KR101427806B1 - Aircraft Voice Command Execution Method and Cockpit Voice Command Recognizer System therefor - Google Patents

Abstract

In the method of executing the voice command of the fighter pilot of the present invention, voice command (fighter command) of the uttered pilot is automatically segmented by word unit so that silence is obscured by the energy processing of the short-range signal and the zero crossing rate processing, The recognition candidate word set for each word is set by determining the voice characteristic parameter from the Mel-Frequency Cepstrum Coefficent. The recognition candidate word set is a connection word matching technique based on a verbal tree, The number of word combinations is greatly reduced and the recognition rate of commands is greatly increased. In particular, the word recognition time is shortened by reducing the number of similarity measurement times.

Description

TECHNICAL FIELD [0001] The present invention relates to a method for executing a voice command of a fighter pilot,

The present invention relates to recognition of a voice command of a fighter pilot, and more particularly, to a method and apparatus for recognizing voice commands of a fighter pilot by using a rule command of a voice command (fighter command) even when a small number of isolated words are misunderstood among a voice command word (fighter command word) The present invention relates to a method of executing a voice command of a fighter pilot and a voice command recognition system of a fighter pilot in which recognition time is shortened by improving voice recognition rate of a pilot voice command (fighter command) by performing verification and correction of a false word from a verbal tree.

In general, the latest electronic technology applied to fighters can significantly improve the combat capabilities of fighters as well as pilots, but proportional to this, pilots are forced to undergo a high degree of coordination and operational skill.

In addition, the pilot is forced to increase the pressure due to the many menus in the control of the fighter's flight attitude and weapon control. Under these conditions, the pilot must press the desired button directly for controlling the flight attitude or weapon control among numerous buttons, I have no choice but to face.

As a method of improving the control efficiency, there is a voice recognition method in which a pilot performs a command using voice. In such a voice recognition method, a fighter voice command recognition system is used.

For example, a fighter voice command recognition system may include a link word recognizer that can recognize a command composed of words uttered by the pilot, and a pilot that utteres a command (fighter command) composed of words while the pilot presses a certain button on the cockpit To distinguish the start and end of the command (fighter command) and to release the button operation of the pilot at the end of the command (fighter command).

Generally, in the above-described fighter voice command recognition system, the start and end of each word are automatically searched for a voice command (fighter command word) input by a pilot voice, and the isolated word recognition is performed for each word, The voice command of the voice command can be recognized.

Therefore, when a fighter voice command recognition system is applied, control efficiency can be greatly improved along with the inconvenience of a fighter pilot who has to operate many buttons.

Korean Patent Publication No. 10-2002-0060975 (July 19, 2002)

The patent document describes an example of a technique relating to a speech recognition method for capturing an enabling part of a plurality of specific speech or template speech and generating a data profile in which the characteristics of the acquired speech are analyzed by analyzing the enabling part .

However, the voice command recognition of a fighter should be able to satisfy the fast response characteristic of speech recognition in other fields. Therefore, in the same manner as in the patent document, which is an on-line use-based technology such as a smartphone-based technology that is a network-connected server-based speech recognition method, a voice command (fighter command) uttered by a pilot must be instantaneously implemented by flight control or weapon control There is a limit to the ability to satisfy the fast response characteristic.

Especially, it is important for the recognition of speech command of a fighter aircraft to satisfy a high recognition rate compared with speech recognition in other fields.

This is because, even if one of the words of the fighter command word composed of plural words is mistakenly recognized even if the word recognition rate of the fighter voice command is 90% or more, the recognition rate of the command is remarkably lowered due to the one word which is recognized erroneously, I can not help it.

Therefore, it is important that the accurate and high recognition rate of all the words of the fighter command word composed of several words are fulfilled in the voice command recognition of the fighter.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and it is an object of the present invention to provide a speech recognition apparatus and method, After a mistaken word of the isolated word is verified and corrected from a certain rule information of the command (fighter command), a combination having the greatest similarity value of the voice command (fighter command) with the end of command recognition is selected as the final recognition result, There is provided a method of executing a voice command of a fighter pilot and a voice command recognition system thereof.

According to another aspect of the present invention, there is provided a method of executing a voice command for a fighter pilot, the method comprising: receiving a fighter command word uttered by a pilot in a cockpit cockpit, An automatic speech segmentation step of automatically dividing the speech signal into n-th frame units;

One frame unit constituting the short interval signal among the first, n, and n-th frame units is selected and a voice characteristic parameter for voice subjected to the silence judgment of the selected frame unit is extracted, A k-th similar word recognition candidate group consisting of parameters;

A similar word recognition candidate group out of the first, ..., k-th similar word recognition candidate groups is selected, the selected similar word recognition candidate group passes through a syntax tree (Verbal Tree) constructed by the fighter command word, ..., K-like word recognition candidates from the first, ..., k-th similar word recognition candidates, and the matching instruction word candidates are respectively selected as the matching commands after the word matching of all the matching instruction word candidates A connection word matching step;

A voice command execution step in which the similarity value is determined in the adaptation commands, the adaptation command having the maximum similarity value is recognized as a final command, and the final command is executed by the fighter command word uttered by the pilot; Is performed.

In the automatic voice segmentation step, the fighter command includes a series of words and numbers.

Wherein the speech word recognizing step comprises the steps of: extracting a speech signal characteristic and extracting a speech characteristic parameter after the frame unit of the first, n, A speech word recognition mode in which the first, ..., k-th similar word recognition candidates are extracted for speech words composed of the speech characteristic parameters; As shown in FIG.

In the voice word recognition mode, the voice characteristic parameter applies an MFC (Mel-Frequency Cepstrum) to the selected frame unit, processes coefficients of the MFC (Mel-Frequency Cepstrum) It is determined by the gathered MFCC (Mel-Frequency Cepstrum Coefficent).

In the linked word matching step, the selected similar word recognition candidates are classified into ignition lengths and similarity rankings, and inadequate models having an inappropriate search length and similarity rank are excluded.

In the connection word matching step, the verbal tree has an end node as an end point, with the start node as a start point when the fighter command is a phrase, and the end node as an end point between the start node and the end node, The word becomes the parent node and the following word forms the arrival path to the child node; When the fighter command is a numeral, a node representing a number is constructed as a self loop including a plurality of digits; And the adaptation command candidate group is selected when passing through the path formed by the syntax tree (Verbal Tree).

A first recognition result candidate group that has passed through the parent node of the path and a second recognition result candidate group that has passed through the child node are selected, and the first recognition result candidate group and the second recognition result candidate group And the fit instruction word candidate group is selected as a result of the combination.

In the voice command execution step, the degree of similarity determination is performed by applying a VD algorithm (Viterbi Decoding Algorithm) using a log function and an exponential function in a HMM (Hidden Markov Model) After the determination of the degree of similarity, when the good command is selected from two or more recognition results, the good command having a relatively high similarity value is selected as the final recognition result after the similarity values are compared.

According to another aspect of the present invention, there is provided a system for recognizing a voice command for a fighter pilot, the system comprising: a receiving module for receiving a voice command voice of a pilot in a cockpit cockpit; A processing module for comparing the similarity of the pilot voice data model read after selecting the command candidate with the read similarity of the pilot voice data model and then determining the final command as a final command; a transmission module for outputting the final command so that the final command is executed in the fighter; A built-in speech recognizer composed of a voice storage module in which the pilot voice data model is embedded; Is included.

The processing module automatically divides the speech command word by word, performs energy processing of the short-term signal on the word unit basis, and performs zero crossing rate processing on the short-term signal to classify the voice except the silence A speech word processing unit;

A voice characteristic parameter is selected as an MFCC (Mel-Frequency Cepstrum Coefficient) for applying a MFC (Mel-Frequency Cepstrum) process to the voice and collecting a coefficient by a MFC (Mel-Frequency Cepstrum) A speech feature extraction unit for selecting the instruction word from the parameter;

A Viterbi Decoding Algorithm of the HMM (Hidden Markov Model) is applied to determine the similarity between the selected command word candidate and the pilot voice data model, a similarity to the pilot voice data model is determined, and a Verbal Tree A similarity measuring unit for determining the selected instruction word candidate as a final instruction word through a matching process based on the comparison result; .

The present invention has the effect of quickly recognizing an isolated word from a voice command word of a pilot (a fighter command word) composed of a plurality of words connected by a certain rule by applying a connection word matching process using a verbal tree.

In addition, the present invention considers only a suitable combination of the words of a voice command (fighter command word) in a connection word matching process using a verbal tree, thereby reducing possible word combinations as the words of a voice command (fighter command word) are matched , And the time required to measure the similarity of the voice command (fighter command) is greatly improved.

In addition, the present invention applies a recognition candidate for each word of a voice command (fighter command word) in a connection word matching process using a verbal tree, and applies it to a mistaken word verification and correction, thereby improving the recognition rate of the command have.

In addition, the present invention relates to a connection word matching process using a verbal tree, wherein a combination of isolated words found from a pilot voice command word (fighter command word) is subjected to misrecognized word verification and correction, The overall recognition rate is greatly increased, and in particular, there is an effect that the recognition result that the voice command (fighter command) is erroneously recognized as a whole is mistaken for the small number of isolated words among the whole words.

FIG. 1 is a flow chart illustrating a voice command execution procedure of a pilot voice command (fighter command word) according to the present invention, FIG. 2 is a flowchart illustrating an operation of a voice recognition system in a fighter aircraft according to the present invention, FIG. 5 is a detailed block diagram of a voice command recognition algorithm of a fighter voice command recognizer according to the present invention. FIG.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings, which illustrate exemplary embodiments of the present invention. The present invention is not limited to these embodiments.

1 shows a voice command execution procedure of a pilot voice command (fighter command) according to the present invention.

In S1, when a voice command (fighter command) uttered by a fighter pilot is received as in S1, S2 recognizes the received voice command (fighter command) as an individual word and then constructs a suitable command word group through application of a verbal tree In S3, one final command is selected by summing up the similarity values for each word. In S4, the selected final command is executed so that the fighter can be controlled by the voice command of the pilot.

In particular, in S 2, the similarity value of each word can be measured by using VD algorithm (Viterbi Decoding Algorithm) to distinguish and recognize words through silent section extraction. After recognizing the first word after measuring the similarity value, A verbal tree is applied to the first word, and a process of recognizing the second word using the same method is performed.

FIG. 2 is a block diagram illustrating a configuration of a voice recognizer system including a built-in voice recognizer 10 installed in a cockpit of a fighter aircraft 100, Flow.

As shown, the onboard speech recognizer 10 includes a receiving module 20 for receiving voice commands (fighter command words) uttered by a fighter pilot, processing for selecting a command word from the voice signal input to the receiving module 20, A voice storage module 40 provided with a pilot voice for selecting an instruction word as a final instruction word, and a transmission module 50 for outputting a final instruction word.

In the present embodiment, the processing module 30 includes a speech word processing unit 31, a speech feature extraction unit 32, and a similarity measure unit 33.

In the speech word processing unit 31, zero crossing rate processing of the short-range signal is performed together with energy processing of the short-range signal, and the voice feature extraction unit 32 performs MFC (Mel-Frequency Cepstrum) A VD algorithm (Viterbi Decoding Algorithm) is applied to the similarity measuring unit 33 and the voice storage module 40 is provided with a HMM (Hidden Markov) Model) is applied to perform the sound training of the pilot. This will be described in detail later.

Therefore, the above-described procedure of S1 is performed in the receiving module 20, the above-described S2 process is performed in the processing module 30, and the voice storage module 40 together with the processing module 30 The above-described procedure of S3 is performed, and in the transmitting module 50, the above-described procedure of S4 is performed.

As a result, the fighter may be able to fire the air-to-air missile 200 with a voice command (fighter command) or control the flight position of the fighter 100, thereby avoiding the inconvenience of manipulating many buttons of the fighter pilot .

3 and 4 show detailed procedures of a voice command execution algorithm of the pilot voice command (fighter command word) according to the present embodiment.

Referring to FIG. 3, when the voice recognizer is turned on as in S10, the fighter command word uttered by the pilot is input to the voice recognizer as in S20.

In this case, the fighter command uttered by the pilot may be an input word of S21, and the input word may be JETTISON, AIR TO AIR, or STATION 2, for example, S21-1.

In addition, the input word may be a number (NUMBER) as in S22.

These fighter commands have the characteristic that several words are connected by a certain rule.

S30 is a process in which a fighter command input to the speech recognizer is temporarily stored in a buffer and data is extracted therefrom. In this process, the fighter command of the pilot is split into first, . At this time, the frame units are grouped into a plurality of isolated words.

S31 is a process of determining silence for the first, nth, and nth frame units, and the recognition process sequentially proceeds from the first frame unit to the nth frame unit.

For this recognition, the energy treatment of the short-term signal of S31-1 and the zero crossing rate treatment of the short-term signal of S31-2 are applied.

The energy processing of the short-term signal is a method using energy having a large energy in the case of speech while having a small energy in the case of noise and silence, and the zero crossing rate processing of the short- Is used.

In step S32, it is determined whether or not the first frame is muted.

Normally, noise and silence have a large zero crossing ratio and a low energy, but a voice has a relatively large energy. On the other hand, the unvoiced sound has a small energy but the zero crossing rate is small, unlike the noise. Therefore, if the short-term energy is large or the short-term energy is small, but the zero crossing rate is small, the short-term signal can be classified as speech, whereas if the short-term energy is small and zero crossing rate is large, .

If it is determined in step S32 that the first frame unit is mute, if the immediately preceding frame is silent as in step S33, the process returns to step S31 by switching to frame unit reselection in step S34.

On the other hand, if it is determined in step S32 that the first frame unit is not mute, the speech signal (number) characteristic of the first frame unit is extracted as in step S40. The same applies to the case where the previous frame unit = not mute as a result of the judgment of S33.

In particular, by extracting a feature vector for each frame unit, a voice signal can be expressed with a relatively small number of parameters, and in particular, a feature that a calculation amount of a subsequent process such as voice recognition, Respectively.

S41 is a process in which an MFC (Mel-Frequency Cepstrum) is applied to a voice signal (number) feature extraction. The MFC is a method of expressing a power spectrum of a short-term signal. In particular, a nonlinear mel- Mel-Scale) by taking a cosine transform in the log power spectrum in the frequency domain.

S42 is a state in which the voice characteristic parameters are determined by collecting the coef- ficients after the MFC processing into the MFCC (Mel-Frequency Cepstrum Coeffi- nent), from which the characteristic of the voice signal (number) possessed by the frame unit is taken.

In this case, the frequency band of the MFCC is divided evenly in the Mel-Scale, unlike the general spectrum having the frequency bands uniformly divided, so that the energy of the frequency band that is particularly sensitive to the human auditory sense can be expressed more Lt; / RTI >

S50 is a process of determining the similarity between the first to n-th frame units and the similar word and the similarity degree of the speech signal (number) taken in the first to the n-th frame units, And extracted as a candidate group. This is defined as the first similar word recognition candidate group.

S51 is applied to all of the speech signals in which the similar word recognition candidate group extraction process of the speech signal composed of the first to the n-th frame units described above is performed, and then the similar word recognition candidate groups are extracted for the speech signal, , And defines it as a first to k-th similar word recognition candidate group.

As can be seen from this, the ignited pilot fighter command is automatically segmented into frames in isolated words, and the energy of the short-term signal per frame unit and the zero crossing rate of the short- , Then the non-silent frame unit is collected into the Mel-Frequency Cepstration (MFCC) via the MFC (Mel-Frequency Cepstrum), and the voice characteristic parameters are determined, According to the first, ... , And k-like word recognition candidates are matched, respectively.

On the other hand, in S60, , The k-th similar word recognition candidate group is a process of executing the connection word matching using the verbal tree, and comparing the words in the fighter command uttered by the pilot with the verbal tree having a certain scheme, Information is used.

Generally, the verbal tree is a list of combat command words that can be uttered by a pilot selected in advance, and is used as a criterion by which a command word of a fighter commanded by a pilot in command can be compared, The voice command recognizer recognizes the voice data currently uttered by the pilot and the final command word can be determined.

Such a verbal tree can be constructed by distinguishing between words and numbers.

FIG. 5 shows an example of a word syntax tree (Verbal Tree), which has a start node (root node) as an end point, an end node (terminal node) as an end point, and a start node (root node) The syntax tree can be constructed by setting the preceding word to be the parent node and the subsequent word to be the child node according to the command rule.

From this, each word of a voice command forms a tree as a child node of the word immediately preceding it, and the word at the head of the command becomes a child node of the start node (root node) It can be seen that it is the parent node of the end node (terminal node) that represents the end of the command.

Therefore, by comparing the above-mentioned syntax tree (Verbal Tree) with the commands uttered by the fighter pilot executing the command, , It can be easily judged whether or not a series of words selected as the k-like word recognition candidates matches the command syntax.

For example, if a word sequence starts from the starting node (root node) and arrives at the end node (terminal node), then the word rule matches the rule of the command, whereas the word string without the path that passes through the syntax tree, Because it is misaligned, the mistaken word correction in the middle can be done.

Meanwhile, FIG. 6 shows an example of a numerical phrase tree such as a command using a number, such as frequency conversion, in which a fighter's command uttered by a pilot is in the middle. In this case, (Verbal Tree) is built accordingly.

In particular, such a verbal tree has a feature that a node representing a number is composed of a self loop in order to express a phrase including a plurality of digits.

4, after the connection word matching using the verbal tree is executed as shown in S60, it is selected as the final command recognition result as in S80, and from this, the fighter command uttered by the pilot is executed as shown in S90 .

At S62, the first, ... , If the first similar word recognition candidate group among the k-th similar word recognition candidate groups is selected, the selected first similar word recognition candidate group selects the end node (terminal node) from the start node (root node) which is the construction path of the syntax tree .

In this case, the inadequate model considering ignition length is excluded as in S63-1, and the inadequate model considering the similarity rank as S63-2 is excluded. The result that the inadequate model is not excluded in the word matching is shown in FIG.

FIG. 7 shows the results of the case where the inadequate model is not excluded in the situation where the pilot has uttered JETTISON, AIR TO AIR, and STATION 2 as the fighter command word. In this case, up to three candidate words for each word are listed, And the word having the largest value is used as the recognition result. However, in this case, the second word AIR TO AIR has a higher probability of being recognized as ENTER.

S63 is a process in which the selected first similar word recognition candidate group is matched with a parent node of a syntax tree (Verbal Tree), and then a first recognition result candidate group is selected as a result, and S64 is continuously matched with a child node, The recognition result candidate process is selected.

This process is done continuously through the verbal tree.

Next, in S65, the second recognition result candidate group and the first recognition result candidate group are combined with each other. In S66, the group of command candidates most suitable for the combination result is selected. At this time, the selected conforming command candidate group is defined as the first conforming command candidate group.

On the other hand, after the first suitable command word candidate group is selected as in S67, and then the same process is performed until the k-like word word recognition candidate group, all of the suitable command word candidates can be selected as a result in S68. 8, ... , And an iteration process in which the candidate group of the k-th fit command is selected.

S69 indicates that all the selected candidate adaptive commands are matched with words suitable for the verbal tree, and FIG. 9 shows such results.

FIG. 9 shows a result of executing a connection word matching technique by using a verbal tree to which the scheme of FIG. 5 is applied.

As shown in the figure, since the words suitable for the verbal tree in the word 1 candidates are JETTISON and SYSTEM MODE, it can be seen that the first recognition result candidates are selected excluding SEVEN. It is also known that JETTISON, AIR TO AIR, SYSTEM MODE, and NAVIGATION are selected as candidate candidates for the second recognition result, which can be combined with the first word suitable for the verbal tree among the candidates of the word 2.

 Then, the commands which are suitable for the verbal tree considering combination of word 1 and word 2 among the candidates of word 3 are JETTISON, AIR TO AIR, STATION 2 and JETTISON, AIR TO AIR, STATION 1, Are selected, and it can be known that they are candidates for the first fit command word.

However, as a result of such a process, it may happen that the command recognition result value is two or more when the command utterance is terminated.

In this case, in the present embodiment, the similarity values of two or more instruction recognition result values are compared, and as a result, a command with a large similarity value is selected as the final recognition result, thereby eliminating crosstalk from two or more instruction recognition results .

Meanwhile, in step S80, the final command recognition result is selected. In step S90, this final command is executed as a fighter command word. The command execution result is a result that the fighter is controlled by the voice command of the pilot as described in FIG. 2 .

S100 is a process of confirming the pilot fighter command speech and voice input again. If the pilot fighter command speech is confirmed and the voice input is inputted again, the process returns to step S10, whereas if there is no pilot fighter command speech, the pilot mode is switched to the standby mode of S200 .

On the other hand, the characteristics of the speaker dependency using the pilot's voice data model for voice recognition may be considered and used for improving the recognition rate of words. However, in the present embodiment, the connection word matching method is used in the speaker independent method, and it is proved that the recognition performance of the fighter command is greatly improved by correcting the isolated word mistake recognition of the speaker independent method.

As described above, in the method of executing the voice command of the fighter pilot according to the present embodiment, the speech command (fighter command) of the uttered pilot is automatically divided into word units, and energy processing of the short-range signal and zero crossing rate processing And the speech characteristic parameters are determined from the MFCC (Mel-Frequency Cepstrum Coefficent), whereby recognition candidates for each word are set, and the recognition candidates for the respective words are set for the connection word matching based on the verbal tree Based word-based instructions, which greatly reduces the number of word combinations and greatly increases the recognition rate of commands. In particular, word recognition time is shortened by reducing the number of similarity measurements.

10: on-board speech recognizer 20: receiving module
30: Processing module 31: Speech word processor
32: voice feature extraction unit 33:
40: Voice storage module 50: Transmission module
70: Voice trainer
100: Fighter aircraft 200: Air-to-air missiles

Claims (9)

An automatic speech segmentation step in which a fighter command uttered by a pilot in a cockpit of a fighter aircraft is input to a speech recognizer, temporarily stored in a buffer, and automatically divided into first, second, and nth frame units of the fighter command word;
One frame unit constituting the short interval signal among the first, n, and n-th frame units is selected and a voice characteristic parameter for voice subjected to the silence judgment of the selected frame unit is extracted, A speech word recognition step of selecting a first, ..., k-th similar word recognition candidate group after a parameter is obtained from each of the first, second, nth frame units;
A similar word recognition candidate group out of the first, ..., k-th similar word recognition candidate groups is selected, the selected similar word recognition candidate group passes through a syntax tree (Verbal Tree) constructed by the fighter command word, ..., K-like word recognition candidates from the first, ..., k-th similar word recognition candidates, and the matching instruction word candidates are respectively selected as the matching commands after the word matching of all the matching instruction word candidates A connection word matching step;
A voice command execution step in which the similarity value is determined in the adaptation commands, the adaptation command having the maximum similarity value is recognized as a final command, and the final command is executed by the fighter command word uttered by the pilot;
Wherein the step of executing the voice command comprises: 2. The method of claim 1, wherein in the automatic speech segmentation step, the fighter command is a series of words and numbers, and the first, How to do it.
The method of claim 1, wherein the speech word recognition step comprises: extracting a speech signal characteristic after the frame unit of the first, second, nth frame unit is determined as a non-silent speech, A speech word recognition mode in which recognition words composed of the speech characteristic parameters are extracted as the first, ..., k-th similar word recognition candidates;
Wherein the step of executing the command comprises:
2. The method according to claim 1, wherein, in the linked word matching step, the selected similar word recognition candidate group includes the ignition length and the similarity rank, and the inappropriate model having the ignition length and the similarity rank that are not suitable is excluded How to run the command.

2. The method of claim 1, wherein, in the connection word matching step, the verbal tree includes an end node as a start point and a start node as an end point when the fighter command is a phrase, According to the rule, the preceding word becomes the parent node and the following word forms the arrival path to the child node;
When the fighter command is a numeral, a node representing a number is constructed as a self loop including a plurality of digits;
Wherein the fit command word candidate is selected when passing through the path of the verbal tree.
6. The method of claim 5, wherein when the path is passed, a first recognition result candidate group that has passed through the parent node of the path and a second recognition result candidate group that has passed through the child node are selected, 2 recognition result candidate groups are combined with each other, and the fit command word candidate group is selected as a result of the combination.
The method according to claim 1, wherein, in the voice command execution step, when the goodness-of-fit command is selected from two or more recognition results after the determination of the similarity degree, a suitable command having a relatively high similarity value is selected as the final recognition result A method for executing a fighter pilot voice command,
A receiving module for receiving a voice command voice of a pilot from a cockpit of a fighter; and a pilot voice data model for selecting a command word of the voice command from the voice signal of the receiving module, comparing the similarity with the read pilot voice data model, And a speech storage module including the pilot speech data model read by the processing module, wherein the speech recognition module comprises: a speech recognition module configured to recognize the final instruction word; / RTI >
The processing module automatically divides the speech command word by word, performs energy processing of the short-term signal on the word unit basis, and performs zero crossing rate processing on the short-term signal to classify the voice except the silence A speech word processing unit; A voice characteristic parameter is selected as an MFCC (Mel-Frequency Cepstrum Coefficient) for applying a MFC (Mel-Frequency Cepstrum) process to the voice and collecting a coefficient by a MFC (Mel-Frequency Cepstrum) A speech feature extraction unit for selecting the instruction word from the parameter; A Viterbi Decoding Algorithm (HMM) of a HMM (Hidden Markov Model) is applied to determine the similarity between the selected command word candidate and the pilot voice data model, a similarity to the pilot voice data model is determined, And a similarity measuring unit for determining the selected command candidate as a final command through a matching process based on the comparison result.
delete

Images