphones 11-1, 11-2, and 11-3. At this time, the microphones 11-1, 11-2, and 11-3 are installed so as to have different sound input angles to receive sound signals, respectively, and a sound signal and a noise signal (noise) .

At this time, the microphones 11-1, 11-2, and 11-3 are installed in an array form, and receive the acoustic signals and convert them into electric signals.

In addition, the microphones 11-1, 11-2, and 11-3 are applied with a beam-forming technique so that the first microphone 11-1 forms a beam in the forward direction, The microphone 11-2 and the third microphone 11-3 are installed to symmetrically form the left and right sides of the beam of the second microphone 11-2.

In this case, the beam-forming technique is a signal processing technique that is mainly used for the purpose of adjusting the direction and sensitivity of a radiation pattern by using an array of a transmitting device or a receiving device. When a signal is transmitted, The intensity of the signal transmitted in the other direction can be reduced.

3 is an exemplary diagram for explaining a pre-processing technique applied to the microphone of FIG.

As shown in FIG. 3, the microphone 11 of the present invention has a dynamic noise source by applying acoustic echo cancellation (AEC) capable of removing a noise recurrence phenomenon from an input signal input from the microphone 11 So that only the input signal input from the microphone 11 is extracted.

In addition, the acoustic echo cancellation (AEC) technique uses a normalized least mean square (NLMS) -based variable learning rate noise elimination algorithm to remove a dynamic noise source such as a speaker and keep the input signal input from the microphone 11 in a natural state It is possible to perform the preprocessing function on the input signal.

That is, the acoustic signal input unit 3 of the present invention receives the acoustic signals from the microphones 11-1, 11-2, and 11-3, and the microphones 11-1, 11-2, The acoustic echo cancellation (AEC) technique can be applied to each of the sound processing units 11-3 to receive the preprocessed sound signals, thereby improving the accuracy of speech recognition.

FIG. 4 is a block diagram showing the speech recognition unit of FIG. 2, and FIG. 5 is an exemplary diagram illustrating a first signal separation algorithm applied to the first speech recognition module of FIG.

As shown in FIG. 4, the speech recognition unit 5 receives the first and second microphones 11-1 and 11-2 input through the sound signal input unit 3 using a predetermined first signal analysis algorithm, The original signals S1 and S2 are separated from the input signals H1 and H2 of the input signals S1 and S2 and the noise signals N1 and N2, A first voice recognition module 51 for detecting a one-way signal X1 of the third microphone 11-3 input through the sound signal input part 3 using a predetermined second signal analysis algorithm, And separates the original signal S3 and the noise signal N3 to add the separated original signal S3 to the first original signal X1 input from the first speech recognition module 51, And a second voice recognition module 52 for detecting the second voice X2.

At this time, the first and second speech recognition modules 51 and 52 may be configured to separate the original signal and the noise signal from the acoustic signal using the same signal separation algorithm, but different signal separation algorithms may be applied , It is more effective in speech recognition because it can offset the disadvantages of each signal separation algorithm.

The first speech recognition module 51 receives the acoustic signals H1 and H2 of the first and second microphones 11-1 and 11-2 from the acoustic signal input unit 3. The first speech recognition module 51 receives the acoustic signals H1 and H2 from the first and second microphones 11-1 and 11-2.

The first speech recognition module 51 separates the inputted sound signal H1 into the original signal S1 and the noise signal N1 using a predetermined first signal separation algorithm, Into the original signal S2 and the noise signal N2.

The first speech recognition module 51 detects the primary signal X1 by summing the separated original signals S1 and S2 when the signals are separated from the acoustic signal.

The second voice recognition module 52 uses the predetermined second signal separation algorithm to generate the original signal S3 and the noise signal N3 from the third microphone 11-3 input through the acoustic signal input unit 3 Separate.

The second speech recognition module 52 also detects the final original signal X2 by summing the original signal S3 separated from the original speech signal and the primary signal X1 input from the first speech recognition module 51. [

In other words, the present invention is configured such that the first speech recognition module 51 and the second speech recognition module 52 separate the original signal and the noise signal from the acoustic signal using different signal separation algorithms, The module 51 detects the primary signal X1 by summing the original signals for the first and second microphones 11-1 and 11-2 and the second voice recognition module 52 detects The first original signal X1 detected by the first speech recognition module 51 so as to detect the final original signal X2 so that two different signal separation algorithms are applied, It is possible not only to cancel out the phenomenon that the original signals are removed together with the noise elimination for each input signal through the speech recognition modules 51 and 52 but also to detect the original voice accurately and precisely Can be applied to each speech recognition module, The advantages of the speech recognition system can be emphasized, but the accuracy of speech recognition can be significantly improved by canceling out the weak points.

In this case, the first signal separation algorithm applied to the first speech recognition module 51 and the second signal separation algorithm applied to the second speech recognition module 52 are configured to have different types of calculation processing.

For example, as shown in FIG. 5, the first signal separation algorithm applied to the first speech recognition module 51 is an algorithm for dividing the sound inputted from the first and second microphones 11-1 and 11-2, (BSS, Blind Source Separation) applying an IE soft-mask algorithm and an IVA algorithm after converting the signals H1 and H2 into a frequency domain (STFT, short-time Fourier transform) Can be applied.

The second signal separation algorithm applied to the second speech recognition module 52 may be a single channel sound source separation method.

For example, the second signal separation algorithm first transforms the sound signal H3 input from the third microphone 11-3 into a frequency domain (STFT), and then separates the signals through ICA (Independent Component Analysis).

In this case, when ICA is applied to the second signal separation algorithm, first, a linear transformation is performed to reduce the dimension of the acoustic signal H1 to the dimension of a sound source, and then a unitary matrix ( B) to obtain the value of the frequency domain of the separated signal, and detects the separated signal through the separation matrix (V * B) detected earlier.

That is, assuming that a soft mask (IE soft-mask) and a blind source separation (BSS) including IVA are applied to the first signal separation algorithm, the first signal separation algorithm has a reverberation time, The second signal separation algorithm assumes that the residual cross-talk component exists in each channel even after the separation so that the separation performance deteriorates. When the ICA is applied to the second signal separation algorithm, Frequency bins are not independent and are vulnerable to static noise.

However, in the present invention, 1) the first speech recognition module 51 separates the original signals S1 and S2 using the first signal separation algorithm, and 2) the second speech recognition module 52 uses the second signal 3) the final original signal X2 is output to the first original signal X1 by the first speech recognition module 51 and the original signal X2 by the second speech recognition module 52 The first signal separation algorithm and the second signal separation algorithm can be canceled from each other, and the detection of the original signal can be repeatedly performed, so that the voice can be accurately and precisely detected The advantage of the signal separation algorithm applied to each speech recognition module is highlighted, but the accuracy of the speech recognition can be significantly improved by canceling out the disadvantages.

Further, when the final original signal X2 is detected by the second speech recognition module 52, the speech recognition unit 5 combines the vowel set in the prefix of the detected original signal, and converts the vowel to the deleted syllable .

For example, the speech recognition unit 5 converts the detected speech into a syllable such as 'Hagada' when the second speech recognition module 52 searches for 'Hong Gil Dong'.

In the present invention, three microphones and two voice recognition modules have been described for convenience of explanation. However, if the number of microphones is four or more and the number of voice recognition modules is three or more, The n-th speech recognition module detects the original signal by using the acoustic signal input from the (n-1) microphone and the original signal input from the (n-1) .

The feature parameter detector 6 analyzes the original signal detected by the speech recognition unit 5 and extracts a feature vector necessary for recognition.

At this time, the feature parameter detector 6 extracts a feature vector from the speech signal input through LPC (Linear Predictive Coding).

The feature parameter detector 6 generates feature parameters for the input speech signal using the extracted feature vectors. In this case, the feature parameter is data obtained by processing the speech signal so that a comparison algorithm with the reference model can be performed.

The characteristic parameters detected by the characteristic parameter detecting section 6 are input to the comparing and matching section 7.

The comparison and matching unit 7 analyzes a predetermined reference model stored in the reference model database unit 8 using a preset comparison algorithm and information on the reference model having the highest similarity to the feature parameter As a speech recognition result.

In other words, the comparison and matching unit 7 generates and compares and analyzes the feature parameters input from the feature parameter detection unit 6 and predetermined reference models in syllable units.

At this time, the reference model database 8 stores preset reference model information.

The comparison and matching unit 7 non-linearly matches the input feature parameter and the reference model in order to compensate for the difference in the speed and length between the input speech and the reference speech, thereby recognizing the speech of the reference model having the highest degree of similarity (Squared Euclidean Distance) of each of the feature parameters and the reference models is calculated using a dynamic time warping (DTW) algorithm for the reference model, and the reference model having the smallest distance is used as the model most similar to the feature parameter .

At this time, for a specific input speech (feature parameter), the reference model may have one reference model or two or more reference models whose degree of similarity with the feature parameter is within a predetermined threshold value. For example, since the signal patterns are similar to each other due to the similarity of pronunciations of 'a', 'ka', 'da' and 'ta', the reference model having similarity to the input speech ' Quot ;, and " k ".

Accordingly, the comparing and matching unit 7 determines the reference model having a high degree of similarity as the input voice if there is one reference model whose similarity with the feature parameter is within a predetermined threshold value.

In addition, the comparison and matching unit 7 performs an analysis with a better recognition rate than the dynamic time warping method when there are two or more reference models whose similarity with a specific parameter is within a predetermined threshold value. Specifically, And then performs a pattern comparison algorithm using the Hidden Markov model. The Hidden Markov Model is a statistical model that determines hidden parameters from the observed parameters based on the assumption that the modeling system is a Markov process with unknown parameters and is a widely used method in the field of speech recognition The detailed description will be omitted.

Further, the comparison and matching unit 7 inputs the voice for the detected reference model to the word determination unit 9. [

The word determination unit 9 searches for words using a character corresponding to the reference model input from the comparison and matching unit 7 as a search word, and performs speech recognition by finally outputting the searched word.

That is, the speech recognition apparatus 1 of the present invention can accurately recognize the speech as the number of reference models stored in the reference model database unit 8 and compared with the feature parameters increases. In this case, the capacity of the reference model database unit 8 In addition to being large, it is necessary to increase the amount of computation when performing comparison algorithm between feature parameters and reference model. In order to derive correct speech recognition result with minimum resource because the system resource is limited when applied to an embedded system, In the invention, the speech recognition based on the prefix is applied.

Especially, in inputting 'ㄱ', 'ㄴ', 'ㄷ', etc. as a voice, it does not use the name of the prefix such as' base ',' A single unified vowel is combined with the first vowel, and the feature parameter is made to correspond to the voice signal having a combination of the vowel and the unified vowel.

FIG. 6 is a flowchart for explaining the operation procedure of FIG. 2. FIG.

The speech recognition method S1 as an operation process of the speech recognition apparatus 1 of the present invention includes a sound signal input step S10, a speech recognition step S20, a feature parameter generation step S30, an analysis step S40, A determination step S50, a phoneme unit pattern analysis step S60, a phoneme determination step S70, and a word determination step S80.

The acoustic signal input step S10 is a step of receiving acoustic signals from the microphones 11-1, 11-2, and 11-3.

The speech recognition step S20 is a step of recognizing speech using the two speech recognition modules as described above with reference to FIG. 4 for the acoustic signals input through the speech signal input step S10.

In addition, in the voice recognition step (S20), speech recognition is performed using a reference model generated based on pronunciation of a combination of vowels common to the initials, so that they have a common vowel such as 'a', 'b' , And a syllable that does not contain a trait.

For example, the voice recognition step S20 converts the detected voice into a voice such as " Hagada " when " Hong Gil-dong " is detected by the voice recognition modules.

The voice recognition step S20 inputs the converted voice signal to the feature parameter generation step S30.

The feature parameter generation step S30 extracts a feature vector from the speech signal input through LPC (Linear Predictive Coding).

The feature parameter detection step S30 generates feature parameters for the input speech signal using the extracted feature vectors. In this case, the feature parameter is data processed so that a voice signal can be compared with a reference model.

In addition, the feature parameters generated by the feature parameter generation step S30 are input to the analysis step S40.

In the analysis step S40, feature parameters input from the feature parameter generation step S30 and predetermined reference models are generated in syllable units and compared and analyzed.

In the analysis step S40, the input pattern and the reference pattern are non-linearly matched to compensate for the difference between the speed and the length between the input voice and the reference voice, and the voice inputted with the voice of the reference pattern having the highest similarity is recognized (Squared Euclidean Distance) of each of the feature parameters and the reference models is calculated using a dynamic time warping (DTW) algorithm. The reference model having the smallest distance is used as a model most similar to the feature parameter .

The determining step S50 is a step of determining whether there are two or more reference models whose similarity with the feature parameter is within a predetermined range by the analysis step S40.

In other words, in the determination step S50, the Euclidean squared distance between the Euclidean squared input pattern of each of the feature parameters and the reference models and the respective reference patterns is calculated in the analysis step S40. As a result, the Euclidean squares It is determined whether or not there are two or more reference models having distances.

That is, since the present input voice is likely to be recognized as two or more similar voices, a more accurate pattern analysis is required.

For example, when the 'a', 'ka', 'da', and 'ta' are similar to each other due to the similarity of pronunciation, signal patterns are compared to each other only by the dynamic time warping method. There is a possibility.

Accordingly, in the present invention, it is determined whether two or more similar reference models exist in the analysis step S40 in the determination step S50, and pattern analysis having a better recognition rate than the dynamic time warping method is performed again when the reference models are two or more .

That is, if the number of similar reference models is two or more, the determination step S50 proceeds with the phoneme unit pattern analysis step S60, and if the similar reference model is one, the phoneme determination step S70 is performed.

The phoneme unit pattern analysis step S60 separates the speech signal into phonemes and performs a phoneme-by-phoneme pattern comparison algorithm in the same manner as the hidden Markov model.

The Hidden Markov Model is a statistical model that determines hidden parameters from the observed parameters based on the assumption that the modeling system is a Markov process with unknown parameters and is a widely used method in the field of speech recognition The detailed description will be omitted.

The phoneme determining step S70 determines the phoneme according to the pattern analysis result performed in the analyzing step S40 or the phoneme unit pattern analyzing step S60.

That is, the phoneme determining step S70 determines the phoneme corresponding to the voice corresponding to the reference model as the inputted phoneme in the analysis step S40 when the similar reference model is one in the determining step S50, S50), the phoneme having the highest similarity is determined as the input phoneme by the phoneme unit pattern analysis step (S60).

For example, if the user inputs the voice 'a' and determines that the reference models corresponding to 'a' and 'ka' are similar in the analysis step S40, the phoneme unit pattern analysis is performed again in step S60 The phoneme portion of the speech signal is separately processed to perform the hidden Markov model, so that the user determines the initial phoneme 'A' to be recognized as the recognized phoneme. In another example, the user inputs the phoneme ' If the similar reference model is recognized as 'I' in step S40, it is determined that 'b' has been input without going through the phoneme unit pattern analysis performing step S60.

The word determination step S80 is a step of searching for a word using the phonemes detected by the phoneme determination step S70 and selecting the final result among the searched words.

As described above, the speech recognition apparatus 1 according to the present invention can first reduce the number of reference patterns compared through the preliminary speech recognition, thereby saving memory and reducing the computational throughput. In addition, By using the dynamic time warping of the pattern basically, it is possible to increase the accuracy and reliability of the speech recognition without adding excessive load to the system by using the hidden Markov model method for the phoneme unit pattern only when the accuracy is required do.

FIG. 7 is a block diagram for explaining a lighting system to which the speech recognition apparatus of FIG. 2 is applied, FIG. 8 is a block diagram showing a control unit installed in the illumination lamp of FIG. 7, and FIG. 9 is an example of FIG.

The lighting system 900 of FIGS. 7 and 8 includes a control unit 911 installed in each of the illumination lamps 910-1, ..., and 910-N installed at intervals on the ceiling of the illumination place, A controller 920 for managing and controlling each of the controllers 911 of the plurality of light sources 910-1 to 910-N and a controller 920 for controlling and controlling the controllers 920 and 910-1 to 910- And a communication network 930 for providing a data movement path between the control units 911 of the mobile stations 910-N.

The communication network 930 supports data communication between the controller 920 of the controller 920 and the illumination lamps 910-1, ..., 910-N, and more specifically, a Zig- A wired / wireless network such as a local area network (LAN) such as Wi-Fi, Bluetooth, and a wide area network (WAN), and a mobile communication network.

The illumination lamps 910-1, ..., and 910-N are installed in the ceiling structure of the illumination site and are spaced apart from each other. In this case, the illumination lamp includes a light source element such as an LED that emits light, a housing, a cover, a diffusion plate, a circuit board, an electric element and the like, and the configuration of the illumination lamp is a widely used technique and therefore a detailed description thereof will be omitted.

Also, a control unit 911 is installed in each of the lamps 910-1, ..., and 910-N.

At this time, the illumination lamps 910-1, ..., and 910-N perform lighting functions at normal times, and blink according to the emergency evacuation path in an unexpected situation, so that they can perform the functions of emergency lights and guidance lights.

Also, the first, second and third microphones 11-1, 11-2, and 11-3 described above are installed on the outer surfaces of the respective lamps 910-1, ..., and 910, The acoustic signals collected by the first, second and third microphones 11-1, 11-2, and 11-3 are input to the controller 911. [

9, the control unit 911 includes an acoustic signal input unit 3, a speech recognition unit 5, a feature parameter detection unit 6, a comparison and matching unit (not shown) for performing the same configuration and operation as those of FIG. 7, a reference model database unit 8, and a word determination unit 9.

The control unit 911 includes a light management unit 912 for managing and controlling the lighting of the LEDs, a communication interface unit 913 for transmitting and receiving data to and from the controller 920, A matching table that matches the control data of the actual lighting for each of the illumination control related characters and the illumination control related comparison characters that can be determined as characters for controlling the illumination light And a memory 914 for searching for a voice word by the word determining unit 9 and detecting a relation between each of the determined voice words and the 'unexpected comparison target characters' And a determination unit 951 that determines that an unexpected event has occurred when the detected association is greater than a threshold value.

If it is determined that an unexpected condition has occurred, the determination unit 951 inputs unexpected confirmation data indicating that an unexpected condition has occurred in the unlit state management unit 912. When the unlit state management unit 912 receives the unexpected data, And the light source element is turned on for a predetermined time t.

Further, when the voice word is determined by the word determination unit 9, the determination unit 951 also detects the association between each of the determined voice words and the 'illumination control related comparison characters' stored in the memory, If the threshold value is exceeded, the matching table is searched and the control data corresponding to the 'illumination control-related comparison object character' is detected and input to the light-off management unit.

For example, the 'illumination control related characters to be compared' may include 'light', 'light', 'raise', 'lower', etc., and the control data of the illumination may be turned off Control data for turning on the power when 'no' is detected, control data for increasing the illumination intensity by one step when 'high' is detected, and control data for increasing the light intensity when 'low' is detected And step-down control data.

The control unit 911 also controls the communication interface unit 913 to transmit the unexpected status confirmation data to the controller 920 when the determination unit 915 determines that an unexpected condition has occurred. At this time, the control unit 911 matches the character to be compared, whose association exceeds the threshold value, to the determination unit 915 and transmits the matching character to the controller 920.

The control unit 911 receives the light control data from the controller 920 through the communication interface unit 913 and inputs the received light control control data to the light management controller 912 to control the controller 920 To be turned on and off according to the control signal.

In this case, the judgment unit 915 judges whether or not an unexpected situation has occurred in accordance with the method disclosed in Korean Patent No. 10-1625121 filed by the applicant of the present application and entitled " Emergency alarm method using voice recognition, , Its recording medium) was applied.

10 is a block diagram showing the controller of Fig.

The controller 920 in Fig. 10 includes a control unit 921, a memory 922, a communication interface unit 923, an unexpected condition processing unit 924, and an illumination lamp management unit 925.

The control unit 921 is an OS (Operating System) of the controller 920 and manages and controls the control objects 922, 923, 924, and 925.

The control unit 921 receives the unexpected situation confirmation data from the control unit 911 of each of the illumination lamps 910-1 through 910-N via the communication interface unit 923, 924 and the illumination lamp management unit 925. [

In the memory 922, voice words received from the control unit 921 are periodically stored.

The memory 922 also stores the location information of the emergency evacuation route of the illumination site. For example, the memory 922 may store location information of an emergency path where evacuation can be made in the event of a fire.

The memory 922 also stores communication identification information and position information of the control units 911 of the illumination lamps 910-1, ..., and 910-N.

The unexpected situation processing unit 925 is driven when receiving the unexpected situation confirmation data from the control unit 921 of each of the illumination lamps 910-1, ..., and 910-N, At the same time, announcements are output through the speakers.

The illumination lamp management unit 925 is driven when an unexpected situation occurs, and manages lighting of the illumination lights 910-1, ..., and 910-N.

Also, the illumination lamp management unit 925 determines a blinking period defined by a period at which the illumination lights 910-1, ..., and 910-N are turned on and off.

11 is an exemplary diagram showing a blinking period determined by the lamp control unit of Fig.

The illumination light management unit 925 uses the position information of each of the illumination lamps 910-1, ..., and 910-N and the predetermined emergency evacuation path as shown in FIG. 11 when an unexpected situation occurs ..., 910-N corresponding to the movement path of the emergency evacuation route, and determines the blinking sequence of the illumination lights 910-1, ..., 910- N (n seconds) are made.

That is, when the illumination lamps 910-1, 910-2, and 910-3 are installed along the path of the emergency evacuation path, the illumination lamp management unit 925 firstly controls the illumination lights 910-1 ), And thereafter determines the blinking period in which the illumination lamp 910-2 is flickered for a predetermined time (n seconds) and finally the illumination lamp 910-3 is blinked for a predetermined time (n seconds) . At this time, the blinking cycle data determined by the illumination lamp management unit 925 is transmitted to the control units 911 of the illumination lights 910-1, ..., and 910-N.

Accordingly, the people at the illuminated place recognize whether or not an unexpected situation occurs by turning off the emergency bell, the skipper, the illumination lights 910-1, ..., 910-N when an unexpected situation occurs, ), ..., (910-N) are visually confirmed in the direction of blinking so that the position of the emergency evacuation route can be easily recognized and the evacuation can be performed quickly.

1: voice recognition device 3: acoustic signal input part
5: voice recognition unit 6: feature parameter detection unit
7: comparison and matching unit 8: reference model database unit
9: word determination unit 51: first speech recognition module
52: second speech recognition module 900: illumination system
910-1, ..., 910-N: illumination lights 911:
912: Lights-out management unit 913:
914: memory 915:
920: Controller 930:

Claims (14)

First, second and third microphones for collecting acoustic signals and converting them into electrical signals;
A reference model database unit storing predetermined reference models;
An acoustic signal input unit receiving acoustic signals obtained by the microphones;
A voice recognition unit for analyzing the acoustic signals input by the acoustic signal input unit and detecting the original signal X2;
A feature parameter generation unit for extracting a feature vector of the original signal X2 detected by the speech recognition unit and generating a feature parameter using the extracted feature vector;
A comparison and matching unit for analyzing the reference models stored in the reference model database unit and the feature parameters generated by the feature parameter generation unit using a predetermined comparison algorithm to detect a reference model having the highest similarity to the feature parameters;
And a word determiner for searching for a word using a character corresponding to the reference model detected by the comparison and matching unit as a search word and finally outputting the searched word to perform speech recognition,
The speech recognition unit
The sound signal of the first microphone is separated into the original signal S1 and the noise signal N1 using a predetermined first signal separation algorithm and the sound signal of the second microphone is converted into the original signal S1 by using the first signal separation algorithm. A first speech recognition module for separating the original signal S1 into a noise signal S2 and a noise signal N2 and for summing the separated original signals S1 and S2 to detect a primary signal X1;
And separating the sound signal of the third microphone into the original signal (S3) and the noise signal (N3) using a predetermined second signal separation algorithm, and then separating the original signal into a first order signal detected by the first speech recognition module Further comprising a second speech recognition module for summing the signal X1 and detecting the final original signal X2,
Wherein the first signal separation algorithm and the second signal separation algorithm separate the original signal (S) and the noise signal (N) from the acoustic signal in different ways. delete The method of claim 1, wherein the first, second, and third microphones are beam-forming techniques, wherein the first microphone forms a beam in a straight-ahead direction, and the second and third microphones form a left- Forming a beam,
Wherein the first, second, and third microphones are applied with Acoustic Echo Cancellation (AEC) for removing a dynamic noise source from input acoustic signals. The speech recognition apparatus of claim 3, wherein when the final original signal (X2) is detected, the speech recognition unit combines the vowel set in the initials of the detected original signal (X2)
The comparing and matching unit
Dynamic Time Warping (DTW) for recognizing the speech of the reference model with the highest similarity by nonlinearly matching the input feature parameter with the reference model to compensate for the difference in the speed and length between the input voice and the reference voice Warping algorithm is used to calculate the Euclidean distance of each feature parameter and reference model. Then, the reference model with the smallest distance is recognized as the model most similar to the feature parameter,
The comparing and matching unit
When there is one reference model whose similarity with the feature parameter is within a predetermined threshold value, the input speech is determined as the reference model with the highest similarity. If there are two or more reference models whose similarity with a specific parameter is within a predetermined threshold, Wherein the input speech is determined as a phoneme having the highest similarity through a pattern comparison algorithm through a hidden Markov model after separating the signal into phonemes. delete delete delete delete delete A lighting system comprising a light source device, a housing, a microphone installed outside the housing, and a lighting unit including a control unit installed inside the housing, the lighting system comprising:
The control unit
A light switch for controlling the light source device;
The control data of the actual illumination for each of the 'illumination control related comparison characters' defined by the characters that can be determined as the characters for controlling the illumination of the voice word and the 'illumination control related comparison characters' And a memory in which the matched matching table is stored,
The control unit analyzes the acoustic signal input from the microphone to determine a voice word. If the determined voice word is any one of the 'illumination control related comparison characters', the control unit searches the matching table to detect the control data The lighting control unit controls the lighting according to the detected control data,
The microphones are composed of first, second and third microphones installed at intervals,
The control unit
A reference model database unit storing predetermined reference models;
An acoustic signal input unit receiving acoustic signals obtained by the first, second and third microphones;
A voice recognition unit for analyzing the acoustic signals input by the acoustic signal input unit and detecting the original signal X2;
A feature parameter generation unit for extracting a feature vector of the original signal X2 detected by the speech recognition unit and generating a feature parameter using the extracted feature vector;
A comparison and matching unit for analyzing the reference models stored in the reference model database unit and the feature parameters generated by the feature parameter generation unit using a predetermined comparison algorithm to detect a reference model having the highest similarity to the feature parameters;
A word determination unit for searching for a word using a character corresponding to the reference model detected by the comparing and matching unit as a search word and finally outputting the searched word to perform speech recognition;
When the detected relation is greater than a threshold value after detecting the association between the voice word determined by the word determination unit and each of the 'illumination control related comparison characters' stored in the memory, the matching table is searched for, Further comprising: a determination unit for detecting control data corresponding to a control-related comparison target character 'and inputting the control data to the light-off management unit;
The speech recognition unit
The sound signal of the first microphone is separated into the original signal S1 and the noise signal N1 using a predetermined first signal separation algorithm and the sound signal of the second microphone is converted into the original signal S1 by using the first signal separation algorithm. A first speech recognition module for separating the original signal S1 into a noise signal S2 and a noise signal N2 and for summing the separated original signals S1 and S2 to detect a primary signal X1;
And separating the sound signal of the third microphone into the original signal (S3) and the noise signal (N3) using a predetermined second signal separation algorithm, and then separating the original signal into a first order signal detected by the first speech recognition module Further comprising a second speech recognition module for summing the signal X1 and detecting the final original signal X2,
Wherein the first signal separation algorithm and the second signal separation algorithm separate the original signal (S) and the noise signal (N) from the acoustic signal in different ways. delete The speech recognition apparatus of claim 10, wherein, when the final original signal (X2) is detected, the speech recognition unit combines the vowel set in the prefix of the detected original signal (X2)
The comparing and matching unit
Dynamic Time Warping (DTW) for recognizing the speech of the reference model with the highest similarity by nonlinearly matching the input feature parameter with the reference model to compensate for the difference in the speed and length between the input voice and the reference voice Warping algorithm is used to calculate the Euclidean distance of each feature parameter and reference model. Then, the reference model with the smallest distance is recognized as the model most similar to the feature parameter,
The comparing and matching unit
When there is one reference model whose similarity with the feature parameter is within a predetermined threshold value, the input speech is determined as the reference model with the highest similarity. If there are two or more reference models whose similarity with a specific parameter is within a predetermined threshold, Wherein the input speech is determined as a phoneme with the highest similarity through a pattern comparison algorithm through a hidden Markov model after separating the signal into phonemes. The memory according to claim 12, wherein the memory further stores 'erroneous comparison-related characters'
The determination unit
When the detected relation is greater than a threshold value after detecting the association between each of the 'spontaneous related comparison characters' stored in the memory and the voice word determined by the word determination unit, it is determined that an unexpected situation has occurred,
Wherein the lighting control unit controls the light source to be turned on in a predetermined manner when the determination unit determines that an unexpected condition has occurred, so that the lighting control unit is used for lighting normally during an unexpected situation, And is used as an emergency light when an unexpected situation occurs. The lighting system according to claim 13, further comprising a controller for managing and controlling the luminaire mechanisms,
Wherein the control unit transmits unexpected condition data indicating that an unexpected condition has occurred to the controller when the determination unit determines that an unexpected condition has occurred,
The controller
The location information of each of the luminaire mechanisms and the predetermined emergency evacuation route, and when the unexpected situation confirmation data is received from any one of the luminaire mechanisms, the location information and the emergency evacuation route of each of the luminaire mechanisms are used And arranging the illumination lamps in the order corresponding to the emergency evacuation path, and then causing the illumination lamps to be flashed in the ordered sequence.

Images