WO2000025299A1 - Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen - Google Patents
Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen Download PDFInfo
- Publication number
- WO2000025299A1 WO2000025299A1 PCT/DE1999/003176 DE9903176W WO0025299A1 WO 2000025299 A1 WO2000025299 A1 WO 2000025299A1 DE 9903176 W DE9903176 W DE 9903176W WO 0025299 A1 WO0025299 A1 WO 0025299A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- classes
- linguistic
- probability
- language model
- determined
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; 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
- G10L15/19—Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules
- G10L15/197—Probabilistic grammars, e.g. word n-grams
Definitions
- the invention relates to a method and an arrangement for class formation for a language model based on linguistic classes by a computer.
- a method for speech recognition is known from [1]. As part of word recognition, it is customary to specify the usability of a sequence of at least one word. A measure of this usability is a probability.
- the probability P (W) for a word sequence W generally characterizes a (statistical) language model in the context of speech recognition, preferably large amounts of vocabulary.
- the probability P (W) (so-called:
- a linguistic lexicon is known from [4]. This is a compilation of as many words of a language as possible available on a computer for the purpose of looking up linguistic properties using a search program. For each word entry (so-called full word form), the linguistic features relevant for this full word form and the appropriate assignments, i.e. the linguistic values, can be found.
- Table 1 Examples of ling. Characteristics and ling. values
- mapping rule F assigns at least one linguistic class to each word, using the following mapping rule F:
- f m is a linguistic characteristic
- m the number of linguistic characteristics
- v m l the number of linguistic characteristics
- v mj the linguistic values of the linguistic
- a special linguistic class represents the class of words with unknown or otherwise mapped linguistic properties.
- class bigrams that is, bigrams applied to linguistic classes
- class Ci represents the correct combination of category, number, case and gender with regard to the example sentence.
- C —C 7 follows for the above class bigram a numerous occurrence, since this combination often occurs in the German language, whereas other class bigrams, eg the combination C2-Cg, are not allowed in the German language because of different geni.
- the class bigram probabilities resulting from the frequencies found in this way are correspondingly high (if they occur frequently) or low (if not permissible).
- the object of the invention is to enable class formation for a language model based on linguistic classes automatically and without the use of expert knowledge.
- a method for class formation for a language model based on linguistic classes is specified by means of a computer, in which a number N classes is determined on the basis of a first mapping rule by means of a predetermined vocabulary with associated linguistic properties.
- K classes (K ⁇ N) are determined from N classes by minimizing a language model entropy H.
- the K classes represent a second mapping rule, the class formation of the language model. It is advantageous here that class formation can be determined fully automatically. Neither a specially trained expert takes on a laborious manual assignment, nor does statistical measures soften the linguistic meaning of the classes.
- the condition that K is less than N significantly reduces the number of classes and thus determines a high-performance language model.
- N classes are determined by determining all possible combinations of linguistic features and associated linguistic values and each of the combinations leads to a separate linguistic class.
- the number N is therefore determined by the maximum possible number of classes (based on the underlying text).
- a method for class formation for a language model based on linguistic classes is also specified by a computer, in which N classes are specified on the basis of a first mapping rule. K classes are determined from the N classes by minimizing a language model entropy. Using the K classes, a second mapping rule for class formation of language models based on linguistic classes is presented.
- the K classes are determined by performing the following steps: a) A number M of the most probable among the N classes are determined as base classes; b) one of the remaining (NM) classes is merged with the base class in which the language model entropy is minimized.
- the M most probable classes are determined.
- the above steps can also be carried out iteratively for several or all remaining (N-M) classes.
- H the language model entropy of the language model
- n the number of words in the text
- W a chain of words W Q , W] _, .., w n , P (W) describe a probability of the occurrence of a sequence of at least two words.
- Another embodiment is that the method described is used to determine a probability of a sequence of at least two words occurring in speech recognition.
- a language has linguistic classes
- At least one of the linguistic classes is assigned to a word.
- a probability P (W) for the occurrence of the sequence of at least two words is obtained using bigrams
- Ci-l a linguistic class belonging to one
- Ci relates to one of the at least one linguistic class which is assigned to the word i from the word sequence W.
- C _ ⁇ the class bigram probability is the probability that the word wi belongs to a first linguistic class on the condition that the previous word wi_ ⁇ belongs to a second linguistic class (see introductory example with explanation).
- Language models based on linguistic classes offer decisive advantages, especially for adaptation.
- the method presented here uses the linguistic properties contained in the language models.
- Probability P (CilC _] _) is adopted in the base language model for the new text.
- the vocabulary for the new domain for which a language model is determined is processed with the aid of a predetermined linguistic lexicon and using a classifier F according to equation (3). At least one linguistic class is automatically determined for each new word from the text. For a detailed description of linguistic classes, linguistic characteristics and linguistic values see [3], for the linguistic lexicon see [4] and / or the introduction.
- An additional development consists in adapting the basic language model based on the determined probability P (wil Ci). This is preferably done in such a way that these determined probabilities P (wil Ci) are included in the basic language model.
- Another further development consists in recognizing a corresponding sequence of at least one word, if the
- Probability P (W) lies above a given limit. If this is not the case, a specified action is carried out. This specified action is e.g. issue an error message or cancel the procedure.
- the text relates to a predetermined application area, a so-called (language, application) domain.
- the method presented requires a new text of only a small extent for the determination of a language model of a new domain.
- an arrangement for class formation for a language model is based on to solve the task Linguistic classes specified, which has a processor unit, which processor unit is set up in such a way that a) a number N classes can be determined by means of a predetermined one, based on a first mapping rule
- K classes are determined from the N classes by minimizing a language model entropy; c) on the basis of the K classes there is a second mapping rule for class formation from language models to linguistic classes.
- an arrangement for class formation for a language model based on linguistic classes in which a processor unit is provided which is set up in such a way that a) N classes can be specified using a first mapping rule; b) K classes are determined from the N classes by minimizing a language model entropy; c) on the basis of the K classes there is a second mapping rule for class formation from language models to linguistic classes.
- Fig.l is a block diagram showing the steps of a method for determining a probability of occurrence comprises a sequence of at least one word in speech recognition by a computer;
- FIG. 5 shows a block diagram with components for the automatic determination of a mapping rule for class formation
- FIG. 6 shows a block diagram for optimizing an existing language model
- FIG. 8 shows a processor unit
- FIG. 4 shows a block diagram with steps of a method for class formation for a language model.
- N classes are determined according to a predetermined condition.
- One possibility is to determine the N classes as all the maximum possible classes by determining all possible combinations of linguistic features and associated linguistic values and each of the combinations resulting in a separate class (cf. step 402).
- K classes are determined using the N classes under Taking into account the condition that a language model entropy is minimized.
- the K classes obtained in this way represent a second mapping rule (cf. step 404), on the basis of which classes are formed according to linguistic classes for a language model.
- an existing language model with a first mapping rule and N predetermined classes is assumed (see step 405).
- the subsequent steps 403 and 404 allow an adaptation of the existing language model by in turn forming a class for a language model which is optimized in relation to the original language model with regard to the linguistic classes.
- mapping rule for class formation based on linguistic properties that minimizes the language model entropy of the language model generated via these classes is sought in the present case (is also referred to below as an optimization criterion).
- an optimization criterion To generate language models on linguistic classes is one
- Mapping rule necessary that assigns at least one linguistic class to each word.
- Class formation is based on linguistic properties. According to the linguistic characteristics according to equation (0-2) and the linguistic values according to equation (0-3), one or more classes are assigned to each word. The linguistic characteristics and the linguistic values are taken from a linguistic lexicon (cf. [4]).
- mapping rule (classifier) F of the linguistic features and their linguistic values on classes is defined according to equation (3).
- a mapping rule L is determined which, given the vocabulary and linguistic properties given from the linguistic lexicon, generates the maximum possible number N of classes:
- the source of knowledge for this is a training corpus representing the given domain and a linguistic lexicon comprising its vocabulary (cf. block 501 in FIG. 5).
- the language model is trained on the maximum classes N (see block 502) (see block 503).
- the optimization takes place in a block 504:
- the maximum possible classes N are combined in such a way that an optimization criterion is fulfilled.
- a language model is determined based on the new K classes (see block 505).
- each class-based language model can be subjected to optimization.
- N classes are specified (cf. block 601, FIG. 6)
- the optimization loop (block 602) and the subsequent calculation of the new language model based on the optimized classes (block 603) are analogous to FIG. 5.
- OPTp ⁇ is searched which minimizes the entropy H (LM) of the language model LM (OPTM).
- the language model is based on the class division determined by OPTJVJ:
- ie C 0 is the union (cluster) of classes from the maximum class set.
- the union takes place via linguistic characteristics and linguistic values of the classes to be united. For example is Ci C * ⁇ il A v B ⁇ (11)
- Equation (1) The speech model entropy H (LM) is given by equation (1), where P (W) can be an approximate value. Equation (4) applies to the value P (W).
- FIG. 7 shows an optimization strategy in the form of a flow chart.
- the classes N are merged. Taking into account all possibilities of merging classes is extremely complex in practice.
- the procedure is therefore preferably different: M be the desired number of optimized classes.
- the probability values of the language model of the N classes are used to determine the most probable M classes among the N classes as base classes.
- the remaining N-M classes form the remaining classes (cf. step 701 in FIG. 7). Within each loop of optimization, one of the remainder classes becomes one
- Base class merged so that an increase in the language model entropy is minimized (see steps 702 to 705). If two classes are merged, the probabilities that are necessary to determine the growth of the language model entropy are recalculated (see step 706).
- a processor unit PRZE is shown in FIG.
- the processor unit PRZE comprises a processor CPU, a memory SPE and an input / output interface IOS, which are connected via an interface IFC in different ways is used: Via a graphics interface, output is visible on a MON monitor and / or output on a PRT printer. An entry is made using a mouse MAS or a keyboard TAST.
- the processor unit PRZE also has a data bus BUS, which ensures the connection of a memory MEM, the processor CPU and the input / output interface IOS.
- additional components can be connected to the data bus BUS, for example additional memory, data storage (hard disk) or scanner.
- a step 101 speech is converted into linguistic classes
- f m is a linguistic characteristic
- m the number of linguistic characteristics
- v ml the number of linguistic characteristics
- v mj the linguistic values of the linguistic
- Characteristic f m , j the number of linguistic values
- F denotes a mapping rule (classifier) of linguistic features and linguistic values on linguistic classes.
- classification rule classifier
- a detailed explanation of the linguistic characteristics and the linguistic values can be found in [3], for example on page 1201 in Table 4 an exemplary listing of linguistic characteristics with associated linguistic values depending on different categories is shown.
- a step 102 at least one of the linguistic classes is assigned to a word. As described in [3], one or more of the linguistic classes can be assigned to a word.
- Ci a linguistic class belonging to a word wi
- Cii_- ⁇ i a linguistic class belonging to a linguistic class belonging to a
- P P (wi I Ci) the conditional word probability
- Equation (4) consists of a cascaded multiplication of three components
- Vocabulary of the text for the new domain is assigned to linguistic classes using a linguistic lexicon using a classifier F as shown in equation (3).
- Each new word is automatically assigned to at least one linguistic class.
- a basic language model includes probabilities for class bigrams [3], whereby this probability represents a grammatical structure on the one hand and is independent of the individual words on the other hand. If it is now assumed that the domain, i.e. the special application-related subject area, has a text structure similar to that of the training text on which the basic language model is based, the probability for class bigrams P (CilCi _] _) is adopted unchanged from the basic language model.
- the probability P (wil Ci) for all words wi that are new with respect to the basic language model must be recalculated and the probability P (wil Ci) (word probability) of the vocabulary present in the basic language model should preferably be adapted accordingly.
- P (wil Ci) word probability
- the probability P (il Ci) for all new words Wi of the new domain is estimated on the basis of the text for the new domain. It is based on a basic language model based on linguistic classes, the newly estimated probability P (wil Ci) preferably in the
- Basic language model is adopted and thus the basic language model is adapted based on the new text.
- Word wi is estimated based on the new text.
- Classifier F 211 and linguistic lexicon 206 is generated using the tagging tool 202 (see detailed Explanations of the tagging tool under [3]) both from a database of large texts 201 a large "tagged" text 203 and from a database of a small text of the new domain (ie the new text) 207 a small "tagged” Text 208 determines.
- a basic language model 205 which is based on linguistic classes, is determined from the large “tagged” text 203 by means of a language model generator 204. As described in detail above, the probability P (CilCi_ ⁇ ) goes unchanged into the language model for the new domain.
- the "tagged" small text 208 is converted using an adaptation tool 209, which estimates the probability P (wil Ci) using the
- carries out "tagged" small text determines a new, preferably adapted, language model 210.
- another language model can also be generated without restriction.
- the classifier F 311 and the linguistic lexicon 306 are used to determine a "tagged" large text 303 from a database of large texts 301 using the tagging tool 302.
- a "basic language model 305" based on linguistic classes is created from the "tagged" large text 303.
- Ci_ ⁇ ) is taken over unchanged from the basic language model 305.
- an adaptation tool 308 is used adapted language model 309 determined.
- an adaptation can include a change or a generation of a language model.
- the adaptation tool 308 calculates the probability P (wil Ci) for new words from the probability P (wi) and renormalizes the probability P (wil Ci) of the basic language model.
- the creation of linguistic classes of a language model for the new domain corresponds to the creation of linguistic classes for the basic language model.
- the classifier F (see equation (3)) of the basic language model is adopted. So the number of linguistic classes k is unchanged.
- the new domain is based on texts with a similar structure to the training texts on which the basic language model is based.
- the probability of the class bigrams P (CilCi_ ⁇ ) and the probability of class unramrams P (C j ) of the basic language model remain unchanged.
- remains to be considered C j (wi)) and the probability P (Cj (i)
- wi) become for the words of the new domain that are not contained in the basic language model are recalculated. Existing probabilities for the words of the basic language model must be rescaled.
- P (w * h) class-independent word probabilities in the new domain.
- the probability P (wft) is given by a word list with word frequencies and the size of the underlying text.
- wi denotes all words of the basic language model that are in class Cj.
- class C j is examined as an example for the following explanations. For the sake of simplicity, this class Cj is referred to as class C in the following.
- equation (12) Another approximation for equation (12) is the sum over the words wi, for which all linguistic classes match the classes of the new word.
- Equation (22) The factor can be interpreted as the quotient of the proportions of old words in linguistic class C and the proportion of old vocabulary in the overall vocabulary.
- the probability P (wil Ci) for words wi that are new with respect to the basic language model is approximated using a corresponding word list.
- the solution strategy is adopted from the procedure described under 'Method 2'.
- the probability P (w ⁇ ) for the new words w- ⁇ , which does not exist here, is approximated. This takes place in particular depending on a main category HC of the respective word h-
- JJC Base language model used. JJC is a number of
- the words ⁇ can be assigned to the special linguistic class of words with unknown or otherwise mapped linguistic properties. Determination of the probability P (Ci_ ⁇
- Index 'i-1' which is subsequently set as index i for simplicity.
- the linguistic class Ci runs through all possible linguistic classes for the word wi.
- the probability P (Ci) is taken from the basic language model (unigram probability for the respective linguistic class of the basic language model).
Abstract
Description
Claims
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE59901575T DE59901575D1 (de) | 1998-10-27 | 1999-10-01 | Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen |
EP99957254A EP1135767B1 (de) | 1998-10-27 | 1999-10-01 | Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen |
US09/844,931 US6640207B2 (en) | 1998-10-27 | 2001-04-27 | Method and configuration for forming classes for a language model based on linguistic classes |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE19849546.3 | 1998-10-27 | ||
DE19849546 | 1998-10-27 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US09/844,931 Continuation US6640207B2 (en) | 1998-10-27 | 2001-04-27 | Method and configuration for forming classes for a language model based on linguistic classes |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2000025299A1 true WO2000025299A1 (de) | 2000-05-04 |
Family
ID=7885827
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/DE1999/003176 WO2000025299A1 (de) | 1998-10-27 | 1999-10-01 | Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen |
Country Status (4)
Country | Link |
---|---|
US (1) | US6640207B2 (de) |
EP (1) | EP1135767B1 (de) |
DE (1) | DE59901575D1 (de) |
WO (1) | WO2000025299A1 (de) |
Families Citing this family (37)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
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 |
WO2002029783A1 (en) * | 2000-09-30 | 2002-04-11 | Intel Corporation | Method and system for using rule-based knowledge to build a class-based domain specific statistical language model |
US7016829B2 (en) * | 2001-05-04 | 2006-03-21 | Microsoft Corporation | Method and apparatus for unsupervised training of natural language processing units |
US7024624B2 (en) * | 2002-01-07 | 2006-04-04 | Kenneth James Hintz | Lexicon-based new idea detector |
US7526424B2 (en) * | 2002-03-20 | 2009-04-28 | Microsoft Corporation | Sentence realization model for a natural language generation system |
JP3768205B2 (ja) * | 2003-05-30 | 2006-04-19 | 沖電気工業株式会社 | 形態素解析装置、形態素解析方法及び形態素解析プログラム |
US7379867B2 (en) * | 2003-06-03 | 2008-05-27 | Microsoft Corporation | Discriminative training of language models for text and speech classification |
US7318022B2 (en) * | 2003-06-12 | 2008-01-08 | Microsoft Corporation | Method and apparatus for training a translation disambiguation classifier |
US7478038B2 (en) * | 2004-03-31 | 2009-01-13 | Microsoft Corporation | Language model adaptation using semantic supervision |
CA2465155C (en) * | 2004-04-21 | 2008-12-09 | Ibm Canada Limited-Ibm Canada Limitee | Recommendations for intelligent data caching |
US7613666B1 (en) | 2004-04-23 | 2009-11-03 | Microsoft Corporation | Generating a class model from a business vocabulary to represent facts expressible in the business vocabulary |
US7761858B2 (en) * | 2004-04-23 | 2010-07-20 | Microsoft Corporation | Semantic programming language |
US7802231B2 (en) * | 2004-04-30 | 2010-09-21 | Microsoft Corporation | Generating programmatic interfaces from natural language expressions of authorizations for provision of information |
US7620935B2 (en) | 2004-04-30 | 2009-11-17 | Microsoft Corporation | Generating programmatic interfaces from natural language expressions of authorizations for request of information |
US7499850B1 (en) * | 2004-06-03 | 2009-03-03 | Microsoft Corporation | Generating a logical model of objects from a representation of linguistic concepts for use in software model generation |
US7613676B2 (en) | 2004-07-27 | 2009-11-03 | Microsoft Corporation | Generating a database model from natural language expressions of business rules |
US8050907B2 (en) * | 2004-07-30 | 2011-11-01 | Microsoft Corporation | Generating software components from business rules expressed in a natural language |
US7634406B2 (en) * | 2004-12-10 | 2009-12-15 | Microsoft Corporation | System and method for identifying semantic intent from acoustic information |
US20060142993A1 (en) * | 2004-12-28 | 2006-06-29 | Sony Corporation | System and method for utilizing distance measures to perform text classification |
US7945437B2 (en) * | 2005-02-03 | 2011-05-17 | Shopping.Com | Systems and methods for using automated translation and other statistical methods to convert a classifier in one language to another language |
US7996219B2 (en) * | 2005-03-21 | 2011-08-09 | At&T Intellectual Property Ii, L.P. | Apparatus and method for model adaptation for spoken language understanding |
US8700404B1 (en) | 2005-08-27 | 2014-04-15 | At&T Intellectual Property Ii, L.P. | System and method for using semantic and syntactic graphs for utterance classification |
EP2026327A4 (de) * | 2006-05-31 | 2012-03-07 | Nec Corp | Sprachenmodell-lernsystem, sprachenmodell-lernverfahren und sprachenmodell-lernprogramm |
US8069032B2 (en) * | 2006-07-27 | 2011-11-29 | Microsoft Corporation | Lightweight windowing method for screening harvested data for novelty |
US8856002B2 (en) * | 2007-04-12 | 2014-10-07 | International Business Machines Corporation | Distance metrics for universal pattern processing tasks |
US20090326924A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Projecting Semantic Information from a Language Independent Syntactic Model |
US20090326925A1 (en) * | 2008-06-27 | 2009-12-31 | Microsoft Corporation | Projecting syntactic information using a bottom-up pattern matching algorithm |
US8447608B1 (en) * | 2008-12-10 | 2013-05-21 | Adobe Systems Incorporated | Custom language models for audio content |
US8725509B1 (en) * | 2009-06-17 | 2014-05-13 | Google Inc. | Back-off language model compression |
US8812299B1 (en) * | 2010-06-24 | 2014-08-19 | Nuance Communications, Inc. | Class-based language model and use |
US9564122B2 (en) * | 2014-03-25 | 2017-02-07 | Nice Ltd. | Language model adaptation based on filtered data |
US9437189B2 (en) | 2014-05-29 | 2016-09-06 | Google Inc. | Generating language models |
US9529898B2 (en) | 2014-08-26 | 2016-12-27 | Google Inc. | Clustering classes in language modeling |
US10242260B1 (en) * | 2017-11-02 | 2019-03-26 | Sap Se | Using machine learning to flag gender biased words within free-form text, such as job descriptions |
US10467339B1 (en) * | 2018-06-28 | 2019-11-05 | Sap Se | Using machine learning and natural language processing to replace gender biased words within free-form text |
US11120367B2 (en) * | 2018-07-30 | 2021-09-14 | International Business Machines Corporation | Validating training data of classifiers |
US11410641B2 (en) * | 2018-11-28 | 2022-08-09 | Google Llc | Training and/or using a language selection model for automatically determining language for speech recognition of spoken utterance |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US3812291A (en) | 1972-06-19 | 1974-05-21 | Scope Inc | Signal pattern encoder and classifier |
US5680509A (en) * | 1994-09-27 | 1997-10-21 | International Business Machines Corporation | Method and apparatus for estimating phone class probabilities a-posteriori using a decision tree |
US5835893A (en) * | 1996-02-15 | 1998-11-10 | Atr Interpreting Telecommunications Research Labs | Class-based word clustering for speech recognition using a three-level balanced hierarchical similarity |
US5839106A (en) * | 1996-12-17 | 1998-11-17 | Apple Computer, Inc. | Large-vocabulary speech recognition using an integrated syntactic and semantic statistical language model |
-
1999
- 1999-10-01 EP EP99957254A patent/EP1135767B1/de not_active Expired - Lifetime
- 1999-10-01 DE DE59901575T patent/DE59901575D1/de not_active Expired - Lifetime
- 1999-10-01 WO PCT/DE1999/003176 patent/WO2000025299A1/de active IP Right Grant
-
2001
- 2001-04-27 US US09/844,931 patent/US6640207B2/en not_active Expired - Fee Related
Non-Patent Citations (3)
Title |
---|
MASATAKI H ET AL: "Variable-order N-gram generation by word-class splitting and consecutive word grouping", 1996 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING CONFERENCE PROCEEDINGS (CAT. NO.96CH35903), 1996 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING CONFERENCE PROCEEDINGS, ATLANTA, GA, USA, 7-10 M, 1996, New York, NY, USA, IEEE, USA, pages 188 - 191 vol. 1, XP002133222, ISBN: 0-7803-3192-3 * |
NEY H ET AL: "ON STRUCTURING PROBABILISTIC DEPENDENCES IN STOCHASTIC LANGUAGE MODELLING", COMPUTER SPEECH AND LANGUAGE,GB,ACADEMIC PRESS, LONDON, vol. 8, no. 1, 1 January 1994 (1994-01-01), pages 1 - 38, XP000452242, ISSN: 0885-2308 * |
NIESLER T R ET AL: "COMPARISON OF PART-OF-SPEECH AND AUTOMATICALLY DERIVED CATEGORY- BASED LANGUAGE MODELS FOR SPEECH RECOGNITION", IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING,US,NEW YORK, NY: IEEE, vol. CONF. 23, 1998, pages 177 - 180, XP000854544, ISBN: 0-7803-4429-4 * |
Also Published As
Publication number | Publication date |
---|---|
EP1135767B1 (de) | 2002-05-29 |
US6640207B2 (en) | 2003-10-28 |
US20010051868A1 (en) | 2001-12-13 |
DE59901575D1 (de) | 2002-07-04 |
EP1135767A1 (de) | 2001-09-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2000025299A1 (de) | Verfahren und anordnung zur klassenbildung für ein sprachmodell basierend auf linguistischen klassen | |
DE69822296T2 (de) | Mustererkennungsregistrierung in einem verteilten system | |
DE602004012909T2 (de) | Verfahren und Vorrichtung zur Modellierung eines Spracherkennungssystems und zur Schätzung einer Wort-Fehlerrate basierend auf einem Text | |
DE69829389T2 (de) | Textnormalisierung unter verwendung einer kontextfreien grammatik | |
DE69814104T2 (de) | Aufteilung von texten und identifizierung von themen | |
DE69908047T2 (de) | Verfahren und System zur automatischen Bestimmung von phonetischen Transkriptionen in Verbindung mit buchstabierten Wörtern | |
DE112014006542B4 (de) | Einrichtung und Verfahren zum Verständnis von einer Benutzerintention | |
DE60033733T2 (de) | Datenbankabfragesystem basierend auf Spracherkennung | |
DE69818231T2 (de) | Verfahren zum diskriminativen training von spracherkennungsmodellen | |
DE60004862T2 (de) | Automatische bestimmung der genauigkeit eines aussprachewörterbuchs in einem spracherkennungssystem | |
DE60111329T2 (de) | Anpassung des phonetischen Kontextes zur Verbesserung der Spracherkennung | |
DE112017004397T5 (de) | System und Verfahren zur Einstufung von hybriden Spracherkennungsergebnissen mit neuronalen Netzwerken | |
DE102008017993B4 (de) | Sprachsuchvorrichtung | |
DE112009002000B4 (de) | Adaptives Verfahren und Vorrichtung zur Umwandlung von Nachrichten zwischen unterschiedlichen Datenformaten | |
DE60316912T2 (de) | Verfahren zur Spracherkennung | |
DE60313706T2 (de) | Spracherkennungs- und -antwortsystem, Spracherkennungs- und -antwortprogramm und zugehöriges Aufzeichnungsmedium | |
EP1273003B1 (de) | Verfahren und vorrichtung zum bestimmen prosodischer markierungen | |
DE102019004300A1 (de) | Verwendung eines dynamischen speichernetzwerks zum verfolgen digitaler dialogzustände und erzeugen von antworten | |
DE112010006037T5 (de) | Spracherkennungsvorrichtung und Navigationssystem | |
DE112013006770B4 (de) | Spracherkennungsvorrichtung und Spracherkennungsverfahren | |
DE60133537T2 (de) | Automatisches umtrainieren eines spracherkennungssystems | |
DE112006000322T5 (de) | Audioerkennungssystem zur Erzeugung von Antwort-Audio unter Verwendung extrahierter Audiodaten | |
EP0987682A2 (de) | Verfahren zur Adaption von linguistischen Sprachmodellen | |
DE602004003609T2 (de) | Lösung der Segmentierungsmehrdeutigkeit bei der Grammatikerstellung | |
EP1016077B1 (de) | Verfahren zur bestimmung einer wahrscheinlichkeit für das auftreten einer folge von mindestens zwei wörtern bei einer spracherkennung |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AK | Designated states |
Kind code of ref document: A1 Designated state(s): US |
|
AL | Designated countries for regional patents |
Kind code of ref document: A1 Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE |
|
DFPE | Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101) | ||
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: 1999957254 Country of ref document: EP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 09844931 Country of ref document: US |
|
WWP | Wipo information: published in national office |
Ref document number: 1999957254 Country of ref document: EP |
|
WWG | Wipo information: grant in national office |
Ref document number: 1999957254 Country of ref document: EP |