KR100855592B1 - Apparatus and method for robust speech recognition of speaker distance character - Google Patents

Abstract

The present invention is to provide a voice recognition apparatus and method that is robust to the speaker distance characteristics, and a distance-specific voice recording unit for receiving the voice input by the near-field voice recording unit and the remote voice recording unit at the same time; An external noise removing unit for receiving the distance-specific speech output from the distance-based voice recording unit and estimating the external noise to remove from the recording voice; An input voice selecting unit which receives the recorded voice from which the external noise has been removed and receives and selects which voice can increase the voice recognition performance among the input voices reflecting the distance characteristics of the remote and short distances; By including the voice selected by the input voice selector to perform the voice recognition; speech recognition unit comprising a, the long-distance speech recognition performance and the near-field speech recognition performance is high at the same time to be able to be robust to external noise. .

Speaker, distance characteristic, voice recognition, voice recording, external noise

Description

Apparatus and method for robust speech recognition of speaker distance character}

1 is a block diagram of a speech recognition apparatus robust to speaker distance characteristics according to an embodiment of the present invention.

FIG. 2 is a detailed block diagram of a voice recording unit for each distance in FIG. 1.

3 is a detailed block diagram of the external noise removing unit of FIG. 1.

FIG. 4 is a detailed block diagram of the input voice selector of FIG. 1.

FIG. 5 is a detailed block diagram of the voice recognition unit of FIG. 1.

6 is a flowchart illustrating a voice recognition method that is robust to speaker distance characteristics according to an embodiment of the present invention.

FIG. 7 is a detailed flowchart illustrating a voice recording operation for each distance in FIG. 6.

FIG. 8 is a detailed flowchart illustrating an external noise removing operation of FIG. 6.

9 is a detailed flowchart illustrating an operation of selecting an input voice in FIG. 6.

FIG. 10 is a detailed flowchart of the voice recognition operation of FIG. 6.

Explanation of symbols on the main parts of the drawings

100: voice recording unit by distance 101: recording voice size classification unit

102: sample speaker input gain calculation unit 103: real user input gain calculation unit

104: input gain designation unit 110: short-range voice recording unit

120: remote voice recording unit 200: external noise removing unit

201: time base feature extractor 202: time base voice detector

203: frequency converter 204: voice frequency characteristic extraction unit

205: noise frequency characteristic extraction unit 206: frequency noise removal unit

207: frequency axis voice detection unit 208: voice detection result selection unit

210: first external noise removing unit 220: second external noise removing unit

300: input voice selector 301: first sound accuracy accuracy confidence value selector

302: First sound quality distortion value selector 303: First SNR numerical selector

304: second sound quality accuracy confidence value selecting unit 305: second sound quality distortion value selecting unit

306: second SNR numerical selection unit 307: voice preselection

400: voice recognition unit 401: feature extraction unit

402: decoder 403: post-processing unit

TECHNICAL FIELD The present invention relates to speech recognition, and more particularly, to a speech recognition apparatus and a method which is robust to a speaker distance characteristic such that the distance speech recognition performance and the near speech recognition performance are both high and suitable to be robust to external noise.

In general, speech recognition refers to speech as a machine, and the most common method is to extract and separate a sound region or an equivalent feature that characterizes a vowel by frequency analysis of the sound wave. . The sonagram is a continuous record of the results of frequency analysis. When trained well, the patterns recorded there can be visually recognized. But if you talk too much or talk too much, you may get errors.

Conventional food recognition devices generally recognize speech that is spoken at a short distance due to external noise, but recently, microphone recognition technology is applied to speech recognition even at a long distance. That is, in the prior art, there is a technology that enables remote speech recognition using a remote microphone and a microphone array technology using a plurality of microphones.

However, far speech recognition technology has weaknesses such as low near speech recognition performance and weakness of external noise. In addition, the microphone array technology has the advantage of being robust against noise, but also has a weakness in short-range speech recognition, an expensive hardware device, and an installation problem that requires complicated and sophisticated system configuration.

As such, the conventional speech recognition apparatus is affected by the distance that can be recognized because of the size of the speech input.

In addition, the far field recognizer has a large speech input that is too large for the near field input, so that the recognition rate decreases significantly. there was.

Therefore, the present invention has been proposed to solve the conventional problems as described above, and an object of the present invention is to provide a speaker distance characteristic capable of achieving both a long distance speech recognition performance and a near speech recognition performance at the same time and being robust to external noise. The present invention provides a robust speech recognition device and a method thereof.

In order to achieve the above object, the voice recognition device robust to the speaker distance characteristic according to an embodiment of the present invention,

A distance-specific voice recording unit for simultaneously receiving and recording voice input by the near voice recording unit and the remote voice recording unit; An external noise removing unit for receiving the distance-specific speech output from the distance-based voice recording unit and estimating the external noise to remove from the recording voice; An input voice selecting unit which receives the recorded voice from which the external noise has been removed and receives and selects which voice can increase the voice recognition performance among the input voices reflecting the distance characteristics of the remote and short distances; And a voice recognition unit for receiving a voice selected by the input voice selection unit and performing voice recognition.

In order to achieve the above object, the voice recognition method robust to the speaker distance characteristic according to an embodiment of the present invention,

A first step of allowing voice recording units by distance to simultaneously receive and record voice input from near and far distances; A second step of estimating external noise from distance-specific voices and removing them from the recorded voice after the first step; After the second step, the input voice selecting unit receives the recorded voice from which the external noise has been removed and checks and selects which voice can improve the voice recognition performance among the input voices reflecting the distance characteristics of the remote and short distances. Wow; And a fourth step of performing voice recognition by receiving the voice selected by the voice recognition unit after the third step.

Hereinafter, an exemplary embodiment according to the present invention as described above, a voice recognition device robust to speaker distance characteristics, and a technical idea of the method will be described with reference to the accompanying drawings.

1 is a block diagram of a speech recognition apparatus robust to speaker distance characteristics according to an embodiment of the present invention.

As shown therein, the voice recording unit 100 for each distance for receiving and recording voices input by the near voice recording unit 110 and the far voice recording unit 120 at the same time; An external noise removing unit 200 which receives the distance-specific voice output to the distance-based voice recording unit 100 and estimates external noise to remove from the recorded voice; An input voice selecting unit which receives the recorded voice from which the external noise has been removed from the external noise removing unit 200 and checks and selects which voice can improve the voice recognition performance among the input voices reflecting the distance characteristics of the far and near distances; 300); And a voice recognition unit 400 which receives voices selected by the input voice selection unit 300 and performs voice recognition.

FIG. 2 is a detailed block diagram of a voice recording unit for each distance in FIG. 1.

As shown in the drawing, the distance-based voice recording unit 110 and the distance-based voice recording unit 120 of the distance-specific voice recording unit 100 respectively receive a sampler's voice to classify the size of the recording voice. Recording voice size classification unit 101; A sampler input gain calculation unit 102 which receives the output of the recording voice size classifying unit 101, calculates an input gain of a sample speaker, outputs an input gain value according to a distance-specific mode, and performs preset parameter setting; An actual user input gain calculator 103 configured to receive a voice of the actual user, calculate an input gain of the actual user, and set a parameter of the actual user; The sample speaker input gain calculator 102 receives an input gain value of a mode for each distance, receives an input gain of an actual user output from the real user input gain calculator 103, and specifies an input gain to determine a distance. And an input gain designation unit 104 for outputting a star voice.

3 is a detailed block diagram of the external noise removing unit of FIG. 1.

As shown in the drawing, the external noise removing unit 200 receives a time-based feature extracting unit 201 for extracting a feature of a time-base by receiving a voice by distance of a short distance or a long distance output from the voice recording unit 100 for each distance. )Wow; A time axis voice detector 202 which receives the output of the time axis feature extractor 201 and detects voice on the time axis; A frequency converter 203 which receives voices of distances or distances output from the distance-based voice recording unit 100 and converts them into frequencies; A voice frequency characteristic extractor 204 which receives the output of the frequency converter 202 and extracts a voice frequency characteristic; A noise frequency characteristic extractor 205 which receives the output of the frequency converter 202 and extracts a noise frequency characteristic; A frequency noise canceller 206 for receiving the outputs of the voice frequency characteristic extractor 204 and the noise frequency characteristic extractor 205 and removing noise characteristics at a frequency side; A frequency axis voice detector 207 which receives the output of the frequency noise remover 206 and detects voice on a frequency axis; And a voice detection result selector 208 which receives the output of the time axis voice detector 202 and the frequency axis voice detector 207 and selects a voice detection result based on a noise characteristic to output voice. It features.

FIG. 4 is a detailed block diagram of the input voice selector of FIG. 1.

As shown in the drawing, the input voice selector 300 receives the recorded voice from which the external noise has been removed from the external noise remover 200, and selects a loudness confidence confidence value for the near-range voice output. A first speech intelligibility confidence level selector 301 which calculates the sound clarity by the clarity of the volume and pitch information of the recording volume; A first sound quality distortion value selecting unit which selects the sound quality distortion value by the degree of causing the discontinuous frequency characteristic in the adjacent voice section when analyzing the frequency characteristic by receiving the output of the first sound accuracy accuracy confidence value selecting unit 301; 302; A first SNR value selector 303 which receives an output of the first sound quality distortion value selector 302 and selects a signal to noise ratio (SNR) value which is a relative magnitude value of noise; The external noise removing unit 200 receives the recorded voice from which the external noise has been removed and selects a loudness confidence confidence value for the far-field voice output, and the sound having a clear voice is calculated based on the magnitude of the recorded volume and the clarity of the pitch information. 2 sound accuracy accuracy confidence level selection unit 304; A second sound quality distortion level selector which receives the output of the second sound accuracy accuracy confidence value selector 304 and selects a sound quality distortion value by a degree of causing discontinuous frequency characteristics in the adjacent voice section when analyzing frequency characteristics; 305; A second SNR value selector 306 that receives an output of the second sound quality distortion value selector 305 and selects an SNR value that is a relative magnitude of noise; And a voice selecting unit 307 that receives the outputs of the first and second SNR value selecting units 303 and 306 and selects a voice input having a high recognition rate and outputs a selected voice.

FIG. 5 is a detailed block diagram of the voice recognition unit of FIG. 1.

As shown therein, the voice recognition unit 400 includes: a feature extracting unit 401 for receiving a voice selected by the input voice selecting unit 300 and extracting unique characteristics for each voice value; A decoder 402 which receives the output of the feature extractor 401 and decodes the speech grammar and the acoustic model; And a post-processing unit 403 for receiving the output of the decoder 402 and performing post-processing in consideration of linguistic characteristics and clarity of speech timing to output a recognition result.

6 is a flowchart illustrating a voice recognition method that is robust to speaker distance characteristics according to an embodiment of the present invention.

As shown therein, the first step ST1 allows the voice recording unit 100 for each distance to simultaneously receive and record the voice input from the near and the far distances; A second step (ST2) of estimating external noise from distance-specific voices and removing them from the recorded voice after the first step; After the second step, the input voice selecting unit 300 receives the recorded voice from which the external noise is removed and checks and selects which voice can improve the voice recognition performance among the input voices reflecting the distance characteristics of the far and near distance. Performing a third step ST3; And a fourth step (ST4) of receiving a voice selected by the voice recognition unit 400 after the third step, and performing voice recognition.

FIG. 7 is a detailed flowchart illustrating a voice recording operation for each distance in FIG. 6.

As shown therein, the first step includes: an eleventh step ST11 of classifying the voice quality of the recorded voice by receiving the voice of the sampler 100; A twelfth step (ST12) of calculating the input gain of the sampler after the eleventh step and outputting an input gain value according to a distance-specific mode (ST12); A thirteenth step (ST13) of receiving a voice of the real user, calculating input gain of the real user, and setting parameters of the real user; A fourteenth step (ST14) of receiving an input gain value of the distance-specific mode in the twelfth step, receiving an input gain value of the real user in the thirteenth step, and specifying an input gain to output voice for each distance; It characterized in that to perform including.

FIG. 8 is a detailed flowchart illustrating an external noise removing operation of FIG. 6.

As shown in the drawing, the second step may include a twenty-first step (ST21) of extracting a characteristic of a time axis by receiving a voice of a short distance or a long distance output from the voice recording unit 100 for each distance; A twenty-second step (ST22) of detecting voice on a time axis after the twenty-first step; A twenty-third step (ST23) of receiving the voice of each distance of a short distance or a long distance output from the voice recording unit 100 for each distance and converting the voice to frequency; A twenty-fourth step (ST24) of extracting characteristics of voice frequency and noise frequency after the twenty-third step; A twenty-fifth step (ST25) for removing a noise characteristic on the frequency side after the twenty-fourth step; A twenty-sixth step (ST26) for detecting voice on the frequency axis after the twenty-fifth step; And performing a twenty-seventh step (ST27) of selecting a voice detection result by reflecting noise characteristics after performing the twenty-second step and the twenty-sixth step (ST27).

9 is a detailed flowchart illustrating an operation of selecting an input voice in FIG. 6.

As shown in the drawing, the third step receives the recorded voice from which the external noise has been removed by the external noise removing unit 200, and selects a loudness confidence confidence value for the near voice output to record voices having a clear voice. A thirty-first step (ST31) for calculating loudness and clarity of pitch information; A thirty-second step (ST32) of selecting sound quality distortion values by the degree of causing discontinuous frequency characteristics in the adjacent voice section when analyzing the frequency characteristic after the thirty-first step; A thirty-third step (ST33) of selecting an SNR value which is a relative magnitude value of noise after the thirty-second step; The external noise removing unit 200 receives the recorded voice from which the external noise has been removed and selects a loudness confidence confidence value for the far-field voice output, and the sound having a clear voice is calculated based on the magnitude of the recorded volume and the clarity of the pitch information. Step 34 (ST34); A thirty-fifth step (ST35) of selecting a sound quality distortion value by the degree of causing a discontinuous frequency characteristic in an adjacent voice section when analyzing the frequency characteristic after the thirty-fourth step; A 36 th step ST36 of selecting an SNR value which is a relative magnitude value of noise after the 35 th step; And a 37 th step (ST37) of selecting a voice input having a high recognition rate after performing the 33rd step and the 36th step and outputting the selected voice.

FIG. 10 is a detailed flowchart of the voice recognition operation of FIG. 6.

As shown in the drawing, the fourth step includes: a 41st step (ST41) of receiving a voice selected by the input voice selecting unit 300 and extracting a unique characteristic for each voice value; A 42nd step (ST42) of performing decoding by the speech grammar and the acoustic model after the 41st step; And a 43rd step (ST43) of outputting a recognition result by performing post-processing in consideration of linguistic characteristics and clarity of speech timing after the 42nd step.

Referring to the accompanying drawings, preferred embodiments of the voice recognition device and the method robust to the speaker distance characteristics according to the present invention configured as described above are as follows. In the following description of the present invention, detailed descriptions of well-known functions or configurations will be omitted if it is determined that the detailed description of the present invention may unnecessarily obscure the subject matter of the present invention. In addition, terms to be described below are terms defined in consideration of functions in the present invention, which may vary according to intention or precedent of a user or an operator, and thus, the meaning of each term should be interpreted based on the contents throughout the present specification. will be.

First of all, the present invention aims to provide high distance speech recognition performance and near speech recognition performance at the same time and to be robust to external noise. To this end, the present invention allows the user to ensure high voice recognition performance regardless of distance.

Thus, in the present invention, a device for solving a decrease in recognition rate according to distance has a feature of performing a plurality of voice input steps by connecting two or more microphones, and a voice capable of automatically securing stable recognition performance among a plurality of voice inputs. Input can be selected.

For example, if you want to control home appliances using wall pad's built-in voice recognition device in a normal home environment, you can give voice commands from the living room sofa about 3 meters and LCD (Liquid Crystal Display). Simultaneously process voice commands at a close range (0.5m) so that the recognition rate is not reduced.

Figure 1 shows the overall configuration of the present invention, Figure 6 shows the operation flow.

Thus, in the present invention, voice recording is performed by separating into two groups. That is, after receiving a voice input simultaneously to the near voice recording unit 110 and the far voice recording unit 120, the external noise removing unit 200 automatically estimates the noise portion of the input voice portion and removes it from the recorded voice. Go through.

After that, the voice recognition unit 400 performs a voice recognition through the input voice selection unit 300 for checking which voices among the input voices reflecting the respective distance characteristics can improve the voice recognition performance, and outputs the recognition result.

The main application of this voice recognition, which is robust to the distance characteristic, is home appliance control device for home network. For example, the distance from the living room sofa to the wall pad is about three meters, allowing the remote voice recorder to resolve the voice input at this distance. On the other hand, the general user wants to experience various information services directly near the wall pad, so that the voice input device at a close distance of about 0.5 meters that can see the LCD window allows the near voice recording apparatus to respond. If there is a distance difference of more than 2 meters, if you do not respond to the recorded voice by distance, the sound may be recorded too loud and distortion may occur, or the sound may be recorded too small to make the sound price unclear. It becomes impossible.

2 is a detailed configuration diagram of the voice recording unit 100 for each distance, and FIG. 6 shows its operation flow.

Therefore, when the voice is recorded by uttering a natural voice at a specified distance, the recognition rate can be optimized if the size of the digitized recorded voice has a value of 5,000 to 20,000 at 16-bit resolution. Therefore, it is necessary to adjust the input gain so that the distance is recorded in the voice recording unit 100 to be an input size of 5,000 to 20,000, respectively.

First, (a) prior parameter setting is performed to set input gain parameters according to the statistical characteristics of many speaker utterances before the final voice input device is completed. The recording voice size classifying unit 101 classifies all the recorded voices according to size by receiving three recordings from a plurality of sample speakers considering male, female, oar, and small at a specified distance from a normal utterance, a loud utterance, and a small utterance. Find out the recording size level and the maximum recording size level. Using this value, the gain value is calculated through the sampler input gain calculation unit 102 so that all recording inputs are numerical values of 5,000 to 20,000. This value is stored as a parameter value corresponding to three modes (large / medium / small) for each distance.

The parameter value obtained in advance from the sampler is applied to the real user as in (b), and is set to one of three mode parameters for each distance so that the input gain designation unit 104 can specify an appropriate input gain value. Do this.

3 and 7 show the detailed configuration and operation of the external noise removing unit 200.

Therefore, external noise canceling devices are required for short distance and long distance, respectively, but the two devices have the same structure.

On the other hand, even if the voice input is faithfully input by distance, even if the noise input is largely input to the recorded voice, the probability of false recognition or recognition is very high. Therefore, it is important to remove the external noise as much as possible and extract the speech part to increase the recognition performance.

External noise rejection is divided into a frequency noise canceller 206, which is a part for removing noise characteristics on the frequency side, and a time-axis speech detector 202 and a frequency-axis speech detector 207, which simultaneously detect speech parts on the time axis and the frequency axis. It is composed. At the same time, the detected voice section reflects the noise characteristic, and the voice detection result selecting unit 208 selects the final voice detection result.

4 and 9 show the detailed configuration and operation of the input voice selection unit 300 for selecting the input voice.

Therefore, the input voice selector 300 selects a voice having a high voice recognition performance from the detected two voice sections.

Since the sound with a clear voice value is calculated by the clarity of the volume and pitch information of the recording volume, the first and second loudness confidence confidence level selectors 301 and 304 select the loudness confidence confidence value for the near or far speech output, respectively. Select.

Since the sound quality distortion value is calculated by the degree of causing the discontinuous frequency characteristic in the adjacent speech section when the frequency characteristic is analyzed, such as when the recording size exceeds 16 bit resolution, the first and second sound quality distortion value selecting units 302 , 305) select sound quality distortion values for near or far speech output.

In addition, the SNR values selected by the first and second SNR value selectors 303 and 306 correspond to signal-to-noise ratios and correspond to relative magnitude values of noise. The voice selection unit 307 selects a voice input having a high recognition rate by using these three values, namely, sound accuracy accuracy, sound quality distortion, and SNR.

5 and 10 show the configuration and operation of the voice recognition unit 400.

Thus, the feature extraction unit 401 extracts unique characteristics for each sound value.

In addition, the decoder 402 performs recognition using a sound model that has been learned in advance and a speech related rule of an input vocabulary.

In addition, the post-processing unit 403 performs post-processing in consideration of the linguistic characteristics and clarity of the timing of speech.

As described above, the present invention provides both the long distance speech recognition performance and the short range speech recognition performance at the same time, and is robust to external noise.

As described above, the speech recognition apparatus robust to the speaker distance characteristic and the method thereof according to the present invention have a long range speech recognition performance and a near speech recognition performance are high at the same time, and has a robust effect on external noise.

In addition, the present invention can manufacture a device capable of speech recognition regardless of the distance.

In addition, the present invention can remove the noise generated at a long distance by using a noise canceling device enables a scenario with excellent service performance. Such a device is configured to analyze a signal input through two or more microphones and to automatically select a voice input having excellent recognition performance internally, so that the installation of the user is not difficult. Therefore, it can be directly applied to the voice interface of a general home network system or a hair robot.

Although the above has been described as being limited to the preferred embodiment of the present invention, the present invention is not limited thereto and various changes, modifications, and equivalents may be used. Therefore, the present invention can be applied by appropriately modifying the above embodiments, it will be obvious that such application also belongs to the scope of the present invention based on the technical idea described in the claims below.

Claims (10)

A distance-specific voice recording unit for simultaneously receiving and recording voice input by the near voice recording unit and the remote voice recording unit; An external noise removing unit for receiving the distance-specific speech output from the distance-based voice recording unit and estimating the external noise to remove from the recording voice; An input voice selecting unit which receives the recorded voice from which the external noise has been removed and receives and selects which voice can increase the voice recognition performance among the input voices reflecting the distance characteristics of the remote and short distances; A voice recognition unit for receiving a voice selected by the input voice selection unit and performing voice recognition; Robust speech recognition device characterized in that configured to include a speaker distance characteristic. The method according to claim 1, The near voice recording unit and the remote voice recording unit of the distance-based voice recording unit, respectively, A recording voice size classifying unit for receiving a sample speaker's voice and classifying the size of the recording voice; A sampler input gain calculator configured to receive an output of the recording voice size classifier, calculate an input gain of a sample speaker, output an input gain value according to a distance-specific mode, and perform preset parameter setting; A real user input gain calculator configured to receive a voice of the real user, calculate an input gain of the real user, and set a parameter of the real user; The sample speaker input gain calculator inputs an input gain value of a distance-specific mode, receives an input gain value of an actual user output from the real user input gain calculator, and specifies an input gain to output voice for each distance. Gain designation unit; Robust speech recognition device characterized in that configured to include a speaker distance characteristic. The method according to claim 1, The external noise removing unit, A time axis feature extraction unit configured to extract a feature of a time axis by receiving voices of distances from a short distance or a long distance output from the distance-based voice recording unit; A time axis voice detector which receives the output of the time axis feature extractor and detects a voice on a time axis; A frequency converter configured to receive voices of distances or distances output from the distance-based voice recording units and convert them into frequencies; A voice frequency characteristic extractor which receives the output of the frequency converter and extracts a voice frequency characteristic; A noise frequency characteristic extractor configured to receive an output of the frequency converter and extract a noise frequency characteristic; A frequency noise canceller which receives the outputs of the voice frequency characteristic extractor and the noise frequency characteristic extractor and removes noise characteristics at a frequency side; A frequency axis voice detector which receives the output of the frequency noise canceller and detects voice on a frequency axis; A voice detection result selection unit which receives the output of the time axis voice detection unit and the frequency axis voice detection unit and selects a voice detection result based on a noise characteristic to output voice; Robust speech recognition device characterized in that configured to include a speaker distance characteristic. The method according to claim 1, The input voice selection unit, The first loudness clarity is calculated by selecting the loudness clarity confidence value for the short-range voice output by receiving the recorded voice from which the external noise is removed from the external noise removing unit, and calculating the vocal clarity based on the loudness of the recorded volume and the clarity of the pitch information. A confidence level selection unit; A first sound quality distortion value selecting unit which receives the output of the first sound accuracy accuracy confidence value selecting unit and selects the sound quality distortion values by the degree of causing discontinuous frequency characteristics in the adjacent voice section when analyzing the frequency characteristics; A first SNR value selector which receives an output of the first sound quality distortion value selector and selects an SNR value that is a relative magnitude of noise; The second loudness clarity that is obtained by inputting the recorded voice from which the external noise has been removed is selected by the external noise removing unit and selecting the confidence value of the loudness accuracy of the far-range voice output, and calculating the vocal clarity based on the loudness of the recorded volume and the clarity of the pitch information. A confidence level selection unit; A second sound quality distortion value selecting unit which receives the output of the second sound accuracy accuracy confidence value selecting unit and selects the sound quality distortion values by the degree of causing discontinuous frequency characteristics in the adjacent voice section when analyzing the frequency characteristics; A second SNR value selector which receives an output of the second sound quality distortion value selector and selects an SNR value that is a relative magnitude of noise; A voice selecting unit for receiving the outputs of the first and second SNR numerical selecting units and selecting a voice input having a high recognition rate to output the selected voice; Robust speech recognition device characterized in that configured to include a speaker distance characteristic. The method according to any one of claims 1 to 4, The voice recognition unit, A feature extraction unit which receives a voice selected by the input voice selecting unit and extracts a unique characteristic for each voice value; A decoder configured to receive the output of the feature extractor and to decode by a speech grammar and an acoustic model; A post-processing unit which receives the output of the decoder and performs post-processing in consideration of linguistic characteristics and clarity of speech timing and outputs a recognition result; Robust speech recognition device characterized in that configured to include a speaker distance characteristic. A first step of allowing voice recording units by distance to simultaneously receive and record voice input from near and far distances; A second step of estimating external noise from distance-specific voices and removing them from the recorded voice after the first step; After the second step, the input voice selecting unit receives the recorded voice from which the external noise has been removed and checks and selects which voice can improve the voice recognition performance among the input voices reflecting the distance characteristics of the remote and short distances. Wow; A fourth step of receiving a voice selected by the voice recognition unit after the third step and performing voice recognition; Robust speech recognition method characterized in that configured to include a speaker distance characteristic. The method according to claim 6, The first step is, An eleventh step of classifying the volume of the recorded voice by receiving the voice of the sampler; A twelfth step of calculating the input gain of the sampler after the eleventh step and outputting an input gain value according to a distance-specific mode to perform pre-parameter setting; Receiving a voice of the real user, calculating an input gain of the real user, and setting parameters of the real user; A fourteenth step of inputting an input gain value of the distance-specific mode in the twelfth step, receiving an input gain value of the real user in the thirteenth step, and specifying an input gain to output voice for each distance; Robust speech recognition method characterized in that configured to include a speaker distance characteristic. The method according to claim 6, The second step, A twenty-first step of extracting a characteristic of a time axis by receiving a voice of a short distance or a long distance output from the voice recording unit by distance; A twenty-second step of detecting voice on a time axis after the twenty-first step; A twenty-third step of receiving a voice for each short distance or long distance output from the voice recording unit for each distance and converting the voice to a frequency; A twenty-fourth step of extracting characteristics of voice frequency and noise frequency after the twenty-third step; A twenty-fifth step of removing noise characteristics on the frequency side after the twenty-fourth step; A twenty-sixth step of detecting voice on the frequency axis after the twenty-fifth step; Performing a twenty-seventh step and a twenty-sixth step, selecting a voice detection result based on a noise characteristic, and outputting a voice; Robust speech recognition method characterized in that configured to include a speaker distance characteristic. The method according to claim 6, The third step, In step 31, the external noise removing unit receives the recorded voice from which the external noise has been removed and selects the loudness accuracy confidence value for the near-range voice output, and calculates the sound quality with the clarity of the recorded volume and the clarity of the pitch information. ; A thirty-second step of selecting a sound quality distortion value by a degree of causing discontinuous frequency characteristics in an adjacent voice section when analyzing the frequency characteristic after the thirty-first step; A thirty-third step of selecting an SNR value that is a relative magnitude value of noise after said thirty-second step; The 34th step of calculating the sound clarity confidence value for the far-range audio output by receiving the recorded voice from the external noise removing unit is removed, the sound clarity is calculated by the loudness of the recording volume and the clarity of the pitch information; ; A thirty-fifth step of selecting sound quality distortion values by the degree of causing discontinuous frequency characteristics in adjacent voice sections when analyzing the frequency characteristic after the thirty-fourth step; A 36th step of selecting an SNR value which is a relative magnitude value of noise after the 35th step; A thirty-seventh step of selecting a voice input having a high recognition rate after outputting the thirty-third step and the thirty-sixth step to output the selected voice; Robust speech recognition method characterized in that configured to include a speaker distance characteristic. The method according to any one of claims 6 to 9, The fourth step, A forty-first step of receiving a voice selected by the input voice selector and extracting a unique characteristic for each voice value; A forty-second step of performing decoding by the speech grammar and the acoustic model after the forty-first step; A 43rd step of performing post-processing in consideration of linguistic characteristics and clarity of speech timing after the 42nd step and outputting a recognition result; Robust speech recognition method characterized in that performed by the speaker distance characteristics.

Images