JP2000029492A - Speech interpretation apparatus, speech interpretation method, and speech recognition apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an efficient speech recognition error correction method, in a speech interpretation apparatus. SOLUTION: This apparatus is provided with a speech recognition section 1001 which recognizes input speeches and converts the speeches into character strings, a speech storage section 1002 which stores the input speeches, a character string storage section 1003 which stores the speech strings converted in the speech recognition section 1001, a language analysis section 1004 which analyzes the character strings stored in the character string storage section 1003, an interpretation section 1005 for making interpretation to other language in accordance with the result of the analysis of the language analysis section 1004, a phrase extraction section 1007 for extracting the phrases of the input speeches corresponding to the portions failed in the analysis in case of the failure of the analysis by the language analysis section 1004 from the speech storage section 1002 and a speech reproduction section 1008 for reproducing the speeches extracted by the phrase extraction section 1007. The speech interpretation apparatus specifies the phrases of the input speeches corresponding to the portions failed in the analysis and, therefore, the user is not required to again utter all of the input speeches.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus for recognizing a spoken voice, and a speech recognition apparatus for recognizing a spoken voice.
The present invention relates to a speech translation device that translates the recognition result into another language.

[0002]

2. Description of the Related Art In recent speech recognition, continuous speech recognition has become possible by introducing a statistical language model in addition to improving the accuracy of an acoustic model. For speech recognition technology,
Nakagawa, "Speech Recognition by Probabilistic Model", edited by IEICE, 1988.

[0003] A speech translator is a device that recognizes an uttered sentence using a continuous speech recognition technique, and translates the recognition result into a different language by machine translation. Documents relating to the speech translation device include, for example, Morimoto, et al., System Configuration and Performance Evaluation of Speech Translation System (ASURA), Transactions of Information Processing Society of Japan, Vol. 9, p
p. 1726-1735, 1996 and the like.

[0004] One of the problems with the speech recognition device and the speech translation device is the problem of recognition errors in speech recognition. In particular, in a speech translator, if an error occurs in speech recognition,
In machine translation, parsing is not possible and correct translation results cannot be obtained. As a countermeasure against a speech recognition error related to the speech translation device, there is a partial translation method that translates only a partial phrase that has been successfully parsed when the syntax analysis of the entire sentence becomes impossible due to the recognition error. This method is described in, for example, Wakita et al., Introduction to Post-Processing Speech Recognition Correct Part Identification Method and Speech Translation Method Using Semantic Similarity, Information Processing Society of Japan Research Report 97-SLP-17, pp. 19-26, 1997
Is disclosed. As another method, there is a method in which a speech recognition result is fed back to a speaker before being subjected to machine translation, and when there is a recognition error, a request for re-speaking is provided.

Further, Japanese Patent Application Laid-Open No. 7-129594 discloses that
There is disclosed an automatic interpreting system that outputs supplementary information of a recognition result together with a speech recognition result when it is determined that an appropriate result is not obtained as a recognition result by a voice recognition unit. For example, as a result of recognizing the utterance "I am gakusei", "wa" in "me"
If either “wa” or “ka” cannot be recognized and ambiguity occurs, information indicating that there is ambiguity is the additional information. Then, in the above automatic interpreting system, in order to inquire the speaker whether "wa" or "ka" is correct, the sentences "I am gakusei" and "Kagahashigakusei" are spoken. Inquiry is made in the form of voice and, if necessary, the inquiry is displayed on the display unit with the character code unchanged.

[0006]

However, in the partial translation method, there is no guarantee that the result of the partial translation is consistent with the intention of the entire utterance sentence, and in some cases,
The content may not be meaningful at all, or may mislead others.

Further, in the method of feeding back the speech recognition result to the speaker before performing the machine translation and requesting re-speech when there is a recognition error, if the speech recognition error is found by the feedback of the speech recognition result, In order to correct only the part where the recognition error has occurred, the whole sentence is usually re-uttered in order to require the translator to specify the corrected part. Since a longer sentence is more likely to cause a recognition error, it is preferable that only a portion where a recognition error occurs can be efficiently corrected.

In the automatic interpretation system disclosed in Japanese Patent Application Laid-Open No. 7-129594, if the recognition result is completely different from the sentence input by the user, the entire sentence must be input.

Accordingly, an object of the present invention is to provide a speech translation apparatus and a speech translation method for efficiently correcting only a part where a speech recognition error has occurred when a speech recognition error has occurred, and to provide a speech translation method when a speech recognition error has occurred. Another object of the present invention is to provide a speech recognition device that efficiently corrects only a portion where a recognition error has occurred.

[0010]

In order to solve the above-mentioned problems, a speech translator according to the present invention recognizes an input speech and converts it into a character string, a speech storage section for storing the input speech, A character string storage unit that stores the character string converted by the voice recognition unit; a language analysis unit that analyzes the character string stored in the character string storage unit; and another language based on the analysis result of the language analysis unit. A translation unit for performing translation to the language analysis unit; a phrase extraction unit configured to extract, from the speech storage unit, a phrase of an input voice corresponding to a part that failed in analysis when the language analysis unit fails in analysis; and the phrase extraction unit. And a sound reproducing unit for reproducing the sound extracted in step (1).

The voice recognition unit recognizes the phrase of the input voice corresponding to the part of the repetition that failed to be analyzed and converts the phrase into a character string. The translation unit includes a character string corresponding to the input voice. Characters obtained by replacing the character string corresponding to the phrase of the input voice corresponding to the part where the analysis has failed with the character string corresponding to the phrase of the input voice corresponding to the part where the repetition of the analysis failed, which has been converted by the voice recognition unit. Translate a column.

The linguistic analysis unit performs an analysis based on a semantic relationship between phrases.

In addition, a prosody information estimating section for extracting prosody information from the input speech is provided, and the phrase extracting section reads, from the speech storage section, a phrase of the input speech corresponding to a portion that failed to be analyzed based on the prosody information. Extract.

Further, the voice reproducing section has a voice quality converting function for converting voice quality, converts the voice quality by the voice quality converting function, and reproduces the voice extracted by the phrase extracting section.

In the character string corresponding to the input voice,
A display unit is provided which highlights and displays a portion corresponding to the phrase of the input voice corresponding to the portion where the analysis has failed.

In order to solve the above-mentioned problems, in the voice translation method of the present invention, a step of recognizing an input voice and converting it into a character string; a step of storing the input voice in a voice storage unit;
Storing the converted character string in a character string storage unit, analyzing the character string stored in the character string storage unit, and, when the analysis fails, corresponding to a part that failed in the analysis. Extracting a phrase of an input voice from the voice storage unit; and reproducing the extracted voice.

In addition, a step of recognizing a phrase of an input voice corresponding to a portion which has been repeated and whose analysis has failed, and converting the phrase into a character string; Translating a character string corresponding to the phrase of the input voice corresponding to the portion of the input voice corresponding to the portion of the voice that has failed to be analyzed that has been converted by the voice recognition unit; May be provided.

The method may further include a step of analyzing a character string stored in the character string storage unit based on a semantic relationship between phrases.

The method may further include the step of extracting prosody information from the input speech, and the step of extracting from the speech storage unit a phrase of the input speech corresponding to the portion that failed to be analyzed based on the prosody information. Good.

The method may further comprise the step of converting and reproducing the voice quality of the extracted voice. Also,
The display unit may include a step of emphasizing and displaying, in the character string corresponding to the input voice, a portion corresponding to the phrase of the input voice corresponding to the portion where the analysis failed.

In order to solve the above-mentioned problems, a voice recognition device of the present invention recognizes an input voice and converts the voice into a character string; a voice storage unit that stores the input voice; A character string storage unit for storing the obtained character string,
A language analysis unit that analyzes the character string stored in the character string storage unit, and, when the language analysis unit fails in the analysis, extracts a phrase of the input voice corresponding to the part that failed in the analysis from the voice storage unit. And a sound reproducing unit that reproduces the sound extracted by the phrase extracting unit.

Further, the voice recognition unit recognizes a phrase of the input voice corresponding to the portion where the repeated analysis failed, converts the phrase into a character string, and corresponds to the input voice stored in the character string storage unit. In the character string, the character string corresponding to the phrase of the input voice corresponding to the part that failed to be analyzed corresponds to the phrase of the input voice corresponding to the part that failed to be analyzed and read back by the voice recognition unit. Replace with a string.

In the character string corresponding to the input voice,
A display unit is provided which highlights and displays a portion corresponding to the phrase of the input voice corresponding to the portion where the analysis has failed.

[0024]

DESCRIPTION OF THE PREFERRED EMBODIMENTS (First Embodiment) An embodiment of the present invention will be described below.

FIG. 1 is a block diagram for explaining the present embodiment. First, an outline of a configuration and a processing flow of the speech translation apparatus will be described with reference to FIG. The input voice is stored in the voice storage unit 1002 and the voice recognition unit 100
In step 1, voice recognition is performed, and the recognition result is output as a character string. The configuration of the voice recognition unit 1001 is shown in FIG. 2 and will be described later in detail. The recognition result processed by the voice recognition unit 1001 is stored in the character string storage unit 1003, and at the same time, sent to the language analysis unit 1004 to perform syntax analysis. The details of the language analysis unit 1004 will be described later with reference to FIG. The result of the syntax analysis by the language analysis unit 1004 is sent to the translation unit 1005, where the translation is performed. For machine translation, case-based translation methods are mainstream. For example, TDMT (Transfer Dri)
ven Machine Translation). For TDMT, see Furuse,
et al, Constituent Bounda
ry Parsingfor Example-Bas
ed Machine Translation, I
n proc. of COLING 94, pp. 10
5-111, 1994 and the like. Translator 10
The result translated in 05 is presented to the user by the translation presenting unit 1006. Specific examples of the translation presenting unit 1006 include a display for presenting a translation result as a character string and a speech synthesizer for outputting as a synthesized speech.

Note that the voice storage unit 1002 and the character string storage unit 1003 may be constituted by the same storage medium, or may be constituted by separate storage media.

FIG. 2 is a block diagram for explaining the speech recognition unit. The voice detection unit 2001 detects a voice section from the input signal. For voice detection, for example, a general method is to calculate short-time power of an input signal, and compare and detect a time series of the power value with a threshold. There is also a method in which a zero cross of a signal waveform is used together. For details of the voice section detection method, see Furui, Digital Voice Processing, Tokai University Press, 1985. Voice detection unit 2001
The audio signal detected by the above is divided at short time intervals (usually tens to several tens of ms) and sent to the feature analysis unit 2002.
The feature analysis unit 2002 extracts a feature vector from the input short-time signal. Generally, an LPC cepstrum is often used as a feature vector used for speech recognition. For the calculation method of the LPC cepstrum, refer to the aforementioned digital signal processing by Furui. Collation unit 2
In step 003, based on the acoustic model 2004, the word dictionary 2005, and the language model 2006, a conditional probability when the time-series data of the input speech is observed for a sentence candidate that can be generated is calculated. The acoustic model 2004 represents a feature parameter sequence of speech divided into phoneme units as a hidden Markov model (HMM). Word dictionary 20
Reference numeral 05 denotes a dictionary for converting a recognition target vocabulary into a sequence of phonemic symbols. The language model 2006 is obtained by modeling the co-occurrence probabilities between words as 2-gram and 3-gram. Regarding the matching method of the matching unit 2003, A *
Numerous methods such as search, beam search, and Viterbi search have been proposed. Regarding A * search, Kawahara et al., Word Spotting in Conversational Speech Using Heuristic Language Model, IEICE, Vol. J78-D-II, No.
7, pp. 1013-1020, Jul. 1995,
For beam search, see Ney et. al, A
Data-driven Organization
of Dynamic Programming Be
am Searchfor Continuous S
peech Recognition, Proc. IC
ASSP87, pp. 833-834, 1987, etc., are helpful. The sentence candidate having the highest probability among the sentence candidate probability values calculated by the matching unit 2003 is estimated as the recognition result for the input speech.

FIG. 3 is a block diagram for explaining the language analysis unit 1004. The linguistic analysis unit 1004 includes a parser 3001, a grammar within a clause 3002, a dependency grammar 3003
It consists of. In the field of language processing, the smallest unit of language is called a morpheme. Therefore, the word sequence obtained as a result of the speech recognition will be referred to as a morpheme sequence here.

A Japanese sentence can be regarded as a concatenation of a unit called a clause composed of one or more content words, zero or more function words, and punctuation marks. The intra-phrase grammar 3002 models the relationship between morpheme concatenation and a phrase. As an example of the intra-phrase grammar, there is a morpheme n-gram model in which a chain of morphemes is formulated as a Markov model.
The expression of the grammar as a morpheme chain probability model is called a probabilistic language model. Meanwhile, dependency grammar 3
003 indicates the relationship between a plurality of clauses called dependency
For example, a context-free grammar (CFG: Context fr)
ee grammar). The parser 3001 converts the morpheme sequence obtained as a result of the speech recognition into the intra-phrase grammar 3002 and the dependency grammar 3002.
Using the two grammars (3), the dependency structure is analyzed for each clause. FIG. 4 shows an example of the analysis result. In this example, for the morpheme sequence, “To Hotel New Otani”, “To the Hotel New Otani”, “I want to go to the Hotel New Otani” Is generated, and the dependency grammar 30
03 defines the dependency relationship between these two phrases. As described above, the linguistic analysis unit 1004 performs linguistic analysis on the input sentence, generates a structured tree having an analysis dependency structure, and outputs the parse tree to the translating unit 1005.

The functions of the translator 1005 and below have already been described with reference to FIG.

The above is the outline of the speech translation processing in the speech translation device.

Next, a description will be given of a problem when a speech recognition error occurs and a method of correcting the speech recognition error in this embodiment.

A problem with speech translators is that speech recognition does not always produce correct recognition results. No matter how good a speech recognition device is made, recognition errors cannot be avoided. When a speech recognition error occurs in the speech translation device, an erroneous translation result is generated, or language analysis fails, and the translation itself cannot be executed. For example, it is assumed that a speech recognition error occurs for an input utterance of “I want to go to Hotel New Otani”, and a recognition result of “I want to go to Hotel Smelt Tama” is generated. At this time, the language analysis unit 10
At 04, the analysis fails. FIG. 5 shows an analysis result when the language analysis fails. In this example, a phrase cannot be generated at the hotel, odor, or Tama where the speech recognition is incorrect. Therefore, a phrase NP (destination) that can take a dependency relationship with the phrase "I want to go" is generated. can not find. If the linguistic analysis unit 1004 fails to generate the analytic tree, the analysis result cannot be passed to the translation unit 1005, so that the translation process cannot be executed.

In such a case, the speech translating apparatus according to the present embodiment estimates a portion that has caused the language analysis to be unexecutable, and requests the user to speak again. Hereinafter, the part that caused the language analysis to fail will be referred to as an ineligible part. For estimating an unqualified part, for example, a part in which a phrase cannot be generated by the grammar in a phrase, or a phrase whose dependency relation is not obtained and which is out of the element of the parse tree is extracted. Also, in the speech recognition correct part specification method described in the introduction of the post-processing speech recognition correct part specification method using semantic similarity and Wakata et al. A part that has not been specified may be specified as an unqualified part. As described in this document, by using semantic attributes, if the utterance “I want to go to a hotel” is misrecognized as “I want to go to a firefly”, Even if the sentence is correct, it is possible to estimate that "to firefly" is an ineligible part.

The ineligible part specified by the language analysis unit 1004 is sent to the phrase extraction unit 1007. The phrase extracting unit 1007 extracts a phrase corresponding to the input ineligible part from the uttered speech stored in the speech storage unit 1002. For the extraction of the phrase, the collation information obtained from the voice recognition unit 1001 is used. In speech recognition, each phoneme is modeled as a state transition model called an HMM, and the HMM is modeled according to the phoneme sequence.
Of the state transition models obtained by combining the phoneme sequences with the highest probability value as the recognition result. In a Viterbi algorithm often used for calculating a score for speech recognition, if a feature vector sequence of an input speech is (Equation 1) and a state transition sequence of an HMM is (Equation 2), the probability value at this time is (Equation 3)
Can be obtained by

[0036]

(Equation 1)

[0037]

(Equation 2)

[0038]

(Equation 3)

According to (Equation 3), at an arbitrary time T = t, the feature vector yi of the input voice is associated with only one state transition. This transition sequence is called a Viterbi path. FIG. 6 is a diagram for explaining the correspondence between the uttered voice and the voice recognition result. The vertical axis is a state transition sequence obtained by combining HMMs corresponding to the phoneme sequence of the speech recognition result. By referring to the Viterbi path 6001 as in this example, the phrase 6003 of the uttered voice corresponding to the ineligible part 6002 can be easily extracted. The partial data of the uttered voice cut out by the phrase extracting unit 1007 is sent to the voice reproducing unit 1008 and reproduced as voice. An audio reproduction unit 1008 that converts the audio data stored as quantized data into an analog signal; a D / A converter; an amplifier that reproduces the converted analog signal as sound;
It is composed of speakers. At this time, there are many users who feel uncomfortable with listening to their own voice which is not usually heard. In order to avoid such discomfort, a voice quality conversion function may be added to the audio reproduction unit 1008. As a voice quality conversion method, it is only necessary to apply a conversion vector which is obtained by associating a feature vector with each of the voices of the speakers A and B according to the same strategy as the speaker adaptation in the voice recognition. For details on voice conversion, see Hashimoto et al., Spectral Mapping for Voice Conversion Using Speaker Selection and Motion Vector Smoothing, IEICE Technical Report, sp95
-1, PP. 1-8, 1995, etc.

The user is required to repeat the voice "from Hotel New Otani" reproduced by the voice reproducing unit 1008. Regarding this repetition, the repetition may be described in a manual or the like, or guidance such as “Please repeat” may be provided.

A display unit for displaying a character string (recognition result) output from the voice recognition unit 1001 is provided.
When reproducing the sentence "To Hotel New Otani" by 08, each inappropriate portion may be highlighted and displayed on the display unit, such as "I want to go to XXX". Highlighting ineligible parts and clarifying where they need to be corrected can help users read back. In addition, by displaying the recognition result, the user can determine whether the phrase that the speech translator has determined to be inappropriate is appropriate or not, and recognize a recognition error in addition to the phrase that the speech translator has determined to be inappropriate. It can be determined whether or not there is. When the user determines that the phrase that the speech translator has determined to be inappropriate is inappropriate, or when the user determines that there is a recognition error other than the phrase that the speech translator has determined to be inappropriate. Alternatively, the user can stop the correction by repetition and select to redo the entire recognition.

When the translation result presentation unit 1006 is constituted by a display, the display may be used as a display unit for displaying the recognition result.

The voice read by the user "Hotel
"Up to New Otani" is input again to the speech recognition unit 1001 and the recognition result "To Hotel New Otani"
Is output to the character string storage unit 1003. At this time, the character string storage unit stores the previous speech recognition result “I want to go to Hotel Odor Tama”. Language analysis unit 1
According to the history information of No. 004, it is known that the portion “Hotel Odor Tama” is an unqualified portion. Therefore, the newly input “until Hotel New Otani” is replaced with the ineligible part “Hotel Smelly Tama”. The morpheme string Hotel New Otani, which is generated in this way, is sent to the language analyzer 1004. The linguistic analysis unit 1004 wants to go to a new input hotel, New Otani, but can perform correct analysis as described with reference to FIG. The result of the correct analysis is sent to the translation unit 1005, and the translation unit 1005 completes the speech translation.

FIG. 7 shows the sequence described above. The left side of the figure is the dialogue sequence between the user and the translation device,
The right side shows the internal operation of the speech translator.

As described above, when the language analysis fails due to a voice recognition error in the voice uttered by the user, the utterance phrase corresponding to the ineligible part is specified, and only that part is identified. Since repetition is performed, a portion where a speech recognition error has occurred can be efficiently corrected, and a correct translation result can be generated.

(Second Embodiment) In the first embodiment, the part causing the analysis error in the language analysis unit is directly associated as an unqualified phrase. However, there is no guarantee that a sentence containing a speech recognition error is necessarily generated in a unit having a semantic unit corresponding to the original utterance intention. For example, "To Hotel New Otani,
If the voice recognition result for the utterance of "I want to go" is incorrectly saying "I want to go to Daejeon", phrases such as "To hotel" and "To Daejeon" are generated. Dependency analysis of "To Hotel" failed. As a result, guidance is given to repeat the phrase "Hotel New", and the "To Hotel New Otani" part cannot be corrected correctly. Next, a second embodiment to which a function for preventing such a malfunction is added will be described.

FIG. 8 is a block diagram for explaining a second embodiment of the present invention. The difference from the first embodiment described with reference to FIG. 1 is that a prosody information estimating unit 8009 is added. The prosody information estimating unit 8009 extracts the prosody information by analyzing the input speech signal. FIG. 9 shows an embodiment of the prosody information estimating section 8009. In this example, a power calculation unit 9001 for calculating short-time power of an input signal, a pitch extraction unit 9002 for extracting a pitch of voice, and a prosody information integration for integrating power information and pitch information to determine a prosody structure of a sentence. It comprises a unit 9003. Pitch extraction refers to estimating the fundamental frequency of vocal fold vibrations. For example, there are a number of extraction methods such as an extraction method based on the number of zero crossings of a voice waveform and a method of extracting from the autocorrelation coefficient of a residual signal in LPC analysis. Is being developed. For details, see the above-mentioned book by Furui, Digital Signal Processing.

FIG. 10 shows an output example extracted by the power calculation unit 9001 and the pitch extraction unit 9002. The upper part is power information, and the middle part is pitch information. In the power information, a part where the value of the power has dropped is estimated to be a pause part. The pitch information is determined by synthesizing a speech tone component (basic intonation component) having a characteristic that the pitch frequency decreases with the passage of time from the beginning of the voice, and a pitch pattern (accent component) unique to a word or a phrase. Prosody information integration unit 9003
Then, the sentence structure based on the prosody information is determined by using such a pause by a pause and a break of a pitch pattern as a unit of a phrase. The lower part of FIG. 10 is an example of a structural tree determined based on each piece of prosody information.

The phrase extraction unit 8007 replaces the error part (the part that caused the language analysis failure) extracted by the language analysis unit 8004 with the error part based on the prosody information obtained from the prosody information estimation unit 8009. Is defined as a playback phrase. For example, for the erroneous part corresponding to “Hotel New”, the utterance voice of “Up to Hotel New Otani”, which is a subtree of the prosodic structure including that part, is used as the repetition target phrase extracted by the phrase extraction unit 8007. Become. As a result,
"To Hotel New Otani" is repeated, and "To Hotel New Otani" is obtained as a new recognition result. This new recognition result is replaced with the corresponding part "to the hotel to Daejeon" in the first speech recognition result "I want to go to the hotel to Daejeon" stored in the character string storage unit 8003. In this manner, the correctly corrected recognition result is subjected to syntax analysis in the language analysis unit 8004, and then translated into another language in the translation unit 8005.

Therefore, in the second embodiment of the present invention, as in the first embodiment, if the speech analysis fails for the voice uttered by the user and the language analysis fails, By specifying the utterance phrase corresponding to the part and repeating the part only, it is possible to efficiently correct the part where the speech recognition error has occurred and generate a correct translation result. Furthermore, even if a phrase is segmented at an inappropriate position in the syntax analysis due to an error in speech recognition, the repetition target portion is determined using the syntax obtained from the prosodic information, thereby It is possible to repeat a phrase.

(Third Embodiment) In the first and second embodiments, the embodiment of the speech translating apparatus has been described. In the present embodiment, the embodiment of the speech recognizing apparatus will be described.

The speech recognition apparatus according to the present embodiment includes a speech recognition unit 1001, a speech storage unit 1002, a character string storage unit 1003, a language analysis unit 1004, and a phrase extraction unit 10 shown in FIG.
07 and an audio reproducing unit 1008. While the input voice is stored in the voice storage unit 1002, voice recognition is performed in the voice recognition unit 1001, and the recognition result is output as a character string. The recognition result processed by the voice recognition unit 1001 is stored in the character string storage unit 1003, and at the same time, sent to the language analysis unit 1004 to perform syntax analysis. The ineligible part specified by the language analysis unit 1004 is sent to the phrase extraction unit 1007. The phrase extraction unit 1007 extracts a phrase corresponding to the input ineligible part from the uttered speech stored in the speech storage unit 1002. The partial data of the uttered voice extracted by the phrase extracting unit 1007 is
To be played back as audio. The user is required to repeat the sound reproduced by the sound reproducing unit 1008. The voice reproduced by the user is input to the voice recognition unit 1001 again, and the recognition result is output to the character string storage unit 1003. Then, the ineligible part is replaced with the recognition result of the repetitive voice. Through such processing, it is possible to efficiently correct the portion where the speech recognition error has occurred.

[0053]

According to the present invention, when a language analysis fails due to a speech recognition error in a speech uttered by a user in a speech translation device, a speech phrase corresponding to an ineligible part. , The user does not need to re-utter all of the input speech. In addition,
By repeating the specified portion, it is possible to efficiently correct the portion where the speech recognition error has occurred, and generate a correct translation result.

Further, with the speech recognition apparatus of the present invention, it is possible to efficiently correct a portion where a speech recognition error has occurred.

[Brief description of the drawings]

FIG. 1 is a block diagram for explaining an embodiment of a speech translation apparatus.

FIG. 2 is a block diagram illustrating a voice recognition unit.

FIG. 3 is a block diagram for explaining a language analysis unit.

FIG. 4 is a diagram for explaining an analysis result by a language analysis unit.

FIG. 5 is a diagram for explaining an example in which the analysis by the language analysis unit has failed.

FIG. 6 is a diagram for explaining the correspondence between an uttered voice and a voice recognition result.

FIG. 7 is a diagram for explaining an operation sequence of an embodiment of the speech translation apparatus.

FIG. 8 is a block diagram for explaining a second embodiment of the speech translation apparatus.

FIG. 9 is a block diagram for explaining a prosody information estimation unit.

FIG. 10 is a diagram for explaining prosody information estimated by a prosody information estimation unit.

[Explanation of symbols]

1001 ... voice recognition unit, 1002 ... voice storage unit, 10
03: character string storage unit, 1004: language analysis unit, 100
5: translation unit, 1006: translation result presentation unit, 1007 ...
Phrase extraction unit, 1008 ... Sound reproduction unit.

Claims (15)

[Claims] A voice recognition unit configured to recognize an input voice and convert the input voice into a character string; a voice storage unit configured to store the input voice; a character string storage unit configured to store a character string converted by the voice recognition unit; A language analysis unit that analyzes a character string stored in a character string storage unit, a translation unit that translates into another language based on the analysis result of the language analysis unit, and a case where the language analysis unit fails in analysis. A speech extraction unit that extracts a phrase of an input speech corresponding to a part that failed to be analyzed from the speech storage unit, and a speech playback unit that plays back the speech extracted by the phrase extraction unit. Translator. 2. The voice recognition unit recognizes a phrase of an input voice corresponding to a portion where repetition of the analysis failed, and converts the phrase into a character string. The translation unit includes a character string corresponding to the input voice. Characters obtained by replacing the character string corresponding to the phrase of the input voice corresponding to the part where the analysis has failed with the character string corresponding to the phrase of the input voice corresponding to the part where the repetition of the analysis failed, which has been converted by the voice recognition unit. 2. The speech translation device according to claim 1, wherein translation is performed on the columns. 3. The speech translation device according to claim 1, wherein said language analysis unit performs analysis based on a semantic relationship between phrases. 4. A prosody information estimating unit for extracting prosody information from an input speech, wherein the phrase extracting unit stores a phrase of the input speech corresponding to a part that failed to be analyzed based on the prosody information in the speech storage unit. The speech translation device according to claim 1, wherein the speech translation device extracts the speech from the speech. 5. The voice reproducing unit according to claim 1, further comprising a voice quality converting function for converting voice quality, wherein the voice quality is converted by the voice quality converting function and the voice extracted by the phrase extracting unit is reproduced. The speech translation device according to claim 1. 6. A display unit for emphasizing and displaying, in a character string corresponding to an input voice, a portion corresponding to a phrase of the input voice corresponding to a portion that has failed in analysis. The speech translation device according to claim 5. 7. A step of recognizing an input voice and converting it into a character string; a step of storing the input voice in a voice storage unit; a step of storing the converted character string in a character string storage unit; Analyzing the character string stored in the storage unit; extracting the phrase of the input voice corresponding to the part of the analysis failure from the voice storage unit when the analysis fails; Reproducing the speech. 8. A step of recognizing a phrase of an input voice corresponding to a portion that has failed to be analyzed and converting it into a character string, and an input corresponding to a portion of the character string corresponding to the input voice that failed to analyze. Translating the character string corresponding to the phrase of the voice replaced with the character string corresponding to the phrase of the input voice corresponding to the part of the input voice that corresponds to the part that failed to be analyzed after being converted by the voice recognition unit; The speech translation method according to claim 7, comprising: 9. The voice according to claim 7, further comprising a step of analyzing a character string stored in the character string storage unit based on a semantic relationship between phrases. Translation method. 10. A method for extracting prosody information from an input voice, and extracting a phrase of the input voice corresponding to a part of which analysis failed based on the prosody information from the voice storage unit. Claims 7 to 9
To any of the speech translation methods. 11. The speech translation method according to claim 7, further comprising the step of converting and reproducing the voice quality of the extracted speech. 12. The display device according to claim 11, further comprising a step of emphasizing and displaying a portion corresponding to the phrase of the input voice corresponding to the portion of the character string corresponding to the input voice which has failed in the analysis. The speech translation method according to any one of claims 7 to 11. 13. A voice recognition unit for recognizing an input voice and converting it into a character string; a voice storage unit for storing the input voice; a character string storage unit for storing a character string converted by the voice recognition unit; A language analysis unit that analyzes a character string stored in a character string storage unit, and, when the language analysis unit fails in analysis, extracts a phrase of an input voice corresponding to a part in which the analysis failed from the voice storage unit. A speech recognition device, comprising: a phrase extraction unit; and a speech reproduction unit that reproduces speech extracted by the phrase extraction unit. 14. The voice recognition unit recognizes a phrase of an input voice corresponding to a part where the repeated analysis failed, converts the phrase into a character string, and corresponds to the input voice stored in the character string storage unit. In the character string, the character string corresponding to the phrase of the input voice corresponding to the part that failed to be analyzed corresponds to the phrase of the input voice corresponding to the part that failed to be analyzed and read back by the voice recognition unit. 14. The speech recognition device according to claim 13, further comprising means for replacing the character string with a character string. 15. A display unit for emphasizing and displaying, in a character string corresponding to an input voice, a portion corresponding to a phrase of the input voice corresponding to a portion that failed to be analyzed. The speech recognition device according to claim 14.