KR20060064493A - Voice interface system and method - Google Patents

Abstract

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice interface system and method, and more particularly, to an audio robot system and method that can be used in applications such as an intelligent robot, and enables natural voice communication and improves voice recognition performance.

The present invention provides a speech recognition module that performs speech recognition using speech data and determines whether an error of the speech recognition is performed; And an H / O error post-processing module for obtaining a speech recognition result through a human operator when it is determined that the speech recognition module has an error in speech recognition.

Description

Voice interface system and method {VOICE INTERFACE SYSTEM AND METHOD}

1 is a view for explaining a voice interface system according to an embodiment of the present invention.

2 is a diagram illustrating a signal processing flow of a voice interface system according to an embodiment of the present invention.

3 is a diagram illustrating an information processing process when a voice recognition result is positively recognized and when it is incorrectly recognized.

FIG. 4 is a diagram for describing a voice interface method that may be performed in the voice interface system illustrated in FIG. 1.

FIG. 5 is a diagram illustrating an example of an H / O error post-processing step in the voice interface method illustrated in FIG. 4.

FIG. 6 is a diagram illustrating an example of a dialogue model step in the voice interface method illustrated in FIG. 4.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice interface system and method, and more particularly, to an audio robot system and method that can be used in applications such as an intelligent robot, and enables natural voice communication and improves voice recognition performance.

Voice recognition is a very convenient function that can control home appliances or terminals with voice or access desired information by voice. Recently, there is an increasing number of applications for intelligent robots, telematics, home networks, and the like. In particular, in the case of intelligent robots, since interfaces such as a keyboard and a mouse are very difficult, interfaces such as voice recognition, video recognition (gesture, text recognition), and sensors (ultrasound, infrared) are known as effective methods. Is known as the most natural interface for the user.

However, in the prior art, the voice interface has a main function of recognizing and performing less than 100 simple voice commands, and the recognition / synthesis engine is also stand-alone and embedded in the robot, thereby limiting resources such as CPU and memory. Interactive voice interface was difficult. Commands were mainly composed of robot driving commands or commands for menu selection, so there was a limit to application services using robots. In addition, the conventional technology is insufficient to cope with the voice recognition error and user error caused the use of the voice interface system rather uncomfortable.

Therefore, the technical problem to be achieved by the present invention is to solve the above problems, as well as the performance of speech recognition error recognition (Recognition error), the human error (Human error) coping method (Realtime) and the user Considering the usability, the present invention proposes a voice interface method and apparatus capable of human-robot interaction so that an intelligent robot can be used in real life.

As a technical means for achieving the above object, a first aspect of the present invention comprises a speech recognition module for performing speech recognition using speech data, and determines whether the performed speech recognition error; And an H / O error post-processing module for obtaining a speech recognition result through a human operator when it is determined that the speech recognition module has an error in speech recognition.

According to a second aspect of the present invention, a voice interface client converts a voice spoken by a user into voice data and transmits the converted voice data to a voice interface server through communication; And a voice interface server that performs voice recognition using voice data transmitted from the voice interface client, and obtains a voice recognition result through a human operator when it is determined that an error in voice recognition is large. to provide.

A third aspect of the invention provides a speech recognition module for performing speech recognition using speech data; A dialogue model module for forming a system response which is a misrecognition result of a speech recognition result performed in the speech recognition module or a question for correcting an error when there is a semantic error; And a speech synthesis module for converting the question into speech data.

A fourth aspect of the present invention provides a voice interface client for converting a voice spoken by a user into voice data and transferring the converted voice data to a voice interface server through communication; And a voice interface server configured to perform voice recognition using voice data transmitted from the voice interface client, and to form a system response that is a question for correcting an error if a voice recognition result is misrecognized or there is a semantic error. Provide a voice interface system.

A fifth aspect of the present invention is a voice recognition step of performing a speech recognition using the speech data, and determining whether the performed speech recognition error; And (b) an H / O error post-processing step of obtaining a speech recognition result through a human operator when it is determined that there is an error in speech recognition in step (a).

A sixth aspect of the present invention includes the steps of (a) performing a speech recognition using speech data; (b) forming a system response which is a question for correcting an error when there is a mistake in the speech recognition result obtained in step (a) or a semantic error; And (c) converting the system response into voice data.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, embodiments of the present invention may be modified in various forms, and the scope of the present invention should not be construed as being limited by the embodiments described below. Embodiments of the present invention are provided to more fully explain the present invention to those skilled in the art.

1 is a view for explaining a voice interface system according to an embodiment of the present invention.

Referring to FIG. 1, the voice interface system includes a voice interface server 10 and voice interface clients 20a, 20b, and 20c.

The voice interface clients 20a, 20b, and 20c convert the voice spoken by the user into voice data and transmit the converted voice data to the voice interface server 10 through communication. The voice interface clients 20a, 20b, and 20c may be robots, such as intelligent robots, and communicate with the voice interface server 10 through wireless or wired communication such as a wireless LAN. The voice interface clients 20a, 20b, and 20c may have an endpoint detection function for distinguishing a start point and an end point of a voice section. In this case, the voice interface clients 20a, 20b, and 20c distinguish between the silent section and the voice section, and transmit the voice data corresponding to the voice section to the voice interface server 10.

The voice interface server 10 performs voice recognition using voice data transferred from the voice interface clients 20a, 20b, and 20c. The voice interface server 10 includes a speech recognition module 11 and may include a human operator error post-processing module 12, a dialogue model module 13, and a speech synthesis module 14. . In addition, the voice interface server 10 may further include a server management module 15. Each module constituting the voice interface server 10 may be configured as a separate server or hardware. In addition, each module constituting the voice interface server 10 may be implemented in the form of a separate program executed in one server or hardware.

The speech recognition module 11 performs speech recognition using the speech data received from the speech interface clients 20a, 20b, and 20c. When the voice interface server 10 includes the H / O error post-processing module 12, the voice recognition module 11 determines whether there is an error in the result of the performed voice recognition, and if there is an error, The result can be notified to the H / O error post-processing module 12.

The human-operator error post-processing module 12 performs a function of obtaining a speech recognition result through a human operator when it is determined that the speech recognition module 11 has an error in speech recognition. More specifically, when it is determined that there is an error of speech recognition in the speech recognition module 11, a person heals the error by directly listening to the voice and inputting an accurate speech recognition result. The H / O error post-processing module 12 has a function of displaying the cumulative number of speech recognition errors for each user, so that the error of a user having a large number of accumulated rejections can be corrected first, thereby minimizing user complaints. have. In addition, the H / O error post-processing module 12 may display the wrong word frequently, allowing the human operator to easily select the positive recognition result, thereby enabling efficient correction. In addition, the H / O error post-processing module 12 displays at least one word that is determined to be closest to the word in which the error occurs, thereby enabling the human operator to quickly find a recognition result among the displayed words, thereby efficiently Modifications can be made possible. In addition, the H / O error post-processing module 12 may display the conversation history, allowing the human operator to select the recognition result more accurately and efficiently. In addition, the H / O error post-processing module 12 has an automatic word indexing function, so that when only a few phonemes are typed, a list of matching words is selected to easily select a recognized word without typing the remaining phonemes. Efficient modifications can be made possible. In addition, the H / O error post-processing module 12 may have a function of varying the speech rate so as to record the recognized speech recognition result after listening to the voice at a high speed, thereby improving the H / O error post-processing speed. have.

 The conversation model module 13 performs a function of forming a system response to correct an error when there is a semantic error in the speech recognition result obtained by the speech recognition module 11 or the H / O error post-processing module 12. do. As an example of semantic error, assume that "[date] + [weather]" is a case where there is no error in semantic relation. In this case, if a voice recognition result of simply "weather" is obtained, there is a semantic error in voice recognition, and it is necessary to ask the user which date the weather is asked. In addition, if a voice recognition result of "father + weather" is obtained, there is also a semantic error in speech recognition, and it is necessary to ask the user which date the weather is asked. As described above, the dialogue model module 13 forms a system response for correcting an error when there is a semantic error, thereby facilitating interactive interaction between the user and the voice interface.

The voice synthesis module 14 converts the system response output from the dialogue model module 13 into voice data and transmits the voice response to the voice interface clients 20a, 20b, and 20c.

The server management module 15 may be used when the speech recognition module 11, the H / O error post-processing module 12, the dialogue model module 13, and the speech synthesis module 14 are each implemented in the form of an independent server. As one module, load balancing can enable real-time processing.

When a robot is used in each home, the voice interface clients 20a, 20b, and 20c may be multiple units in each home, and each home requests information to the voice interface server 10 through communication such as wireless LAN. The voice interface server 10 provides the result of the information processing according to the voice data transmitted from the voice interface clients 20a, 20b, and 20c. This allows the user to purchase low cost voice interface clients 20a, 20b, and 20c, and to provide services in real time by in charge of various information processing through voice interface server 10. FIG. Information transfer between voice interface server 10 and voice interface clients 20a, 20b, and 20c preferably uses packets.

2 is a diagram illustrating a signal processing flow of a voice interface system according to an exemplary embodiment of the present invention, and FIG. 3 illustrates an information processing process when a voice recognition result is positively recognized and when a false recognition is performed. It is a figure which shows.

Referring to FIGS. 2 and 3, when the voice recognition result is positively recognized, the information processing process includes the steps of the user 30 uttering a voice command such as "what is the schedule today" (S11), and the voice interface client. 20, after detecting the voice section among the voice data uttered by the user 30 and delivering the detected voice data (S12), the voice recognition module 11 transmits the “today” using the transmitted voice data. In step S13, the dialogue model module 13 performs the voice recognized result of "voice" based on the voice recognition result. "Who is it?" Forming a system response (S14), the speech synthesis module 14 converts the system response into voice data (S15), and the voice interface client 20 converts the user into voice data according to the system response. The step (S16) to ignite.

In addition, the information processing process when the voice recognition result is recognized is a step (S21) in which the user 30 utters a voice command such as "what is the schedule today" (S21), the voice interface client 20 is the user ( In step S22, the voice data detected by the voice section of the voice data uttered by the voiced voice is transmitted (S22), and the voice recognition module 11 performs the voice recognition that is judged as a false recognition using the transferred voice data. In step S23, the error is corrected by the human operator in the H / O error post-processing module 12 to form a voice recognition result of "today" and "schedule" (S24), and the dialogue model module 13 Based on this speech recognized result, "Who's schedule?" Forming a system response (S25), the speech synthesis module 14 converts the system response into voice data (S26), and the voice interface client 20 to the user according to the system response converted into voice data. Ignite step (S27).

FIG. 4 is a diagram for describing a voice interface method that may be performed in the voice interface system illustrated in FIG. 1.

Referring to FIG. 4, the voice interface method includes a voice enhancement step S31, a voice endpoint detection step S32, a voice / non-voice verification step S33, a voice feature extraction step S34, a real time noise compensation step S35, Keywords search step (S36), online speaker adaptation step (S37), speech verification step (S38), H / O error post-processing step (S39), dialogue model step (S40) and speech synthesis step (S41). Among the above steps, the voice enhancement step S31 and the voice endpoint detection step S32 may be performed in the voice interface client, and the remaining steps may be performed in the voice interface server. If the voice endpoint detection step S32 is composed of two steps, the first step of the two steps may be performed in the voice interface client, and the second step may be performed in the voice interface server. Voice enhancement step (S31), voice endpoint detection step (S32), voice / non-voice verification step (S33) voice feature extraction step (S34), real-time noise compensation step (S35), keyword search step (S36), online speaker adaptation Step S37 and speech verification step S38 may be referred to as speech recognition step S42 for convenience. Speech / non-speech verification step (S33) Speech feature extraction step (S34), real-time noise compensation step (S35), keyword search step (S36), online speaker adaptation step (S37) and speech verification step (S38) is a speech recognition module The H / O error post-processing step (S39) is performed in the H / O error post-processing module, the dialogue model step (S40) is performed in the dialog model module, and the speech synthesis step (S41) is performed in the speech synthesis module. Can be.

Speech enhancement step (S31) mainly removes stationary background noise to increase the intelligibility of speech relative to background noise. Array signal processing and Wiener filter are performed. Function and so on.

The voice endpoint detection step S32 is a step of distinguishing between the silent section and the voice section. For example, two steps of voice endpoint detection may be performed. In the second stage of speech endpoint detection, the first stage of performing the first speech endpoint detection using the energy information of the speech and the statistical model are more precisely using the global speech absent probability (GSAP) from the result of the first stage. And a second step of detecting the endpoint. The first step of the two voice endpoint detection steps may be performed by the voice interface client, and the second step may be performed by the voice interface server.

In the voice / non-voice verification step (S33), the endpoint-detected voice is subjected to a voice recognition noise recognition process through a GMM (Gaussian Mixture Model) -based voice / non-voice verification method, etc. If it is determined that the voice is not performed, the subsequent steps are performed.

In the voice feature extraction step S34, feature parameters (eg, a filter bank, a meltstring, etc.) are extracted from the voice.

In the noise compensation step S35, non-stationary background noise for the speech section is removed in real time using the IMM (Interactive Multiple Model) method. The final feature parameter without noise is used to calculate the probability value from the HMM (Hidden Markov Model) acoustic model. From this, the result of the recognition is compared by comparing the probability value between the lexical candidate words.

In the key word search step (S36), when the recognition target vocabulary is large (for example, 1000 words or more), the time required for recognition increases, so that a high speed using a tree search or the like can be output in real time. Use the search method. When speech commands are spoken, not only isolated words but also short words can be spoken, and the key words included in the short words are extracted to recognize only the key words, so that the user's speech can be greatly increased compared to isolated words.

In the online speaker adaptation step (S37), through the online speaker adaptation, the speaker's speech characteristics are reflected in real time to the speaker-independent speech characteristics that have been previously modeled, thereby preventing the degradation of the recognition performance due to the speaker-independent acoustic modeling.

In the speech verification step (S38), it is verified whether the speech recognition result is correct or incorrect. In general, when misrecognition occurs, if the misrecognized result is used in the system response as it is, the user's satisfaction is greatly reduced, and the voice recognition function may cause inconvenience rather than convenience. In order to solve the user's dissatisfaction due to the voice recognition error, the voice recognition result is verified once again, and the rejection function that delivers it to the system response only when it is believed to be the recognized result and the other voice is made to the user once again is very convenient for user convenience. It is important. The utterance verification step S38 may be performed by using a score value extracted from various Log Likelihood Ratio (LLR) values (eg, anti-model LLR score, N-best LLR score, combination of LLR score, and word duration). A second step of increasing the reliability of speech verification using the intermediate results and metadata (e.g., SNR, gender, age, syllable number, phonological structure, pitch, speech rate, dialect, etc.) output for each recognition module step. It may include. The voice interface server decides whether to proceed to the next step, dialogue model / human operator, or to request recurrence from the user, based on the final result of speech verification.

The H / O error post-processing step S39 is a step performed when speech recognition is misidentified in the speech verification step S38, and a human operator who is a human corrects the recognition error to a positive recognition result.

The dialogue model step (S40) receives the speech recognition result directly from the speech recognition module, or receives the modified speech recognition result from the H / O error post-processing step (S39). Example: "Daily Father's Day" means no semantic error, or "Weather Father's Day" means semantic error.), And responds to the missing semantic word (key word) again. Step to generate.

Speech synthesis step S41 forms speech data in accordance with the system response. In this case, the system may analyze the aspect from the system response sentence and convert it into a dialogue style synthesized sound according to the speaker's intention.

FIG. 5 is a diagram illustrating an example of an H / O error post-processing step in the voice interface method illustrated in FIG. 4. In the H / O error post-processing step, it is very important to process a plurality of false recognition results inputted from multiple voice interface clients so as to respond to a large number of users quickly. Accordingly, the present invention proposes a method in which a human operator can efficiently correct a misrecognition result as a positive recognition result.

The H / O error post-processing step may include a rejection count display step (S51). Rejection count display step (S51) accumulates the number of rejection (Rejection) in the Utterance verification step in the DB 41, so that the user errors by accumulating the cumulative number of rejections is preferentially corrected, user complaints Steps to minimize this.

The H / O error post-processing step may include a frequently incorrect word display step S52. This step registers frequently wrong words in the DB 42 and displays them, thereby enabling efficient modification by allowing the operator to easily select the correct recognition result.

The H / O error post-processing step may include displaying a best recognition result (S53). In this step, by displaying a plurality of words closest to the recognition result of the misrecognized word, it is possible to quickly find the correct recognition result among the displayed words.

The H / O error post-processing step may include a conversation history display step S54. In this step, the log of the conversation between the user and the voice interface system is displayed, allowing the operator to select the recognition result more accurately.

The H / O error post-processing step may include a word automatic indexing step (S55). In this step, when only a few phonemes are typed so that the correct word can be found quickly, a list of matching words is selected so that the correct word can be selected immediately without typing the remaining phonemes.

The H / O error post-processing step may include a variable speech rate step (S56). Since the time for listening to the voice is proportional to the length of the voice, the time for the user to respond to the positive recognition result through the H / O error post-processing step becomes longer, so that the speech can be heard in the variable speech rate step S56. We want to improve the H / O error post-processing speed by significantly reducing the length of speech to the extent possible.

FIG. 6 is a diagram illustrating an example of a dialogue model step in the voice interface method illustrated in FIG. 4.

The conversation model step includes a semantic error post-processing step S61, a search conversation area limit step S62, and a response dialog sentence generation step S63.

In the semantic error post-processing step (S61), it is verified whether there is a semantically error in the speech-recognized result, and if there is an error, the user is requested to speak again the missing semantic word (key word). At this time, when there is such semantic ambiguity, when a form other than a predetermined rule is input using the semantic relation rule table as shown in Table 1, the dialogue is performed in the most similar form.

The vocabulary spoken by the user will be limited according to the dialogue sentence generated at the dialogue model stage. For example, if the dialogue sentence generated at the dialogue model stage is a sentence asking for timing, the answer will be a word within a certain range such as date and time. Therefore, in the search dialogue area limiting step S62, the target range of the key word to be searched in the above key word search step is reduced, and as a result, the recognition rate is improved.

In response dialogue generation step S63, a system response is generated.

Table 2 shows the operation states of the input / output between the user, the voice interface client (robot), and the voice interface server over time.

In the initial stage, the robot (client) and the server are in a standby state, and the background noise input from the robot is sent to the server to adapt to the environment in real time. When the user calls "robot" from a distance, the robot estimates the speaker position through the array microphone, removes the noise, and detects the voice section and sends it to the server. The server recognizes who speaks by performing speaker recognition, and loads speaker's personal information to adapt the acoustic characteristics according to the speaker's characteristics. The robot turns to the estimated speaker position and moves up to 50cm away from the user. And it receives the voice synthesis sound from the server and outputs to the user, "What can I do for you? Hong Gil Dong". At this time, the robot performs face-tracking to look at the speaker's lips through image recognition, and extracts multimodal information along with voice information.

The user requests information from the robot (eg, “How is the weather today?”) And the robot sends noise to the server after noise reduction and voice endpoint detection as before. The voice sent to the server extracts key words (eg, “today's weather”) contained in the short text and performs voice recognition. At this time, it includes a barge-in processing function so that speech recognition is possible while the synthesized sound is output. Voice recognition results are performed through online speaker adaptation, voice recognition is performed, and the verification results are verified again through utterance verification. Based on the utterance verification results, they are directly inputted into the dialogue model or the H / O error post-processing process. The result is corrected and the final recognition result is input into the dialogue model. The dialogue model generates a system response (eg, “Weather in Daejeon today”) for the query requested by the user and outputs it through the dialogue synthesizer. After that, an interactive voice interface is established between the user and the robot through the same process. When the final termination signal (eg, “OK”) is given, the robot and the server wait in an initial state.

Voice interface device and method according to the present invention has the advantage that the voice recognition error can be minimized by performing the H / O error post-processing.

In addition, the speech interface device and method according to the present invention forms an appropriate system response using a conversational model, so that when a semantic error or speech recognition error occurs, a proper question is presented to the user and the answer is obtained. The advantage is that it can cope with errors or user errors appropriately.

In addition, the voice interface device and method according to the present invention has the advantage of being able to personalize the accuracy and speed of speech recognition by reducing the range of key words searched for during speech recognition according to the system response formed in the dialogue model.

In addition, the apparatus and method for voice interface according to the present invention includes at least one that is determined to be the closest to the cumulative number of user-specific speech recognition errors, frequently incorrect words, and errors occurring in performing H / O error post-processing. By displaying at least one of a word and a conversation history, having a word automatic indexing function, or having a variable speech rate function, the efficiency of H / O error post-processing can be improved.

In addition, the voice interface device and method according to the present invention has the advantage of having a client / server structure, thereby enabling a low-cost purchase of a client, in particular a robot client.

Claims (30)

A voice recognition module which performs voice recognition using voice data and determines whether an error of the voice recognition is performed; And And a H / O error post-processing module for obtaining a speech recognition result through a human operator when it is determined that there is an error in speech recognition in the speech recognition module. The method of claim 1, The H / O error post-processing module is configured to display at least one of a cumulative number of voice recognition errors for each user, a word that is frequently wrong, at least one word that is determined to be closest to an error word, and a conversation history. . The method of claim 1, The H / O error post-processing module has a word automatic indexing function. The method of claim 1, The H / O error post-processing module has a speech interface variable function. The method of claim 1, A dialogue model module for forming a system response which is a question for correcting an error when there is a semantic error in a speech recognition result obtained by the speech recognition module or the H / O error post-processing module; And And a speech synthesis module for converting the system response into speech data. The method of claim 5, wherein And the voice recognition module searches only words in a range corresponding to a system response formed in the conversation model module. A voice interface client for converting a voice spoken by a user into voice data and transferring the converted voice data to a voice interface server through communication; And And a voice interface server that performs voice recognition using voice data transmitted from the voice interface client, and obtains a voice recognition result through a human operator when it is determined that an error in voice recognition is large. The method of claim 7, wherein The voice interface server is a voice interface server according to any one of claims 1 to 6. The method of claim 7, wherein And the voice interface client has an endpoint detection function of voice data converted from voice spoken by the user. The method of claim 7, wherein And the voice interface client is a robot. A speech recognition module for performing speech recognition using speech data; A dialogue model module for forming a system response that is a misrecognition result of a speech recognition result performed in the speech recognition module or a question for correcting an error when there is a semantic error; And And a speech synthesis module for converting the question into speech data. The method of claim 11, And the voice recognition module searches only words in a range corresponding to a question formed in the conversation model module. A voice interface client for converting a voice spoken by a user into voice data and transferring the converted voice data to a voice interface server through communication; And A voice including a voice interface server configured to perform voice recognition using voice data transmitted from the voice interface client, and to form a system response that is a question for correcting an error if a voice recognition result is misrecognized or there is a semantic error. Interface system. The method of claim 13, The voice interface server is a voice interface server according to claim 11 or 12. (a) performing voice recognition using voice data and determining whether an error of the performed voice recognition is performed; And and (b) an H / O error post-processing step of obtaining a speech recognition result through a human operator when it is determined in step (a) that there is an error in speech recognition. The method of claim 15, Step (a) is (a1) extracting feature parameters from the voice data; (a2) searching and obtaining a keyword from the extracted feature parameters; And (a3) determining whether an error of speech recognition is performed by determining whether the key word obtained in step (a2) is a result of a positive recognition or a result of a misrecognition. The method of claim 16, Step (a3) is Determining whether an error of speech recognition is made by using a score value extracted from at least one type of LLR value; And A speech recognition method comprising the step of determining whether the speech recognition error using the metadata. The method of claim 16, Step (a) is (a4) further comprising the step of reflecting the speaker's voice characteristics in real time to the speaker independent speech characteristics. The method of claim 16, Step (a) is (a5) The voice recognition method performed before step (a1), further comprising a voice endpoint detection step of distinguishing between the silent section and the voice section of the voice data. The method of claim 19, Step (a5) is Detecting a voice endpoint using energy information of the voice; And Detecting a speech endpoint using GSAP. The method of claim 19, Step (a) is (a6) The voice recognition method performed before the step (a1), further comprising verifying whether the endpoint detected voice data is voice or noise. The method of claim 19, Step (a) is (a7) The voice recognition method performed before the step (a5), further comprising the step of removing static background noise from the voice data. The method of claim 16, Step (a) is (a8) further comprising the step of removing the non-static background noise from the feature parameter extracted in the step (a1). The method of claim 15, Step (b) is And displaying at least one of a cumulative number of speech recognition errors for each user, a word frequently mistaken, at least one word determined to be closest to the word having the error, and a conversation history. The method of claim 15, Step (b) is When the at least one phoneme is typed, listing a word that includes the typed phoneme. The method of claim 15, Step (b) is A speech recognition method comprising the step of varying the speech rate. The method of claim 15, (c) forming a question for correcting an error if there is a semantic error in the speech recognition result obtained in step (a) or step (b); And (d) converting the question into speech data. The method of claim 27, Step (c) is (c1) determining whether there is a semantic error in the speech recognition result obtained in step (a) or step (b); (c2) forming the question; And (c3) controlling to search only key words of a range corresponding to the question in the speech recognition performed later. (a) a voice recognition step of performing voice recognition using voice data; (b) forming a system response which is a question for correcting an error when there is a mistake in the speech recognition result obtained in step (a) or a semantic error; And (c) converting the system response into voice data. The method of claim 29, Step (b) is (b1) determining whether a speech recognition result obtained in step (a) has a misrecognition or a semantic error; (b2) forming the system response; And (b3) controlling to search only words in a range corresponding to the system response in the speech recognition performed later.

Images