KR20080061901A - Efficient Speech Recognition Method and System by Robot Input / Output Device - Google Patents

Abstract

A method and a system for effectively recognizing voice by an input/output device of a robot are provided to process a non-registered word effectively by filtering out a mute sound or noise, and improve the speed by saving resources and time related to processing of a mute sound. A contact sensor allows a user to perform manipulation. A microphone inputs a voice of a user. A voice recognition system(3) recognizes a user voice inputted via the microphone. A voice recognition system control module controls the voice recognition system according to an event occurring through manipulation of the contact sensor. An LED lamp(5) displays an operational state of the voice recognition system. A speaker outputs recognized voice information through the voice recognition system. A user guide providing unit allows a user to learn a vocalization method indirectly through a user feedback with respect to the voice input results. A microphone volume adjusting function(2) adjusts an input volume of the microphone when voice recognition is successful. A speaker volume adjusting function(7) adjusts the volume of the speaker by manipulating the contact sensor.

Description

Efficient Speech Recognition Method and System by Robot Input / Output Device {System and Method of Effcient Speech Recognition by Input / Output Device of Robot}

1 is a overall configuration of the speech recognition system according to the present invention

2 is a flowchart of a speech recognition engine algorithm according to the present invention.

<Explanation of symbols for the main parts of the drawings>

Touch sensor (1), voice input device (2), voice recognition system (3), voice recognition system control module (4), LED lamp (5), LCD display (6), voice output device (7), user Guide providing unit 8, end point detector 10, feature extractor 11, noise removing filter 12, lexical recognition means 13, post-processing means 14, lexical recognition means 13, grammar file ( Grammar File) (15), Grammar (16), Phoneme Model DB (17), Phoneme Model Dictionary (18)

The present invention relates to an efficient voice recognition method and system by an input / output device of a robot, and more particularly, a voice recognition engine is controlled by an event method by a contact sensor of a robot, and an LED is formed on the ear of the robot to blink the LED. You can see the operating status of the voice recognition engine through the status, LCD display unit is provided to indicate the success or failure after the voice recognition, and transmits the event generated by the touch sensor to the voice recognition engine control module to provide resources for the microphone The present invention relates to an efficient speech recognition method and system by an input / output device of a robot that attempts to recognize speech exclusively and outputs recognized speech.

The service robot in which the conventional voice recognition function is implemented performs a specific behavior by recognizing a user's voice command. At this time, the speech recognition is generally applied to the degree of 'speaker independent isolated word recognition' (recognition of specific words determined regardless of the speaker) or 'speaker independent key word recognition' (recognition of specific words in a sentence) applicable at the current technology level. The situation is being adopted.

In addition, these methods of speech recognition pre-store the corresponding voice DB for the predefined words inside the robot, and then compare the result with the pre-stored word DB of the robot when a voice input of a specific word comes in from the outside. Outputs

Among the various types of AIBO, the model that amplifies the most voice recognition is Party AIBO.In order to recognize many kinds of voices, it is divided into several modes arbitrarily, and then the switch mounted on the sole of the AIBO's sole is set arbitrarily. Press to enter each mode.

Examples of the voice commands for each mode include voice commands of forward, backward, right, left, good and stop in modes related to movement, and voices of kick, down, goal and nice revenue in modes related to soccer. There are commands, and in the game-related modes, voice commands recognized by each mode are different, such as scissors, rocks, and beam commands.

As described above, the reason for separating the voice command for each mode is that if the number of voice commands recognized at one time is large, the accuracy of the voice recognition decreases due to the limitation of the voice recognition technology and the time for performing the voice recognition becomes long.

However, as described above, a method of reducing the number of voices recognized at once by arbitrarily separating the modes is used. However, since it is necessary to press a predetermined switch to set the mode, it is difficult to understand intuitively and also to check the current mode. Or there is a problem that the cost increases because the LED must be attached.

In order to solve these problems, researches on voice recognition technology are continuously conducted. Hereinafter, a look at the prior patent technology for the voice recognition technology.

Japanese Publication No. 1994-034181 relates to a speech recognition method having a low memory capacity or a small amount of calculation of a probability of misrecognition among words which are partially similar, and a means for time-dividing an input speech signal. And a means for obtaining a plurality of time series acoustic parameters from the time-differentiated speech signal, a means for obtaining acoustic parameter distances between viewpoints from the plurality of time series acoustic parameters, and a series of means obtained for obtaining the acoustic parameter distances. Time normalized speech recognition obtained by means for dividing an acoustic parameter distance into a predetermined length and obtaining acoustic parameters of the splitting point as the plurality of time series acoustic parameters by speech recognition parameters, and means for obtaining the speech recognition parameters. Parameters and Previously Registered Standards Means for matching process for La and m, and is characterized in that means for outputting a result of the matching process,

Korean Patent Application No. 10-2004-0109128 relates to an intelligent robot voice recognition service apparatus and method using a language model and a dialogue model for each region, and includes a voice recognition unit for performing voice recognition on a user's voice query, A conversation processing unit for generating a response sentence using a conversation model corresponding to a voice query, a speech synthesis unit for converting the response sentence into a voice and outputting it to a user, and a display unit for receiving the response sentence from the conversation processing unit and displaying the response sentence to the user. It features

Korean Patent Application No. 10-2001-0005875 relates to a robot control apparatus and method having a voice recognition function that provides children with more attachment and interest in robot toys. Voice recognition means for receiving and recognizing a voice of a person, a microprocessor unit for controlling the overall operation of the robot by analyzing the recognized voice, a pulse width modulation means for generating a pulse width modulation signal under the control of a microprocessor unit, and a pulse width modulation Motor driving means for controlling the operation of the robot according to the signal, to detect and prevent risk factors such as collision and fall when the robot performs the movement operation, and to detect when the robot falls to the left and right Infrared sensor group that provides signals to the microprocessor unit, voice corresponding to the operation of the robot Includes voice generation means for outputting a call to the speaker, and thus is configured to control the operation of the robot by voice command,

Korean Patent Application No. 10-2001-0059001 relates to a security device and method for a robot toy having a voice recognition function, and includes a transmission device for transmitting a code value and a security mode data for a voice command, and data transmitted and installed in a robot. First high frequency receiving means for receiving; Infrared sensor group for detecting the collision of the robot, fall and fall to the left, right, human body sensing means for detecting the human body when the robot moves freely, data received from the first high-frequency receiving means and infrared sensor group and human body sensing means A microprocessor that controls the overall operation of the robot based on the detection signal obtained from the robot, and drives a plurality of motors installed in the joints and neck joints of each leg of the robot under the control of the microprocessor to perform the free movement and movement response of the robot Motor driving means, high-frequency transmission means for receiving alarm data from the microprocessor and transmitting the alarm data when the human body is detected while the security mode is set. An alarm to send a voice message by dialing a number, It is configured to control the operation of the robot by sex command,

Korean Patent Application No. 10-2002-0013641 subdivides the gain level of a voice signal to perform voice recognition by selecting a voice signal without clipping and having a high signal noise ratio, thereby maximizing the voice recognition rate without being constrained by the distance of the sound source. The present invention relates to a voice recognition device and a method for an intelligent robot, comprising: a microphone part configured to have a predetermined number on each of the left and right sides, and to amplify the voice signal by a plurality of fixed gains, each of which is divided into a predetermined level by receiving a voice signal; An analog / digital converter for converting a plurality of output amplified signals into digital signals, and a buffer unit for storing a plurality of digital signals respectively output from the analog / digital converter; It comprises a voice processing unit for receiving a plurality of digital signals stored in the buffer unit for processing the voice signal and a voice recognition unit for recognizing a voice command by the signal output from the voice processing unit,

Korean Patent Application No. 10-2001-0080831 relates to a voice recognition method of a robot that enables effective recognition of a user's command by distinguishing and implementing a voice recognition DB according to the attitude of the robot. When the voice command is stored in the first step of switching to the voice recognition mode, the second step of determining the current posture of the robot to detect the word DB according to the posture, and the voice command from the words of the detected word DB If a matched word exists, the third process returns to the autonomous mode after executing the command with reference to the current robot state.

As described above, voice recognition is limited to words registered in advance due to the limitation of current voice recognition technology. In addition, noise recognition robust speech recognition requires effective concentration of resources necessary for recognition and computation.

In order to solve the above problems, rather than attempting to improve the performance of the relatively slow speech recognition engine itself, the present invention generally provides information on the start / end of the speech engine, which is various devices mounted on a robot. Or a touch sensor for minimizing errors, an ear LED that informs the timing of speech expression, facial expression and speech synthesizer based on speech recognition results, and a microphone / speaker system according to external noise or user's voice size. In order to provide an efficient voice recognition method and system by the input / output device of a robot, the purpose is to provide an efficient method.

According to the present invention, the voice recognition engine is controlled by an event method by a contact sensor of the robot, and the operation state of the voice recognition engine can be known through the blinking state of the LED attached to the robot's ear, and the event generated by the touch sensor The present invention relates to an efficient voice recognition method and system by an input / output device of a robot, which transmits to a recognition engine control module to perform a voice recognition by monopolizing resources for a microphone and outputs a result of the recognized voice on an LCD screen.

In addition, the present invention by adding a microphone volume control function (2) to adjust the microphone input volume, and a speaker volume control function (7) to adjust the volume of the speaker,

Upon successful voice input, if the input voice is a little low, the microphone volume is slightly increased. If the voice volume is a little louder, the microphone volume is slightly lowered to allow adaptive voice recognition according to the situation.

When the voice is recognized, the sound from the speaker (including the voice) is minimized to re-input the microphone to reduce external noise.

An efficient voice recognition system by an input / output device of a robot according to the present invention includes a touch sensor 1 operable by a user, a microphone (not shown) for user's voice input, and a user voice input through the microphone. The voice recognition system 3, the voice recognition system control module 4 for controlling the voice recognition system according to an event generated through the operation of the touch sensor 1, and the operating state of the voice recognition system 3 LED lamp 5 for displaying, a speaker (not shown) for outputting voice information recognized through the voice recognition system 3, and a user feedback on a voice input result to indirectly learn voice utterance. It consists of a user guide providing unit (8),

Microphone volume control function (2) for adjusting the input volume of the microphone upon successful voice recognition, and speaker volume control function (7) for adjusting the volume of the speaker by the touch sensor (1) is configured do.

If necessary, it may be configured to include a voice recognition result display means 6, such as an LCD, so as to display the success or failure after the voice recognition ends.

Speech recognition system (3) is characterized in that the speaker independent isolated word recognition, speech DB matching method, the entire sentence is treated as a single isolated word and divided into registered words, non-registered words,

In more detail, an endpoint detector 10 for detecting only a speech section from an input signal, a feature extractor 11 for extracting a feature parameter effective for recognition from the speech signal detected through the endpoint detector 10, and an input; A noise removing filter 12 for removing a factor (breathing sound, lips sound, etc.) that degrades recognition performance among voice signals, lexical recognition means 13 for checking whether the input voice signal corresponds to the recognition vocabulary, and the recognition process. Post-processing means 14 for determining whether to accept or reject the recognized result using the confidence measure of the similarity of the regular model (general recognition model) aligned by the anti-model (pillar model). It is configured to include).

The LED lamp (5) is mounted on the robot's ear is turned on / off according to the user's touch sensor operation, to provide an accurate time point for the start of speech.

The LCD display unit 6 is in charge of the process and the result of the speech recognition. For example, a facial expression when a voice recognition is attempted, a facial expression when a voice recognition success and a voice recognition failure are output.

The user guide is configured to indirectly learn the speech method with user feedback on the results. For example, I'm not sure, the voice is too loud, the voice is too small, and so on.

Looking at the efficient voice recognition method by the input and output device of the robot according to the present invention, the user generating a voice recognition event through the operation of the touch sensor of the robot, the operating state of the voice recognition engine controlled according to the voice recognition event Flashing the LED lamp formed on the robot's ear so as to know the signal; and transmitting the voice recognition event to the voice recognition engine control module to attempt to recognize the voice by monopolizing the resources for the microphone; A voice input step of inputting a voice command through the unit;

When a voice is input through the voice input step, a voice section detection step of detecting only a voice section from the input signal, and a feature extraction step of extracting a feature parameter valid for recognition from the input voice signal. , Noise elimination step of eliminating factors (respiratory sounds, lip sounds, etc.) deteriorating recognition performance of input voice signal, lexical recognition step of checking whether input voice signal corresponds to recognition vocabulary, positive model sorted by recognition process The post-processing step of determining whether to accept or reject the recognized result using the confidence measure of the similarity of the general recognition model and the similarity by the anti-model. And an output step of outputting a word.

The present invention as described above operates the start and end of voice recognition in the event method by the touch sensor. While the voice recognition is in progress, it is displayed through the LED of the ear so that the user can know the operation of the voice recognition engine. In addition, after the voice recognition ends, the LCD window indicates success.

In addition, the event generated by the touch sensor is transmitted to the voice recognition engine control module, and the voice recognition is attempted by monopolizing the resources for the microphone, and the voice synthesis stop command is given to the voice synthesis engine control module.

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

1 is an overall configuration diagram of a voice recognition system according to the present invention, Figure 2 is a detailed configuration diagram of a voice recognition system according to the present invention, the touch sensor (1), voice input device (2), voice recognition system (3) ), Voice recognition system control module (4), LED lamp (5), LCD display (6), voice output device (7), user guide provider (8), endpoint detector (10), feature extractor (11) ), Noise reduction filter 12, lexical recognition means 13, post-processing means 14, lexical recognition means 13, Grammar File 15, Grammar 16, Phoneme Model DB 17 , Phoneme model dictionary 18.

As shown in FIG. 1, an efficient voice recognition system by an input / output device of a robot according to the present invention includes a touch sensor 1 operable by a user, a microphone (not shown) for user's voice input, and the microphone. A voice recognition system 3 for recognizing a user voice input through the voice recognition system, a voice recognition system control module 4 for controlling the voice recognition system according to an event generated through manipulation of the touch sensor 1, and the voice recognition LED lamp 5 indicating the operating state of the system 3, a speaker (not shown) for outputting voice information recognized through the voice recognition system 3, and a voice feedback method with user feedback on voice input results. Consists of a user guide providing unit (8) to indirectly learn,

Microphone volume control function (2) for adjusting the input volume of the microphone when the voice recognition is successful, and speaker volume control function (7) for adjusting the volume of the speaker by the touch sensor (1) is configured to include It is characterized by.

If necessary, it may be configured to include a voice recognition result display means 6, such as an LCD, so as to display the success or failure after the voice recognition.

The speech recognition system (3) is characterized in that the speaker independent isolated word recognition, speech DB matching method, the entire sentence is processed as one isolated word and divided into registered words, non-registered words,

In more detail, as shown in FIG. 2, an end point detector 10 for detecting only a voice section from an input signal and a feature parameter valid for recognition from a voice signal detected through the end point detector 10 are extracted. An extractor 11, a noise removing filter 12 that removes factors (such as breathing sounds, lip sounds, etc.) that reduce recognition performance among the input voice signals, and lexical recognition means for checking whether the input voice signal corresponds to the recognition vocabulary. (13) and accept or reject the recognized result using the confidence measure of the similarity of the regular model (general recognition model) aligned by the recognition process and the similarity by the anti-model (pillar model). It comprises a post-processing means 14 for determining the.

The lexical recognition means 13 searches the grammar 16 and the phonemes stored in the phoneme model DB 17 from the grammar file 15 to recognize the vocabulary.

The microphone volume control function (2) is a function to adjust the microphone input volume when the voice recognition is successful, if the voice volume input when the success is a little low, while adjusting the microphone volume a little higher, If it is judged that the voice volume is a little louder, the microphone volume is adjusted a little lower so that it is adapted to the situation for effective voice recognition.

The speaker volume control function (7) is a function that allows the user to control the volume of the speaker by operating the touch sensor (1), by minimizing the phenomenon that the sound coming from the speaker (including voice) is re-input into the microphone during voice recognition It is a function to reduce external noise. In practice, the speaker volume is zero.

Looking at the efficient voice recognition method by the input and output device of the robot according to the present invention, the user generating a voice recognition event through the operation of the contact sensor of the robot, the operating state of the voice recognition engine controlled according to the voice recognition event Flashing the LED lamp formed on the robot's ear so as to know the signal; and transmitting the voice recognition event to the voice recognition engine control module to attempt to recognize the voice by monopolizing the resources for the microphone; A voice input step of inputting a voice command through the unit;

When a voice is input through the voice input step, a voice section detection step of detecting only a voice section from the input signal, and a feature extraction to extract a feature parameter valid for recognition from the input voice signal. Step, a noise removing step of removing a factor (breathing sound, lip sound, etc.) degrading recognition performance among input voice signals, a lexical recognition step of checking whether the input voice signal corresponds to a recognition vocabulary, and a recognition process. A post-processing step of determining whether to accept or reject the recognized result using the confidence measure of the similarity of the aligned regular model (general recognition model) and the similarity by the anti-model (pillar model), and voice An output step of outputting a corresponding word upon successful recognition,

If necessary, the operation may be performed by displaying the success or failure of the voice recognition on the voice recognition result display unit of the robot.

The speech section detection step is a process of detecting only the speech section from the input signal, which is a very important process because it is the first process to deliver the signal used for speech recognition. It must be implemented with an algorithm that requires near processing speed. We extract statistical features that are relatively simple in the calculation process from the input signal and distinguish between speech and silence sections. A commonly used method is a method of determining the voice interval and the silent interval by obtaining an energy value in every section of the input signal and comparing it with a threshold predetermined by statistics. This method shows excellent performance in relatively high signal-to-noise ratio (SNR) environments, but degrades significantly in noisy environments. The Zero Crossing Rate (ZCR) along with the energy value is obtained and used as a feature. This is to consider the frequency characteristics. As ambient noise changes over time, it is essential to adjust the threshold accordingly. For reliable recognition, it is important to minimize false rejection of speech sections as silence sections, even if the silence section tends to be false acceptance. Therefore, false acceptance ), It should include a recognition process that can be treated as an Out Of Vocabulary or classified as either silence or noise.

The feature extraction step is a process of extracting a feature parameter valid for recognition from an input voice signal. When the same word is pronounced by several people, even though the meaning of the word is the same, the voice waveforms are not the same. Even if the same person pronounces the same word continuously at the same time, the voice waveforms are not the same. The reason for this phenomenon is that the speech waveform includes not only the meaning information of the voice but also information such as the tone of the speaker and the emotional state. Therefore, feature extraction of speech is a process of extracting information representing semantic information from speech.

The noise removing step is a process of removing a factor (breathing sound, lips sound, etc.) that degrades the recognition performance of the input voice signal. In general, factors that degrade recognition performance include an out-of-vocabulary (OVO) problem or a noise problem rather than a recognized vocabulary. Noise can be categorized into background noises based on the surrounding environment, and vocal noises such as breath sounds and lip sounds. Background noise can be solved by the noise canceling function when extracting feature parameters. Therefore, the recognition rate can be improved by applying the human noise filler model.

The post-processing step determines whether to accept or reject the recognized results using the confidence measure of the similarity of the normal model (general recognition model) aligned by the recognition process and the similarity by the anti-model (pillar model). To decide. Reliability is a measure of how reliable the result is for a speech recognition result, and means a relative value of a phoneme or word that is a recognized result and the probability that the word is spoken from other phonemes or words. If the confidence value is larger than a predetermined threshold, the recognition result is accepted. In general, the regular model is used as a tri-phone and the anti-model is used as a context-independent model. The model also uses a Gaussian Mixture Model. This post-processing step may prevent early recognition as well as misrecognition by outputting the most similar words when the user speaks out of Vocabulary (OOV) that is not in the recognition vocabulary.

In order to solve the above problems, rather than attempting to improve the performance of the relatively slow speech recognition engine itself, the present invention generally provides information on the start / end of the speech engine, which is various devices mounted on a robot. Or a touch sensor for minimizing errors, an ear LED that informs the timing of speech expression, facial expression and speech synthesizer based on speech recognition results, and a microphone / speaker system according to external noise or user's voice size. In addition, the robot's input / output device enables efficient voice recognition. In addition, in order to increase the probability of success of speech recognition, considering the characteristics of the algorithm that should mitigate the classification threshold of misrecognition, the probability of speech recognition is reduced while increasing the probability of speech recognition as much as possible. It can be effectively processed and speed can be achieved by saving resources and time related to the silent process.

Claims (10)

In an efficient voice recognition system by an input / output device of a robot, a touch sensor (1) operable by a user, a microphone (not shown) for user's voice input, and voice recognition to recognize a user's voice input through the microphone The system 3, a voice recognition system control module 4 for controlling the voice recognition system according to an event generated through the operation of the contact sensor 1, and an LED indicating an operating state of the voice recognition system 3; A user guide system for indirectly learning a speech method using a lamp 5, a speaker (not shown) for outputting voice information recognized through the voice recognition system 3, and user feedback on a voice input result. It consists of study (8), Microphone volume control function (2) for adjusting the input volume of the microphone when the speech recognition is successful, and speaker volume control function (7) for adjusting the volume of the speaker by the touch sensor (1) is provided. Efficient voice recognition system by the input and output device of the robot, characterized in that. The voice recognition system of claim 1, further comprising a voice recognition result display means (6) such as an LCD to display success after the voice recognition ends. 2. The speech recognition system (3) according to claim 1, wherein the speech recognition system (3) extracts a feature parameter effective for recognition from an end point detector (10) for detecting only a speech section from an input signal and a speech signal detected through the end point detector (10). A feature extractor 11, a noise removing filter 12 for removing a factor (respiratory sound, lip sound, etc.) deteriorating the recognition performance among the input voice signals, and a vocabulary for checking whether the input voice signal corresponds to the recognition vocabulary. Accept or reject the recognized result using the similarity between the recognition means 13 and the positive model (general recognition model) arranged by the recognition process and the confidence measure of the similarity by the anti-model (pillar model). Efficient speech recognition system by the input and output device of the robot, characterized in that it comprises a post-processing means for determining whether to do. The method of claim 3, wherein the lexical recognition means (13) retrieves the grammar (16) from the grammar file (15) and the phonemes stored in the phoneme model DB (17) through a phoneme model dictionary (18) to recognize the vocabulary. Efficient voice recognition system by the input and output device of the robot, characterized in that. In the efficient voice recognition method by the input and output device of the robot, the user generating a voice recognition event through the operation of the touch sensor of the robot, so that the user can know the operating state of the voice recognition engine controlled according to the voice recognition event Flashing the LED lamp formed on the robot's ear, transmitting the voice recognition event to the voice recognition engine control module, and attempting voice recognition by monopolizing the resources for the microphone; Voice input step of inputting, When a voice is input through the voice input step, a voice section detection step of detecting only a voice section from the input signal, and a feature extraction to extract a feature parameter valid for recognition from the input voice signal. Step, a noise removing step of removing a factor (breathing sound, lip sound, etc.) degrading recognition performance among input voice signals, a lexical recognition step of checking whether the input voice signal corresponds to a recognition vocabulary, and a recognition process. A post-processing step of determining whether to accept or reject the recognized result using the confidence measure of the similarity of the aligned regular model (general recognition model) and the similarity by the anti-model (pillar model), and voice Efficiency by the input / output device of the robot, characterized in that it comprises an output step of outputting a corresponding word upon successful recognition The method of voice recognition. The method according to claim 5, wherein the method comprises the step of displaying the success or failure of the voice recognition on the voice recognition result display means of the robot. The method of claim 5, wherein the detecting of the speech segment uses a method of distinguishing between the speech segment and the silent segment by extracting statistical features having a relatively simple calculation process from the input signal. . [6] The method of claim 5, wherein the detecting of the speech section further comprises a recognition process to process the silence section into a word that is not in the dictionary or to distinguish between silence or noise when the silence section is divided into a speech section. Efficient voice recognition method by robot input / output device. The method of claim 5, wherein the noise removing step applies a human noise filler model. The method according to claim 5, wherein the post-processing step uses the reliability measure of the similarity of the positive model aligned by the recognition process and the confidence measure of the similarity by the anti-model.

Images