WO2004047075A1 - 音声処理装置および方法、記録媒体並びにプログラム - Google Patents
音声処理装置および方法、記録媒体並びにプログラム Download PDFInfo
- Publication number
- WO2004047075A1 WO2004047075A1 PCT/JP2003/014342 JP0314342W WO2004047075A1 WO 2004047075 A1 WO2004047075 A1 WO 2004047075A1 JP 0314342 W JP0314342 W JP 0314342W WO 2004047075 A1 WO2004047075 A1 WO 2004047075A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- word
- speech
- unknown word
- unknown
- network
- Prior art date
Links
- 238000012545 processing Methods 0.000 title claims abstract description 60
- 238000000034 method Methods 0.000 title claims abstract description 56
- 230000008569 process Effects 0.000 claims description 36
- 238000003672 processing method Methods 0.000 claims description 5
- 230000004044 response Effects 0.000 claims description 5
- 238000012217 deletion Methods 0.000 abstract description 17
- 230000037430 deletion Effects 0.000 abstract description 17
- 230000003247 decreasing effect Effects 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 13
- 230000007704 transition Effects 0.000 description 13
- 238000000605 extraction Methods 0.000 description 10
- 238000003780 insertion Methods 0.000 description 10
- 230000037431 insertion Effects 0.000 description 10
- 238000006243 chemical reaction Methods 0.000 description 6
- 238000006467 substitution reaction Methods 0.000 description 6
- 230000005236 sound signal Effects 0.000 description 5
- 230000014509 gene expression Effects 0.000 description 4
- 239000000284 extract Substances 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 239000004065 semiconductor Substances 0.000 description 3
- 238000001228 spectrum Methods 0.000 description 3
- 102100033806 Alpha-protein kinase 3 Human genes 0.000 description 2
- 101710082399 Alpha-protein kinase 3 Proteins 0.000 description 2
- 244000205754 Colocasia esculenta Species 0.000 description 2
- 235000006481 Colocasia esculenta Nutrition 0.000 description 2
- 230000009471 action Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- NLZUEZXRPGMBCV-UHFFFAOYSA-N Butylhydroxytoluene Chemical compound CC1=CC(C(C)(C)C)=C(O)C(C(C)(C)C)=C1 NLZUEZXRPGMBCV-UHFFFAOYSA-N 0.000 description 1
- 235000016496 Panda oleosa Nutrition 0.000 description 1
- 240000000220 Panda oleosa Species 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 239000012634 fragment Substances 0.000 description 1
- 230000010365 information processing Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
- G10L15/18—Speech classification or search using natural language modelling
- G10L15/183—Speech classification or search using natural language modelling using context dependencies, e.g. language models
- G10L15/187—Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/08—Speech classification or search
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
- G10L2015/022—Demisyllables, biphones or triphones being the recognition units
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/02—Feature extraction for speech recognition; Selection of recognition unit
- G10L2015/027—Syllables being the recognition units
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
- G10L2015/0631—Creating reference templates; Clustering
Definitions
- Audio processing apparatus and method recording medium, and program
- the present invention relates to a voice processing device, a voice processing method, a recording medium, and a program, and in particular, during recognition of a voice signal that is continuously input, an unknown word included in the input voice signal.
- the present invention relates to a voice processing device, a voice processing method, a program, and a recording medium that can be easily registered. Background art
- speech recognition is performed on the speech in units shorter than words (subwords), such as phonology and syllables, and a sequence of syllables is given to the speech (that is, reading a kana).
- the score of each syllable is calculated, and the score of an unknown word (00V: 0ut Of Vocabulary) is estimated by assigning an appropriate penalty to this score.
- word speech recognition considering the possibility that words other than normal word candidates are unknown words, the score obtained as described above is used for words other than normal word candidates.
- the unknown word part is recognized as an unknown word if the unknown word score is worse than the correct word score and better than the incorrect word score. Is done.
- the sub-word sequence for example, syllable series
- the sub-word sequence is referred to based on the time information in the syllable typewriter of the unknown word part. This makes it possible to estimate syllable strings assigned to unknown words (for example, by I ssam Bazz i and James R. Glass).
- Non-Patent Document 2 Hiroaki Ogikubo, Shigehiko Onishi, Hiroshi Yamamoto, and Genro Kikui, subword model Efficient search method for unregistered word recognition, Journal of Information Processing Society of Japan, published in July 2000, Vol. 43, No. 7, p. 2082-2090 (hereinafter referred to as Non-Patent Document 3) See).
- the time of the word boundary of the word string obtained as a result of the word string search and the temporal boundary of the subword sequence obtained as a result of the search of the subword series are not necessarily coincident.
- the recognition result by word speech recognition is wordl ⁇ 00V> word2
- the word string boundary and sub-word sequence (Syl l to Syl8) boundary mismatch may occur.
- ⁇ 00V> is a symbol representing an unknown word.
- the boundary before time oov> corresponds to the middle of Syl4
- the boundary after time corresponds to the middle of Syl7.
- Syl4 and Syl7 which are sub-nodes of the boundary where the mismatch occurs, may or may not be included in 00V>. Therefore, in order to obtain a pronunciation of ⁇ 00V>, the subword boundary must be determined.
- a method of determining the boundary of sub-nodes As a method of determining the boundary of sub-nodes, a method of determining the boundary using a sup- ply sequence is known. The method, that is, the method for obtaining oov> pronunciation by subword sequence is described below with reference to FIG.
- 00V> pronunciation acquisition method based on subword sequence when normal speech recognition and recognition by syllable typewriter are completed, each syllable is continued for syllables including the syllable typewriter time at both ends of 00V>. If 50% or more of the time is included in the section of 00V>, the syllable is made part of 00V>. For example, as shown in FIG. 2, it is assumed that a part of the recognized word string is an array of word 1, 00V>, and word 2. Also, suppose that a part of the subword sequence by the syllable typewriter is an array of syllable i, syllable j, and syllable k.
- L1 length of syllable i corresponding to word 1 side (time)> L 2 (length of time corresponding to 00VM rule of syllable i (time)), so syllable i is Is not included in 00V>.
- L3 the length of the part corresponding to the 00VM rule of syllable k (time)
- L 4 the length of the part corresponding to word 2 side of syllable k (time)
- Syllable k is determined to be included in 00V>.
- Figures 3 and 4 show the results of experiments using the “00V” pronunciation acquisition method according to the subword sequence of FIG.
- the method of obtaining 00V> pronunciation was tested on 7 5 2 types of utterances in a travel domain of 12 people (6 men and women) (such as hotel check-in and restaurant order).
- conditions were set for features, acoustic models, and language models.
- Features include 16 bit, 16 KHz sampling, 1 O msec frame period, 25 msec frame length, and 1 2nd order MFCC (Mel Frequency Cepstrum Coffi cients), and It is set to the first order regression coefficient (25th dimension) of MFCC from 0th order to 1st order.
- the acoustic model is 1 6 mixture, 1 0 0 O ti ed -state H (Hidden Markov Model), and the language model is subword trigram, cut-off trigram 5, bigram 5.
- 3 14 syllables and syllable chains were used as subwords.
- As the language model we used phonological trigrams learned with a 6-year corpus of Nikkei Shimbun.
- Fig. 4 shows the performance of recognition accuracy, substitution error, deletion error, and insertion error of the subword sequence when applying the method of obtaining pronunciation of sub-word sequence of Fig. 2>00V> pronunciation.
- a substitution error is an error that replaces a syllable different from the original syllable
- a deletion error is an error that does not recognize a syllable that should be recognized originally (deletion error).
- a syllable that should not be recognized is an error that appears in the recognition result.
- the recognition accuracy Acc is obtained from Eq. (1) using the total number of syllables N, the number of correct answers NJ, and the number of insertion errors N–I.
- the recognition accuracy with the sub-word sequence in Fig. 2 is 00.2> 4%, and the rate of occurrence of deletion errors and insertion errors is 33.3%, respectively. % And 4.1%.
- the word recognition must be performed while taking into account the syllable boundaries when performing continuous word recognition.
- the recognition accuracy is not as high as 40.2%, and the occurrence rate of deletion errors is as high as 33.3%.
- the deletion error rate was 33.3%, while the insertion error rate was 4.1%, and the balance between the two was poor.
- the present invention has been made in view of such a situation, and an object of the present invention is to improve the recognition rate of speech recognition and to balance the occurrence of deletion errors and insertion errors.
- the speech processing apparatus of the present invention includes a recognition unit that recognizes continuous input speech, an unknown word determination unit that determines whether or not an unknown word is included in the recognition result recognized by the recognition unit, and an unknown word Network generation for generating a network having a path including a subword at a time corresponding to the word boundary of the unknown word and a path not including the subword when the determination means determines that the recognition result includes an unknown word If the means and the unknown word determination means determine that the unknown word is included, the acquisition means for acquiring the word corresponding to the unknown word and the word acquired by the acquisition means are related to other information. And registration means for registering.
- the speech processing apparatus further includes pattern determination means for determining whether or not the recognition result matches a specific pattern, and the registration means determines that the recognition result matches the specific pattern by the pattern determination means. If it is judged, the word can be registered.
- Response generation means for generating a response corresponding to the input voice can be further provided.
- the registration means can register the word in association with the category as other information.
- the registration unit can register other information in association with the pattern determined to be matched by the pattern determination unit.
- the acquisition means can acquire words by clustering unknown words.
- the network generated by the network generation means can be a network of words and syllables.
- the speech processing apparatus converts an unknown word based on a selection means that selects a path having a high score on the network based on matching between the input speech and the network, and a network that includes the path selected by the selection means. It is possible to further include pronunciation acquisition means for acquiring the corresponding pronunciation.
- the speech processing apparatus further includes a comparison unit that compares an acoustic score when a predetermined section of the input speech is matched with a known word and recognized with a syllable typewriter, and the comparison unit includes a syllable typewriter. If the acoustic score recognized in is superior, it can be estimated that the section is an unknown word.
- the comparison means can make a comparison after correcting the acoustic score when the syllable typifier recognizes the acoustic score when matching with a known word.
- the speech processing method of the present invention includes a recognition step for recognizing continuous input speech, a determination step for determining whether or not an unknown word is included in the recognition result recognized by the processing of the recognition step, and a determination step If it is determined that the recognition result contains an unknown word in the process of, the network generation that generates a network having a path that includes the subword of the time corresponding to the word boundary of the unknown word and a path that does not include the subword If it is determined that the unknown word is included by the processing of the step and the determination step, the acquisition step of acquiring the word corresponding to the unknown word, and the word acquired by the processing of the acquisition step as other information And a registration step of registering in association with each other. '
- the recording medium program of the present invention includes a recognition step for recognizing continuous input speech, a determination step for determining whether or not an unknown word is included in the recognition result recognized by the processing of the recognition step, and a determination step.
- Network that generates a network that includes a path that includes a sub-word of the time corresponding to the word boundary of the unknown word and a path that does not include that sub-word if it is determined that the recognition result contains an unknown word
- the program of the present invention includes a recognition step for recognizing continuous input speech, a determination step for determining whether or not an unknown word is included in the recognition result recognized by the processing of the recognition step, and a process of the determination step If it is determined that an unknown word is included in the recognition result, a network generation that generates a network having a path that includes a subword at a time corresponding to the word boundary of the unknown word and a path that does not include the subword. If it is determined that the unknown word is included in the step and the determination step, the acquisition step acquires a word corresponding to the unknown word, and associates the word acquired in the acquisition step with other information. And a registration step of registering in a computer.
- the recording medium, and the program according to the present invention when continuous input speech is recognized and an unknown word is included in the recognition result, a network is generated, and based on the network. The pronunciation of the unknown word is acquired, the word corresponding to the unknown word is acquired, and the word is registered in association with other information.
- FIG. 1 is a diagram for explaining a mismatch between a word string boundary and a subword sequence boundary.
- FIG. 2 is a diagram for explaining a method for obtaining the “00V> pronunciation” according to the subword sequence.
- Figure 3 shows the experimental conditions when using the sub-word sequence of Fig. 2.
- Fig. 4 is a diagram showing the results of an experiment using the ⁇ 00V> pronunciation acquisition method according to the subword sequence of Fig. 2.
- FIG. 5 is a block diagram showing a configuration example of an embodiment of a dialogue system to which the present invention is applied.
- FIG. 6 is a diagram showing an example of a cluster state.
- FIG. 7 is a diagram showing word registration.
- FIG. 8 is a block diagram illustrating a configuration example of the voice recognition unit in FIG.
- FIG. 9 is a flowchart for explaining the operation of the dialogue system of FIG.
- FIG. 10 is a flowchart for explaining the details of the speech recognition processing in step S 22 of FIG.
- FIG. 11 is a flowchart for explaining details of the word string generation processing in step S 54 in FIG.
- Figure 12 shows an example of grammar used in the language model database.
- Figure 13 shows an example of a language model based on a finite state automaton.
- FIG. 14 is a diagram showing an example of language score calculation using tri_gram.
- FIG. 15 shows an example of the tri_gram database.
- FIG. 16 is a diagram for explaining network generation.
- FIG. 17 is a diagram showing experimental results to which the present invention is applied.
- FIG. 18 is a diagram showing experimental results to which the present invention is applied.
- FIG. 19 is a flowchart for explaining details of the word acquisition process in step S 28 of FIG.
- FIG. 20 is a diagram illustrating an example of a template.
- Figure 21 shows an example of a grammar incorporating a syllable typewriter.
- FIG. 22 is a block diagram showing a configuration example of an embodiment of a computer to which the present invention is applied.
- FIG. 5 shows a configuration example of an embodiment of a dialogue system to which the present invention is applied.
- This dialogue system is a system that is installed, for example, in the mouth pot, and performs dialogue with the user (human) by voice. For example, when voice is input, the name is extracted from the voice and registered. It has become.
- the speech recognition unit 1 receives a speech signal based on the user's utterance, and the speech recognition unit 1 recognizes the input speech signal and the text as a result of the speech recognition. Other accompanying information is output to the dialog control unit 3 and the word acquisition unit 4 as necessary.
- the word acquisition unit 4 automatically stores acoustic features for words that are not registered in the recognition dictionary of the speech recognition unit 1 so that the speech of the words can be recognized thereafter.
- the word acquisition unit 4 obtains pronunciations corresponding to the input speech by the syllable typewriter and classifies them into several clusters.
- Each cluster has ID and a representative syllable sequence, and is managed by ID. The cluster state at this time will be described with reference to FIG.
- the word acquisition unit 4 classifies the three voices into three clusters, “Aka” cluster 2 1, “Ao” cluster 2 2, and “Midori” cluster 2 3, respectively.
- Each cluster has a representative syllable sequence (in the example of Fig. 6, "a / force”, “a / o”, “me / do / li") and ID (in the example of Fig. 6, "1 ”,“ 2 ”,“ 3 ”).
- the word acquisition unit 4 classifies the input voice into “Aka” cluster 2 1 and the new cluster is Do not generate.
- the word acquisition unit 4 newly generates a cluster 2 4 corresponding to “Kuro”. Add a typical syllable sequence (in the example of Fig. 6 "Ku / Kuchi") and an ID (in the example of Fig. 6 "4").
- the associative memory 2 stores information such as a category such as whether the registered name (unknown word) is a user name or a character name.
- a category such as whether the registered name (unknown word) is a user name or a character name.
- the cluster ID and the category name are stored in correspondence.
- the cluster IDs “1”, “3”, “4” correspond to the “user name” category
- the cluster ID “2” corresponds to the “character name” category. Has been.
- the dialogue control unit 3 understands the content of the user's utterance from the output of the speech recognition unit 1, and controls the registration of names (unknown words) based on the understanding result. Further, the dialog control unit 3 controls the subsequent dialogs so that the registered name can be recognized based on the registered name information stored in the associative memory unit 2.
- FIG. 8 shows a configuration example of the speech recognition unit 1.
- the voice recognition unit 1 includes a microphone 41, an AD conversion unit 42, a feature extraction unit 43, a matching unit 44, a syllable typewriter 45, a control unit 46, a network generation unit 47, and an acoustic model database.
- 5 1 Dictionary database 5 2, and language model Database 5 3
- the user's utterance is input to the microphone 41, and the microphone 41 converts the utterance into an audio signal as an electric signal.
- This audio signal is AD
- the AD conversion unit 42 samples the audio signal that is an analog signal from the microphone 41 1, quantizes it, and converts it into audio data that is a digital signal. This audio data is supplied to the feature quantity extraction unit 43.
- the feature quantity extraction unit 43 for the audio data from the AD conversion unit 42, for each appropriate frame, for example, feature parameters such as spectrum, power linear prediction coefficient, cepstrum coefficient, line spectrum pair, etc. Is extracted and supplied to the matching section 4 4 and the syllable type writer section 4 5.
- the matching unit 4 4 Based on the feature parameters from the feature extraction unit 4 3, the matching unit 4 4 refers to the acoustic model / record database 5 1, dictionary database 5 2, and language model / record database 5 3 as necessary.
- the word string closest to the speech (input speech) input to the microphone 41 is obtained and output to the network generation unit 47.
- the acoustic model database 51 stores acoustic models representing acoustic features such as individual phonemes and syllables in a speech language for speech recognition.
- acoustic model for example, HMM (Hidden Markov Model) can be used.
- the dictionary database 52 stores a word dictionary in which information related to pronunciation is described for each word (phrase) to be recognized, and a model in which phonological and syllable chain relationships are described.
- the word here is a unit that is more convenient to treat as a single unit in the recognition process, and does not necessarily match the linguistic word.
- “Taro-kun” may be treated as a single word, or “Taro” and “Kun”.
- a larger unit such as “Kontaro ha Taro” may be treated as one word.
- a syllable is one that is acoustically more convenient to handle as a unit. It is not necessarily consistent with phonetic syllables. For example, “Tokyo”
- the “To” part may be represented by two syllable symbols “To / u”, or “To:” using the symbol “To:” which is the long sound of “To”.
- you may prepare a symbol for silence, which is also used for “silence before speech”, “short silence between speeches”, “silence of speech”, and “tsu”. It is possible to classify them finely like “silence” and prepare symbols for each.
- the language model database 53 describes information on how each word registered in the word dictionary of the dictionary database 52 is linked (connected).
- the syllable typewriter unit 4 5 acquires a syllable sequence corresponding to the input speech based on the feature parameters supplied from the feature amount extraction unit 4 3, outputs the syllable sequence to the matching unit 4 4, and generates a network generation unit 4. Also output to 7.
- the syllable sequence “ ⁇ / Tan / No / Na / Ma / e / ha / o / ga / ⁇ / de / su” is obtained from the voice “My name is Ogawa”.
- Existing syllable typewriters can be used.
- any syllable sequence can be used instead, as long as it can acquire a syllable sequence for any speech.
- any syllable sequence can be used instead, as long as it can acquire a syllable sequence for any speech.
- Japanese phonemes for example, Japanese phonemes
- the control unit 46 controls the operation of the 0 conversion unit 4 2, the feature amount extraction unit 4 3, the matching unit 4 4, and the syllable typewriter unit 4 5.
- the network generation unit 47 generates a network of words and syllables based on the word string and the subword sequence (syllable sequence) based on the syllable sequence acquired from the syllable typewriter unit 45. That is, a path that includes the syllable at the time corresponding to the boundary immediately before 00V> and a path that does not include it, and a path that includes the syllable at the time corresponding to the boundary immediately after ⁇ oov>, and does not include it. Generate path and output to matching unit 4 4 The
- step S 21 when the user inputs sound into the microphone 41, the microphone mouthphone 41 converts the utterance into a sound signal as an electrical signal.
- step S 22 the speech recognition unit 1 executes speech recognition processing.
- step S 51 the audio signal generated by the microphone 4 1 is converted into audio data as a digital signal by the AD conversion unit 42 and supplied to the feature amount extraction unit 43.
- step S 52 the feature quantity extraction unit 43 receives the audio data from the AD conversion unit 42. Then, the feature quantity extraction unit 4 3 proceeds to step S 53 and extracts, for example, feature parameters such as spectrum, power, and their temporal change amount for each appropriate frame, and the matching unit 4 4 To supply.
- step S 54 the matching unit 44 performs concatenation of some of the word models stored in the dictionary database 52 and executes word string generation processing.
- the words constituting this word string include not only known words registered in the dictionary database 52 but also “KUV>” which is a symbol representing an unknown word not registered. This word string generation process will be described in detail with reference to FIG.
- step S81 the matching unit 44 calculates the acoustic score in both cases for a certain section of the input speech. That is, the result obtained by the syllable typewriter unit 45 based on the acoustic score obtained by matching a certain section of the input speech with a known word registered in the dictionary database 52 and the input speech ( In this case, the acoustic scores of ⁇ / t / si / no / na / ma / e / ha / o / ga / ⁇ / de / s ⁇ are calculated respectively.
- the acoustic score represents how close the word sequence that is a candidate for the speech recognition result and the input speech are as sounds.
- a partial section of the input speech and known words registered in the dictionary database 52 Is compared with the acoustic score of the result of the syllable typewriter unit 45, but matching with a known word is performed in units of words, and the matching in the syllable typewriter unit 45 is Since it is performed in syllable units and the scales are different, it is difficult to compare them as they are (generally, the acoustic score in syllable units has a larger value). Therefore, in order to be able to compare with matching scales, the matching unit 44 corrects the acoustic score obtained by the syllable typewriter unit 45 in step S82.
- a process of multiplying the acoustic score from the syllable typewriter unit 45 by a coefficient or subtracting a certain value or a value proportional to the frame length is performed.
- this process since this process is relative, it can also be performed on the acoustic score resulting from matching with known words. The details of this process are disclosed, for example, in the document “EUR0SPEECH99 Volume IV, Page 49-52” as “00V-Detection in Large Vocabulary System Using Automatically Defined Word-Fragments as FilesersJ”. .
- the matching unit 4 4 compares the two acoustic scores in step S 83 (determines whether the acoustic score resulting from the recognition by the syllable typewriter unit 45 is higher (excellent). ) If the acoustic score of the result recognized by the syllable typewriter unit 4 5 is higher, the process proceeds to step S84, and the matching unit 4 4 satisfies the section 00V> (Out Of Vocabulary) (unknown word) Estimated.
- step S83 If it is determined in step S83 that the acoustic score of the result recognized by the syllable typewriter unit 45 is lower than the acoustic score of the matching result with the known word, the process proceeds to step S86.
- the matching unit 44 estimates that the section is a known word.
- the acoustic score of “o / ga / ⁇ ” output from the syllable typewriter unit 45 is compared with the acoustic score when matched with a known word, If the acoustic score of “o / ga / ⁇ ” is higher, “kuo> (o / ga / ⁇ )” is output as the word corresponding to the speech segment, and the acoustic score of the known word is output. If the core is higher, the known word is output as the word corresponding to the speech segment.
- step S85 the matching unit 44 generates preferentially n word strings (concatenated several word models) that are estimated to have a high acoustic score.
- step S 55 the syllable typewriter unit 45 is independent of the process in step S 54, and applies the feature parameters extracted in the process in step S 53. Recognize phonological units and output syllable sequences.
- the syllable typewriter section 4 5 will read “ ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / ⁇ / su”.
- step S 56 the matching unit 44 calculates an acoustic score for each word string generated in step S 54.
- the likelihood is calculated by inputting speech feature parameters for each word sequence (concatenated word models). The method is used.
- a word string that includes oov> it is not possible to obtain the acoustic score of the speech segment corresponding to oov> with the existing method (because there is no word model corresponding to 00V> in advance). .
- the acoustic score of the section is extracted from the recognition result of the syllable typewriter, and the value obtained by correcting the value is used as the acoustic score of oov>. It is further integrated with the acoustic score of other known word parts, which is used as the acoustic score of the word string.
- step S57 the matching unit 44 extracts the upper m word strings (m ⁇ n) having a high acoustic score and uses them as candidate word strings.
- step S 58 the matching unit 4 4 refers to the language model database 53 and calculates a language score for each candidate word string.
- the language score indicates how well the word sequence that is a candidate for the recognition result is suitable as a word. Here we explain in detail how to calculate this language score.
- a grammar corresponding to an unknown word or a finite state automaton (FSA: Finite State Automaton) and tri-grara (which is one of the statistical language models) corresponding to unknown words.
- FSA Finite State Automaton
- tri-grara which is one of the statistical language models
- a word string that fits in the grammar stored in the database is accepted, such as / oku 00VV / ku end> (parsed with this grammar) 1S “ ⁇ start> / you /
- Word strings that do not apply to the grammar stored in the database are not accepted (this grammar is not parsed).
- “ ⁇ Start>” and “ ⁇ End>” are special symbols representing silence before and after the utterance, respectively.
- a parser is used to calculate the language score using this grammar. The parser splits the word sequence into a word sequence that can accept grammar and a word sequence that cannot. That is, for example, a language score of 1 is given to an acceptable word string, and a language score of 0 is given to an unacceptable word string.
- Language score is calculated, and language score 1 (acceptance) is output together.
- whether or not the grammar of the word string can be accepted is determined in advance by the equivalent grammar ( It can also be realized by converting to a finite state automaton (hereinafter referred to as FSA) and determining whether each word string can be accepted by the FSA.
- FSA finite state automaton
- An example of converting the grammar in Figure 12 into an equivalent FSA is shown in Figure 13.
- An FSA is a directed graph consisting of states (nodes) and paths (arcs). As shown in Figure 13, S 1 is the start state and S 16 is the end state. Also, "$ ACTI0N" is actually registered with the name of the action, as in Fig. 12.
- paths with “ ⁇ ” are special transitions that do not consume words (hereinafter referred to as ⁇ transitions). That is, for example, in “ ⁇ start> / I / ha / 00V> / is / end>”, the transition from the initial state S1 to the state S2 consumes the ⁇ head>, and from the state S2 Transition to state S 3 consumes “I”, but the transition from state S 3 to state S 5 is an ⁇ transition, so no words are consumed. That is, it is possible to skip from state S 3 to state S 5 and transition to the next state S 6.
- Whether or not a predetermined word string can be accepted by this FSA is determined by whether or not it can start from the initial state S 1 and reach the end state S 16.
- a statistical language model is a language model that determines the generation probability of the word string and uses it as a language score. That is, for example, in the language model 7 1 in Fig. 1
- the language score of “/ 00V> / is / ⁇ term>” is expressed by the probability of the word string, as shown in the second line. This is further expressed as a product of conditional probabilities, as shown in lines 3-6. For example, “P (no
- the expressions shown in the third to sixth lines in FIG. 14 are approximated with conditional probabilities of three consecutive words as shown in the seventh to ninth lines. These probability values are obtained by referring to the tri-grara database 81 as shown in Fig. 15. This tri-gram database 81 is obtained by analyzing a large amount of text in advance.
- w 1 w 2) of three consecutive words w l, w 2, w 3 is represented. For example, if the three words w l., W 2, w 3 forces are “ku >>,“ me ”,“ no ”, the probability value is 0.12,“ I ”,
- the probability value is set to 0.01, and if “ ⁇ oov>", “is”, or “termination>”, the probability value is set to 0.87. ing.
- the language score can be calculated for a word string including ⁇ oov> by performing entry processing on the language model in the manner of 00V. Therefore, the symbol oov> can be output in the recognition result. Also, when using other types of language models, the language score can be calculated for word strings containing 00V> by performing entry processing for 00VM.
- a language score can be calculated by using a mechanism that maps ⁇ oov> to an appropriate word in the language model. For example, even when using a tri_gram database that does not have “P (00V> I I am)”, the database is accessed with “P (Ogawa I I am)” and the probability described there is “ The language score can be calculated by considering it as the value of “P (ku 00V> I I)”.
- the matching unit 4 4 integrates the acoustic score and the language score in step S 59.
- the matching unit 44 selects the candidate word string having the best score based on the score obtained by integrating both the acoustic score and the language score obtained in step S59. And output as a recognition result.
- step S 59 the integration process in step S 59 is performed. If the language score is 0, the word string is deleted, and if the language score is other than 0, it is left as it is. You may make it the process of leaving.
- step S 2 3 the control unit 4 6 of the speech recognition unit 1 identifies an unknown word in the recognized word string. It is determined whether or not it is included. If it is determined in step S 2 3 that an unknown word is included, the network generation unit 47 generates a word based on the word string and the subword sequence based on the syllable sequence acquired from the syllable type writer unit 45. And a network of subwords (eg syllables).
- a path that includes a syllable at the time corresponding to the boundary immediately before ⁇ 00V> and a path that does not include it, and a path that includes the syllable at the time corresponding to the boundary immediately after 00V> and does not include it.
- a path is generated and output to the matching unit 4 4.
- step S 2 5 the matching section 4 4 starts from the syllable typewriter section 4 5 Based on the subword sequence based on the supplied syllable sequence and the network supplied from the network generation unit 47, the speech input from the user is matched with the network.
- Figure 16 illustrates network creation
- the matching unit 4 4 selects a candidate word string having the best score based on the score obtained by integrating both the obtained acoustic score and language score, and outputs it as a recognition result. For example, FIG. As shown in the above, “ku start> / I /// name / ha / ku 00V> / is / ku end>”.
- the syllable typewriter unit 45 as shown in Fig.
- the syllable “ha” is the word “ha” in the word string. Or included in the word “00V>” in the word string.
- a network of words and syllables is generated.
- a path that includes the syllable at the time corresponding to the boundary immediately before ⁇ oov>, a path that does not include it, and a path that includes the syllable at the time corresponding to the boundary immediately after ⁇ oov>, and a path that does not include it. is generated.
- the predetermined syllable “c” and syllable “ ⁇ ” in the subword sequence Syllable typewriter section corresponding to frame 4 5 Let Ebrator time be t 0 , respectively, and let the time of word strings corresponding to the syllable typewriter times t 0 and tl be t 3 and t 3 , respectively. Also, as shown in Fig. 16C, the interval in Fig. 16!), That is, the interval from 00V> to the word "I" immediately after 00V> is 00V> It is composed of syllables “ha” “o” “ga” “ ⁇ ”that may correspond to the pronunciation of>
- My -name- is -ha-ga- ⁇ -
- My -name- is -ha-ga-
- step S 2 6 the matching unit 4 4 determines the path on the network with a high score (for example, in FIG. 16 based on the result of matching between the voice input from the user and the network). Select path 9 1, path 10 2 and path 10 3).
- step S 2 7 the matching unit 4 4 determines the path on the network selected in step S 2 6 (for example, path 9 1 in FIG. 16, and path 1 0 2 and path 1 0 3). Based on the subword sequence output from the syllable typewriter unit 45, the corresponding pronunciation (“o / ga / ⁇ ” in the example of Fig. 16) is acquired.
- Figures 17 and 18 show the experimental results of applying the present invention. Since the condition setting is the same as the setting in FIG. 3, its description is omitted.
- Figure 17 shows the performance of syllable sequence recognition accuracy, substitution error, deletion error, and insertion error when this method is applied. Each value is expressed as a percentage. The detailed description thereof is the same as in FIG. In the results of Fig. 17, the recognition accuracy is 48.5%. Compared with Fig. 4, the recognition accuracy is better than the recognition accuracy of the subword sequence 00V> 4.2% in the pronunciation acquisition method. doing. The occurrence rates of deletion errors and insertion errors are 11.6% and 8.0%, respectively. The deletion error and insertion error in the ⁇ 00V> pronunciation acquisition method using the subword sequence in Fig. 4 are shown. Comparing the occurrence rate of 33.3% with 4.1%, the occurrence balance of deletion errors and insertion errors is improved (the difference between the two is getting smaller).
- FIG. 18 shows an example of a part of the recognition result assigned to 00V> according to the present invention for the utterance of one speaker.
- “Kazumi” was recognized as “Kazumi j, and no errors occurred.” For example, “Kazumi” was recognized as “Katsu two” in the second time. And “Mi” have substitution errors. In the third round, “Kazumi” was recognized as “Kazumi” and no errors occurred. In the 4th time, “Power Tsumi” is recognized as “Katsumi” and a substitution error has occurred in “Zu”. In the 5th meeting, “Kazumi” is recognized as “Kasumi” and replaced by “zu”. An error has occurred.
- control unit 46 controls the word acquisition unit 4 to execute word acquisition processing in step S 28 and acquire the unknown word.
- step S 1 1 the word acquisition unit 4 extracts feature parameters of the unknown word «00V» from the speech recognition unit 1, in particular, feature parameters related to time information.
- step S 1 1 2 the word acquisition unit 4 determines whether or not the unknown word belongs to the already acquired cluster. When it is determined that it does not belong to the already acquired cluster, the word acquiring unit 4 generates a new cluster corresponding to the unknown word in step S 1 1 3.
- step S 1 1 4 the word acquisition unit 4 outputs the ID of the cluster to which the unknown word belongs to the matching unit 4 4 of the speech recognition unit 1.
- step S 1 1 2 If it is determined in step S 1 1 2 that the unknown word belongs to the already acquired cluster, there is no need to generate a new cluster, so the word acquisition unit 4 skips step S 1 1 3 and Proceed to S 1 1 4, and output the ID of the acquired cluster to which the unknown word belongs to the matching unit 4 4.
- step S 2 9 the dialog control unit 3 determines that the word string acquired in the process of step S 2 8 matches the template. Determine whether or not. That is, it is determined here whether the word string of the recognition result means registration of some name. If it is determined in step S 29 that the word string of the recognition result matches the template, in step S 30, the dialogue control unit 3 stores the name cluster ID in the associative storage unit 2. And store the corresponding categories.
- This template 1 2 1 indicates that when the word string of the recognition result matches the regular expression on the left side of the figure, the operation on the right side of the figure is executed.
- the recognition result is the word string “ku start> / I / no / name / ha / ku 00V> (o / ga / ⁇ ) / do / ku end>”
- the character generated from this recognition result The column “My name is 00V>” matches the second regular expression in Figure 20. Therefore, the corresponding operation “register cluster ID corresponding to 00V> as a user name” is executed. That is, if the cluster ID of “ku 00V> (o / ga / ⁇ )” is “1”, the category name of cluster ID “1” is registered as “user name” as shown in FIG.
- the template 1 2 1 and the associative memory unit 2 can be simplified. For example, it is possible to store only the cluster ID in the associative memory unit 2 by storing the ID of the template 1 2 1 as “if the recognition result nikku 00V> is included”.
- the dialog control unit 3 reflects the information registered in the associative memory unit 2 in the subsequent conversation determination process. For example, on the side of the dialogue system, “Determine whether or not the name of the dialogue character is included in the user's utterance.
- the dialogue control unit 3 refers to the information in the associative memory unit 2 and corresponds to the dialogue character. Equivalent to a word (entry whose category name is “character name”) or user name You can get words (entries whose category name is “username”).
- step S 3 In 1 the dialogue control unit 3 generates a response corresponding to the input voice. That is, in this case, the name (unknown word) registration process is not performed, and a predetermined process corresponding to the input voice from the user is performed.
- the part corresponding to "$ 00V" is output as multiple symbols.
- the recognition result of “My name is Ogawa” becomes “ ⁇ start> / I / no / name / has / o / ga / ⁇ / is / termination>”. The result is " ⁇
- step S2 3 in Fig. 9 the processing after step S2 3 in Fig. 9 is executed in the same way as when using the syllable typewriter. can do.
- step S 2 8 in the flow chart of FIG. 9 may be omitted.
- step S 28 is omitted, the process in step S 30 is performed by dialog control unit 3 in associative memory unit 2. The process is to store the support series and the category in association with each other.
- a network of words with syllables is formed, but the likelihood for the input speech is maximized. Since it is only necessary to be able to select an optimal subword sequence on the network, it is possible to create a network of subwords and words other than syllables, such as phonemes and phonemes, as well as syllables.
- FIG. 22 shows a configuration example of a personal computer 150 that executes the above-described processing.
- the Sonanore computer 1 5 0 has a CPU (Central Processing Unit) 1 5 1 built-in.
- An input / output interface 1 5 5 is connected to the CPU 1 5 1 through a bus 1 5 4.
- the node 15 4 is connected to a read only memory (R0M) 15 2 and a random access memory (RAM) 15 3.
- R0M read only memory
- RAM random access memory
- the input / output interface 1 5 5 includes an input unit 1 5 7 composed of input devices such as a mouse, a keyboard, a microphone and an AD converter operated by a user, a display, a speaker, and a DA converter.
- the output unit 1 5 6 composed of output devices such as is connected.
- the input / output interface 1 5 5 includes a storage unit 1 5 8 composed of a hard disk drive for storing programs and various data, and a communication unit 1 5 9 for communicating data via a network represented by the Internet. Is connected.
- the I / O interface 1 5 5 has a drive 1 6 0 that reads and writes data to and from a recording medium such as a magnetic disk 1 7 1, an optical disk 1 7 2, a magneto-optical disk 1 7 3, and a semiconductor memory 1 7 4 Connected as needed.
- a recording medium such as a magnetic disk 1 7 1, an optical disk 1 7 2, a magneto-optical disk 1 7 3, and a semiconductor memory 1 7 4 Connected as needed.
- An audio processing program for causing the personal computer 1550 to execute an operation as an audio processing apparatus to which the present invention is applied includes a magnetic disk 1 71 (including a floppy disk), an optical disk 1 7 2 (CD-ROM (Compact Di (including sc-Read Only Memory), DVD (Digital Versatile Disc), magneto-optical disk 1 7 3 (including MD (Mini Disc)), or semiconductor memory 1 7 4 stored in personal computer 1 5 0, read by the drive 1 60, and installed in the hard disk drive built in the storage unit 1 5 8.
- the voice processing program installed in the storage unit 1 5 8 is transferred from the storage unit 1 5 8 to the RAM I according to the command of the CPU 15 1 corresponding to the command from the user input to the input unit 1 5 7. 5 Spoken by 3 and executed.
- the series of processes described above can be executed by hardware, or can be executed by software.
- the program that constitutes the software must execute various functions by installing a computer embedded in a dedicated hard disk or various programs. For example, it can be installed from a network or recording medium on a general-purpose personal computer.
- this recording medium is distributed to provide a program to the user separately from the main body of the device, and the magnetic disc 1 7 1, optical disc 1 7 2, R0M 1 5 in which programs are recorded that are provided not only to package media consisting of the magneto-optical disk 1 7 3 or semiconductor memory 1 7 4 but also to the user in a pre-installed state in the main body of the device. 2 and hard disk included in the storage unit 1 5 8.
- the steps for describing a program recorded on a recording medium are not limited to processes performed in time series according to the described order, but are not necessarily processed in time series. Or it includes processing that is executed individually.
- system means a logical collection of a plurality of devices, and it does not matter whether or not the devices of each configuration are in the same casing.
- words can be registered by voice.
- the registration can be executed without making the user aware of the registration mode.
- the registered words can be reflected in subsequent dialogues.
- the recognition rate of voice recognition can be improved. Furthermore, it is possible to reduce deletion errors when acquiring unknown words, and to improve the occurrence balance of deletion errors and insertion errors. As a result, it is possible to prevent the user from feeling unwise about a robot having a continuous speech recognition system.
Landscapes
- Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Multimedia (AREA)
- Artificial Intelligence (AREA)
- Machine Translation (AREA)
Abstract
Description
Claims
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP03772681A EP1460615B1 (en) | 2002-11-21 | 2003-11-12 | Voice processing device and method, recording medium, and program |
US10/502,169 US7653541B2 (en) | 2002-11-21 | 2003-11-12 | Speech processing device and method, and program for recognition of out-of-vocabulary words in continuous speech |
DE60318385T DE60318385T2 (de) | 2002-11-21 | 2003-11-12 | Sprachverarbeitungseinrichtung und -verfahren, aufzeichnungsmedium und programm |
AU2003280726A AU2003280726A1 (en) | 2002-11-21 | 2003-11-12 | Voice processing device and method, recording medium, and program |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2002337892A JP4072718B2 (ja) | 2002-11-21 | 2002-11-21 | 音声処理装置および方法、記録媒体並びにプログラム |
JP2002-337892 | 2002-11-21 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2004047075A1 true WO2004047075A1 (ja) | 2004-06-03 |
Family
ID=32321860
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2003/014342 WO2004047075A1 (ja) | 2002-11-21 | 2003-11-12 | 音声処理装置および方法、記録媒体並びにプログラム |
Country Status (8)
Country | Link |
---|---|
US (1) | US7653541B2 (ja) |
EP (1) | EP1460615B1 (ja) |
JP (1) | JP4072718B2 (ja) |
KR (1) | KR101014086B1 (ja) |
CN (1) | CN100354929C (ja) |
AU (1) | AU2003280726A1 (ja) |
DE (1) | DE60318385T2 (ja) |
WO (1) | WO2004047075A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8155958B2 (en) | 2005-11-08 | 2012-04-10 | Nec Corporation | Speech-to-text system, speech-to-text method, and speech-to-text program |
Families Citing this family (18)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7813928B2 (en) * | 2004-06-10 | 2010-10-12 | Panasonic Corporation | Speech recognition device, speech recognition method, and program |
JP4301102B2 (ja) * | 2004-07-22 | 2009-07-22 | ソニー株式会社 | 音声処理装置および音声処理方法、プログラム、並びに記録媒体 |
RU2008127509A (ru) * | 2005-12-08 | 2010-01-20 | Конинклейке Филипс Электроникс Н.В. (Nl) | Система распознавания речи с огромным словарем |
WO2007097390A1 (ja) * | 2006-02-23 | 2007-08-30 | Nec Corporation | 音声認識システム、音声認識結果出力方法、及び音声認識結果出力プログラム |
US9245526B2 (en) * | 2006-04-25 | 2016-01-26 | General Motors Llc | Dynamic clustering of nametags in an automated speech recognition system |
JP5366169B2 (ja) * | 2006-11-30 | 2013-12-11 | 独立行政法人産業技術総合研究所 | 音声認識システム及び音声認識システム用プログラム |
US8868410B2 (en) * | 2007-08-31 | 2014-10-21 | National Institute Of Information And Communications Technology | Non-dialogue-based and dialogue-based learning apparatus by substituting for uttered words undefined in a dictionary with word-graphs comprising of words defined in the dictionary |
CN101452702B (zh) * | 2007-12-05 | 2011-08-10 | 财团法人工业技术研究院 | 语音模型的调整方法及其调整模块 |
US8326631B1 (en) * | 2008-04-02 | 2012-12-04 | Verint Americas, Inc. | Systems and methods for speech indexing |
JP5810912B2 (ja) * | 2011-12-28 | 2015-11-11 | 富士通株式会社 | 音声認識装置、音声認識方法及び音声認識プログラム |
CN103076893B (zh) * | 2012-12-31 | 2016-08-17 | 百度在线网络技术(北京)有限公司 | 一种用于实现语音输入的方法与设备 |
CN104157285B (zh) * | 2013-05-14 | 2016-01-20 | 腾讯科技(深圳)有限公司 | 语音识别方法、装置及电子设备 |
KR102117082B1 (ko) | 2014-12-29 | 2020-05-29 | 삼성전자주식회사 | 음성 인식 방법 및 음성 인식 장치 |
JP6655835B2 (ja) * | 2016-06-16 | 2020-02-26 | パナソニックIpマネジメント株式会社 | 対話処理方法、対話処理システム、及びプログラム |
JP6711343B2 (ja) | 2017-12-05 | 2020-06-17 | カシオ計算機株式会社 | 音声処理装置、音声処理方法及びプログラム |
US11232783B2 (en) | 2018-09-12 | 2022-01-25 | Samsung Electronics Co., Ltd. | System and method for dynamic cluster personalization |
US20220335951A1 (en) * | 2019-09-27 | 2022-10-20 | Nec Corporation | Speech recognition device, speech recognition method, and program |
JP6807586B1 (ja) * | 2020-01-30 | 2021-01-06 | ベルフェイス株式会社 | 情報処理装置、情報処理方法及びプログラム |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20010037200A1 (en) * | 2000-03-02 | 2001-11-01 | Hiroaki Ogawa | Voice recognition apparatus and method, and recording medium |
Family Cites Families (20)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0384707A (ja) * | 1989-08-28 | 1991-04-10 | Victor Co Of Japan Ltd | 垂直磁気記録方式 |
JPH06266386A (ja) | 1993-03-16 | 1994-09-22 | Nippon Telegr & Teleph Corp <Ntt> | ワードスポッティング方法 |
MX9703138A (es) * | 1994-11-01 | 1997-06-28 | British Telecomm | Reconocimiento de lenguaje. |
JP3576272B2 (ja) | 1995-06-22 | 2004-10-13 | シャープ株式会社 | 音声認識装置および方法 |
JP2886117B2 (ja) | 1995-09-11 | 1999-04-26 | 株式会社エイ・ティ・アール音声翻訳通信研究所 | 音声認識装置 |
JP3459712B2 (ja) * | 1995-11-01 | 2003-10-27 | キヤノン株式会社 | 音声認識方法及び装置及びコンピュータ制御装置 |
US5960395A (en) * | 1996-02-09 | 1999-09-28 | Canon Kabushiki Kaisha | Pattern matching method, apparatus and computer readable memory medium for speech recognition using dynamic programming |
JPH09274496A (ja) | 1996-04-09 | 1997-10-21 | Fuji Xerox Co Ltd | 音声認識装置 |
JP2999726B2 (ja) | 1996-09-18 | 2000-01-17 | 株式会社エイ・ティ・アール音声翻訳通信研究所 | 連続音声認識装置 |
JP3428309B2 (ja) | 1996-09-24 | 2003-07-22 | 松下電器産業株式会社 | 音声認識装置 |
US5956668A (en) * | 1997-07-18 | 1999-09-21 | At&T Corp. | Method and apparatus for speech translation with unrecognized segments |
JPH1185179A (ja) | 1997-09-16 | 1999-03-30 | Nippon Telegr & Teleph Corp <Ntt> | 統計的言語モデルの作成方法 |
US5983180A (en) * | 1997-10-23 | 1999-11-09 | Softsound Limited | Recognition of sequential data using finite state sequence models organized in a tree structure |
JP3914709B2 (ja) * | 1997-11-27 | 2007-05-16 | 株式会社ルネサステクノロジ | 音声認識方法およびシステム |
JP3476008B2 (ja) * | 1999-09-10 | 2003-12-10 | インターナショナル・ビジネス・マシーンズ・コーポレーション | 音声情報の登録方法、認識文字列の特定方法、音声認識装置、音声情報の登録のためのソフトウエア・プロダクトを格納した記憶媒体、及び認識文字列の特定のためのソフトウエア・プロダクトを格納した記憶媒体 |
US7310600B1 (en) * | 1999-10-28 | 2007-12-18 | Canon Kabushiki Kaisha | Language recognition using a similarity measure |
JP4543294B2 (ja) * | 2000-03-14 | 2010-09-15 | ソニー株式会社 | 音声認識装置および音声認識方法、並びに記録媒体 |
DE60016722T2 (de) * | 2000-06-07 | 2005-12-15 | Sony International (Europe) Gmbh | Spracherkennung in zwei Durchgängen mit Restriktion des aktiven Vokabulars |
DE10042942C2 (de) * | 2000-08-31 | 2003-05-08 | Siemens Ag | Verfahren zur Sprachsynthese |
GB0027178D0 (en) * | 2000-11-07 | 2000-12-27 | Canon Kk | Speech processing system |
-
2002
- 2002-11-21 JP JP2002337892A patent/JP4072718B2/ja not_active Expired - Fee Related
-
2003
- 2003-11-12 CN CNB2003801003011A patent/CN100354929C/zh not_active Expired - Fee Related
- 2003-11-12 US US10/502,169 patent/US7653541B2/en not_active Expired - Fee Related
- 2003-11-12 KR KR1020047011216A patent/KR101014086B1/ko not_active IP Right Cessation
- 2003-11-12 EP EP03772681A patent/EP1460615B1/en not_active Expired - Fee Related
- 2003-11-12 DE DE60318385T patent/DE60318385T2/de not_active Expired - Lifetime
- 2003-11-12 WO PCT/JP2003/014342 patent/WO2004047075A1/ja active IP Right Grant
- 2003-11-12 AU AU2003280726A patent/AU2003280726A1/en not_active Abandoned
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20010037200A1 (en) * | 2000-03-02 | 2001-11-01 | Hiroaki Ogawa | Voice recognition apparatus and method, and recording medium |
Non-Patent Citations (3)
Title |
---|
NAKAWATASE KIMOTO NAKAGAWA: "Tokeiteki shuho ni yoru jisho mitoroku no kakutokuho", THE TRANSACTIONS OF THE INSTITUTE OF ELECTRONICS, INFORMATION AND COMMUNICATION ENGINEER, D-II, vol. J81-D-II, no. 2, 25 February 1998 (1998-02-25), pages 238 - 248, XP002986474 * |
OGAWA: "Robot taiwa ni okeru shizen na shinki goi no kakutoku", THE JAPANESE SOCIETY FOR ARTIFICIAL INTELLIGENCE AI CHALLENGE KENKYUKAI (DAI 16 KAI), 22 November 2002 (2002-11-22), pages 13 - 18, XP002986473 * |
See also references of EP1460615A4 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8155958B2 (en) | 2005-11-08 | 2012-04-10 | Nec Corporation | Speech-to-text system, speech-to-text method, and speech-to-text program |
Also Published As
Publication number | Publication date |
---|---|
JP4072718B2 (ja) | 2008-04-09 |
DE60318385D1 (de) | 2008-02-14 |
US20050143998A1 (en) | 2005-06-30 |
EP1460615B1 (en) | 2008-01-02 |
KR101014086B1 (ko) | 2011-02-14 |
KR20050083547A (ko) | 2005-08-26 |
DE60318385T2 (de) | 2009-01-02 |
JP2004170765A (ja) | 2004-06-17 |
EP1460615A1 (en) | 2004-09-22 |
EP1460615A4 (en) | 2007-04-25 |
CN1692405A (zh) | 2005-11-02 |
CN100354929C (zh) | 2007-12-12 |
US7653541B2 (en) | 2010-01-26 |
AU2003280726A1 (en) | 2004-06-15 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP4301102B2 (ja) | 音声処理装置および音声処理方法、プログラム、並びに記録媒体 | |
EP1575030B1 (en) | New-word pronunciation learning using a pronunciation graph | |
Hirsimaki et al. | Importance of high-order n-gram models in morph-based speech recognition | |
US7676365B2 (en) | Method and apparatus for constructing and using syllable-like unit language models | |
WO2004047075A1 (ja) | 音声処理装置および方法、記録媒体並びにプログラム | |
JP2003316386A (ja) | 音声認識方法および音声認識装置および音声認識プログラム | |
JP2001188781A (ja) | 会話処理装置および方法、並びに記録媒体 | |
JP2019012095A (ja) | 音素認識辞書生成装置および音素認識装置ならびにそれらのプログラム | |
Livescu | Analysis and modeling of non-native speech for automatic speech recognition | |
Rabiner et al. | Statistical methods for the recognition and understanding of speech | |
Abate et al. | Syllable-based speech recognition for Amharic | |
JP4600706B2 (ja) | 音声認識装置および音声認識方法、並びに記録媒体 | |
Lee et al. | A survey on automatic speech recognition with an illustrative example on continuous speech recognition of Mandarin | |
JPH08123470A (ja) | 音声認識装置 | |
Gauvain et al. | Large vocabulary continuous speech recognition: from laboratory systems towards real-world applications | |
JP4048473B2 (ja) | 音声処理装置および音声処理方法、並びにプログラムおよび記録媒体 | |
JPH1097293A (ja) | 音声認識用単語辞書作成装置及び連続音声認識装置 | |
Syadida et al. | Sphinx4 for indonesian continuous speech recognition system | |
Szarvas et al. | Finite-state transducer based Hungarian LVCSR with explicit modeling of phonological changes | |
JPH09114482A (ja) | 音声認識のための話者適応化方法 | |
Tjalve | Accent features and idiodictionaries: on improving accuracy for accented speakers in ASR | |
KR100608644B1 (ko) | 가변 길이의 연결숫자음 인식방법 | |
JP2000315095A (ja) | 音声認識装置、音声認識方法及び音声認識制御プログラムを記録した記録媒体 | |
JPH0981185A (ja) | 連続音声認識装置 | |
Şahin | Language modeling for Turkish continuous speech recognition |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): AE AG AL AM AT AU AZ BA BB BG BR BY BZ CA CH CN CO CR CU CZ DE DK DM DZ EC EE EG ES FI GB GD GE GH GM HR HU ID IL IN IS KE KG KP KR KZ LC LK LR LS LT LU LV MA MD MG MK MN MW MX MZ NI NO NZ OM PG PH PL PT RO RU SC SD SE SG SK SL SY TJ TM TN TR TT TZ UA UG US UZ VC VN YU ZA ZM ZW |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): BW GH GM KE LS MW MZ SD SL SZ TZ UG ZM ZW AM AZ BY KG KZ MD RU TJ TM AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HU IE IT LU MC NL PT RO SE SI SK TR BF BJ CF CG CI CM GA GN GQ GW ML MR NE SN TD TG |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2003772681 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 10502169 Country of ref document: US Ref document number: 1020047011216 Country of ref document: KR |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application | ||
WWE | Wipo information: entry into national phase |
Ref document number: 20038A03011 Country of ref document: CN |
|
WWP | Wipo information: published in national office |
Ref document number: 2003772681 Country of ref document: EP |
|
WWG | Wipo information: grant in national office |
Ref document number: 2003772681 Country of ref document: EP |