CA2275774C - Selection of superwords based on criteria relevant to both speech recognition and understanding - Google Patents
Selection of superwords based on criteria relevant to both speech recognition and understanding Download PDFInfo
- Publication number
- CA2275774C CA2275774C CA002275774A CA2275774A CA2275774C CA 2275774 C CA2275774 C CA 2275774C CA 002275774 A CA002275774 A CA 002275774A CA 2275774 A CA2275774 A CA 2275774A CA 2275774 C CA2275774 C CA 2275774C
- Authority
- CA
- Canada
- Prior art keywords
- database
- meaningful
- superwords
- phrases
- superword
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Lifetime
Links
- 238000000034 method Methods 0.000 claims description 21
- 238000005259 measurement Methods 0.000 claims description 13
- 238000012549 training Methods 0.000 claims description 11
- 230000007423 decrease Effects 0.000 claims description 7
- 230000003247 decreasing effect Effects 0.000 claims description 6
- 239000012634 fragment Substances 0.000 description 5
- 238000012360 testing method Methods 0.000 description 4
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 210000003050 axon Anatomy 0.000 description 1
- 101150087654 chrnd gene Proteins 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000012790 confirmation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000010348 incorporation Methods 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008450 motivation Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000026683 transduction Effects 0.000 description 1
- 238000010361 transduction Methods 0.000 description 1
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
-
- 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
Abstract
This invention is directed to the selection of superwords based on a criteri on relevant to speech recognition and understanding. Superwords are used to ref er to those word combinations which are so often spoken that they are recognize d as units or should have models to reflect them in the language model. The selected superwords are placed in a lexicon along with selected meaningful phrases. The lexicon is then used by a speech recognizer to improve recognition of input speech utterances for the proper routing of a user's ta sk objectives.
Description
SELECTION OF SUPERTnIORDS BASED ON CRITERIA RELEVANT
TO BOTH SPEECH RECOGNITION AND UNDERSTANDING
BACKGROUND OF THE INVENTION
1. Field of Invention This inveni~ion relates to the selection of superwords and meaningful phrases based on a criterion relevant to both speech recognition acrd understand.
TO BOTH SPEECH RECOGNITION AND UNDERSTANDING
BACKGROUND OF THE INVENTION
1. Field of Invention This inveni~ion relates to the selection of superwords and meaningful phrases based on a criterion relevant to both speech recognition acrd understand.
2. Description of Related Art Currently, there are applications of speech recognition that provide a methodology for automated task selection where t:he targeted ta:>k is recognized in the natural speech of a user making such a selection. A
fundamental aspect of this method is a determination of a set of meaningful phrases. :>uch meaningful phrases are determined by a grammatical inference algorithm which operates orr a predeterrn;~rn ed corpus of speech utterances, each such utterance being associated with a specific task objecti~re, and wherein each utterance is marked with its associ~~ted task objective.
The determination of the meaningful phrases used in the above application is founded in the concept of combining a measure of commonality of words and/or structure within the language -- I.e., how often groupings of things co-occur. -- with a measure of significance to a defined task for such a grouping. That commonality measure within the language can be manifested as the mutual information in n-grams derived from a database of training speech utterances and the measure of usefulness to a task is manifested as a salience measure.
Mutual information ("MI"), which measures the likelihood of co-occurrence for two or more words, involves only the language itself. For example, given War and Peace in the original Russian, one could compute the mutual information for all the possible pairings of words in that I0 text without ever understanding a word of the language in which it is written. In contrast, computing salience involves both the language and its extra-linguistic associations to a device's environment. Through the use of such a combination of MI and a salience factor, meaningful 15 phrases are selected which have both a positive MI
(indicating relative strong association among the words comprising the phrase) and a high salience value.
Such methods are based upon the probability that separate sets of salient words occur in the particular input 20 utterance. For example, the salient phrases "made a long distance" would be determined as a meaningful phrases by that grammatical inference algorithm based on their individual mutual information and salience values.
In addition, while the task goal involves 25 recognizing meaningful words and phrases, this is typically accomplished via a large vocabulary recognizer, constrained by stochastic language models, such as an n-gram model. One approach to such modeling to constrain the recognizer is to train a stochastic finite state grammar represented by a 30 Variable Ngram Stochastic Automaton (VNSA). A VNSA is a non-deterministic automaton that allows for parsing any possible sequence of words drawn from a given vocabulary.
Traditionally, such n-gram language models for speech recognition assume words as the basic lexical unit.
The order of a VNSA network is t1:= maximum number of words that can be predicted as occurring a_ter the occurrence of a particular word in an utterance. :'hus, using conditional probabilities, VNSAs have been uses to approximate standard n-gram language models yielding similar performance to standard bigram and trigram models. aowever, when the "n" in the n-gram becomes large, a database for predicting the occurrence of words in response to _he appearance of a word in an utterance, becomes large and unmanageable. In addition, the occurrence of words which are not strongly recurrent in the language may be assigned mistakenly high probabilities, and thus generating _ number of misdetections in recognized speech.
Thus, a method to create _~zger units for language modeling is needed in order to prc:.-.ete the efficient use of n-gram language models for speech =cognition and for using these longer units along with mear._ngful words and phrases for language recognition and understanding.
SUMMARY OF THE I':':~NTION
A method and apparatus ~~= using superwords and meaningful phrases for both =peech recognition and understanding is provided. For superwords, a training corpus is initialized with the _::itial language model corresponding to a stochastic n-gra:~ model and words. Rank candidate symbol pairs are generated based on a correlation coefficient where the probability o-_' a word occurring denotes the probability of further words occurring sequentially. The resulting superword is selected on_y if the probabilities occur sequentially and the traininc set perplexity decreases by the incorporation of the lager lexical unit (or superword) into the model.
In addition, meaningful ~:ords and phrases are selected which have both a mutua_ information value that indicates a relatively strong asstoiation among the words comprising the phrase, and a high salience value. The selected meaningfu_L words and phrases are input along with the superwords ir..to a common lexicon which provides input to a speech recognizes and interpretation module.
The lexicon consists of all of tr:e meaningful words and phrases, existing superwords, and newly generated superwords. The language model is then retrained using this lexicon.
In accordance with one aspect of the present invention there is provided a method for determining superwords and meaningful phrases, comprising the steps of: determining superwords from a set of candidate phrases contained in a database of sequences of words, symbols and/or sounds based on commonality measurements;
determining meaningful phrases based on commonality measurements; and inserting the meaningful phrases into the database.
In accordan<:e with another aspect of the present invention there is provided an apparatus that determines superwords and meaningful phrases, comprising: a database of commonly occurring words, symbols and/or sounds, the database having a perplexity value;
superword selecting means for selecting superwords based on commonality measurement's, the selected superwords being input to the database; and meaningful phrase selecting means for selecting meaningful phrases based on commonality measurements, the selected meaningful phrases being input to the database.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention :..s dc~sc~rbed in detail with reference to the following drawings wherein like numerals reference like elements, and wherein:
4a Fig. 1A and 1B ar:-a f:Lowcharts for generating and selecting superwords and meaningful phrases;
Fig. 2 is a chart sl~owirig the false rejection rate for entropy-based and saL:fence-based phrases;
Fig. 3 is a block ~iag~_~am of a speech recognizer; and Fig. 4 is a functional block diagram of a system in which the method for sel.ec.t.ing superwords may be employed.
DETAILED DESCRIPTION OF PREFERRED EMBC)DIMENT'i N-gram language models for speech recognition are currently implemented by us:ing words as the basic lexical unit. However, there axon several motivations for choosing longer Lznits for language modeling. First, not all languages have a predefined word unit (e. g.
Chinese). Second, many word groups or phrases are strongly recurrent in thf.~ Languages and can be through as a single lexical entry, e.g. 'area code' , '.I would Like to' or 'New Jersey'. Third, for any model of a fixed order, we can selectively enhance the conditional probabilities by using variable l.engt~h units to capture long spanning dependencies. The use of these longer lexical units is t::he basis f~:~r the generation of superwords.
The method for automa=_cally generating and selecting such variable length un'_=~, or superwords is based on minimization of the language perplexity PP(T) on a training corpus T.
Perplexity can be descrit=as follows. Natural language can be viewed as an i-'~r:nation source W whose output is a word sequence. Thus, -.:~ can associate W to the entropy H(W) as:
H(W)= -log P(wi,wZ,...,wn) (1) n where wl,w2,..., w~ is the actual text corpus ) .
Perplexity (PP) is used .., speech recognition to measure the difficulty of a recoc-_tion task relative to a given language model. In partic-~_ar, perplexity measures the "Average Branching Factor" (i.e_., the average number of words, symbols or sounds that ca.-. follow any given word, symbol or sound). The larger the ~~.::nber of words that could possibly follow one or more words, symbols or sounds, the higher the perplexity value. Fo= example, there are many more words that could follow she word "to" (higher perplexity value) than can follow _ze phrase "I would like to" (lower perplexity value).
The Perplexity of a lang~.:age model is defined as follows:
PP=2Lp where LP is the logprob:
LP=- n log P (wI, w2,..., wn) ( 2 ) where n is the size of the cc=bus W=wl,wz,...,w~ used to estimate PP and P ( wi,w2,..., w~) is t::--__ probability estimate of the corpus W. While there has ~~en other research into automatically acquiring entropy-r~~~~cing phrases, this work differs significantly in the language model components and optimization parameters.
On the other hand, the determination of meaningful phrases used is founded in the concept of combining a measure of commonality of words and/or structure within the language -- i.e., how often groupings of things co-occur -- with a measure of significance to a defined task for such a grouping. That commonality measure within the language can be manifested as the mutual information in n-grams derived from a database of training speech utterances and the measure of usefulness to a task is manifested as a salience measure.
As shown in Figures 1A and 1B, the phrase acquisition method is an iterative process which generates a set of Entropy-based superwords (Ephr) and a set of Salience-based meaningful words and phrases (Sphr). In particular, given a fixed model order n and a training corpus T, the method proceeds as follows.
At step S100, the initial parameters are set so that K equals the number of superword candidates generated at each iteration, and M equals the number of iterations, TIi is set as the initial training corpus T, and x,11 is set as the language model of order n corresponding to a stochastic n-gram model on words trained from that corpus. At step S105, a counter m is advanced from 1 to M, M being the number of iterations set in the parameters at step 5100.
At step 5110, for each iteration, candidate symbol-pairs (x, y) are ranked and generated based on a correlation coefficient P (x.YJ - P (x.YJ l l P (xJ + P (yJ 7 ( 3 ) where P(xJ denotes the probability of the event x and P(x,yJ
denotes the probability of the symbols x and y occurring sequentially. At the first iteration, x and y are both WO 99/22363 ~ PCT/US98/22399 words, in subsequent iterations they are potentially larger units. Observe that 0 __<p(x,y) <_0.5. This correlation measure provides advantages with respect to ease of scaling and thresholding.
At step 5115, a counter is advanced for each candidate phrase so that k - 1 to K, where K is the set of superword candidates. At step S120, the current training corpus Tm,k-i is filtered by replacing each occurrence of a superword with the superword unit (x y)k. The new filtered superword set is denoted by Tmx.
At step S125, a new language model (still of order n) is trained from Tmk. The newly trained model is denoted by ~,mk. At step 5130, a test is performed to determine whether adding the superword candidate decreases perplexity ( i . a . , whether PP (.Z.mk , Tmx) < PP ( ~,m,x-i , Tm,x-i) ) . I f perplexity is decreased, the process is continued. If perplexity is not decreased, the superword candidate is rejected at step S135 by setting Tmk = Tm,x-i~ Thus, a phrase x y is selected only i f P (x, y) - P (x) . P (y) ( i . a . , P (y ~ x) _ I) I and the training set perplexity is decreased by incorporating this larger lexical unit into the model.
At step 5140, the method tests whether any additional superword candidates, k, are required to be examined. At step S145, after all of the superword candidates are examined, the next iteration is performed, m, until m = M. The resulting set of superwords generated is designated as Ephr.
Next, at step S150, a set of salience-based phrases Sphr are selected as candidate phrases. At step 5155, the training corpus T1 is initialized and set as TMx. At step S160, a counter m is advanced from 1 to M, M being the number of iterations to be performed for each candidate phrase in Sphr.
At step 5165, the current training corpus Tm is filtered by replacing each occurrence of the phrase with the phrase unit (x y)m. This new filtered set is denoted by Tm+i ~ At step 5170, the next candidate phrase is retrieved until m = M. The resulting set of salience-based meaningful phrases is designated Sphr.
Finally, at step S175, a final language model from the filtered corpus TM plus the original T is trained with the lexicon comprising all original words plus the phrases contained in the sets Ephr and Sphr. This preserves the granularity of the original lexicon, generating alternate paths comprising both the new phrases plus their original word sequences. I.e., if the words "long" and "distance"
only occur together in the corpus leading to the acquisition of the phrase "long distance", this final step preserves the possibility of the recognized words occurring separately in some test utterance.
A decision is then made as to how to classify an utterance in a particularly straightforward manner. A
speech recognizes is applied to an utterance, producing a single best word recognition output. This ASR output is then searched for occurrences of the salient phrase fragments. In case of fragment overlap, some parsing is required. The method for parsing is a simple one in that longer fragments are selected over shorter ones proceeding left to right in the utterance. This yields a transduction from the utterance s to a sequence of associated call-types.
To each of these fragments fi is associated the peak value and location of the a posteriors distribution.
Pi = maxk P (Ck~ f~~ ( 4 ) ki = arg maxk P (Ck[fij ( 5 ) Thus, for each utterance s we have a sequence (fi ki, pij .
The decision rule is to select the call-type of the fragment with maximum pi, i . a . select CKrs~ where i (s) = arg maxi pi ( 6 ) K (sJ = ki rs~ ( ~ ) If this overall peak is less than some threshold, PT, then the utterance is rej ected and classified as other, i . a . if Pi rs! < Pz .
Figure 2 shows the difference in the "False Rejection Rate" (the number of transactions misdirected or misunderstood by a machine based on inputs from a user) for speech recognizers using a lexicon with generated superwords versus a lexicon using generated superwords and meaningful phrases. The lexicon containing the generated superwords plus meaningful phrases produces fewer false rejections than the lexicon with only generated superwords and therefore leads to better understanding of a user's language patterns.
Figure 3 shows a configuration for a Speech Recognizer 315 receiving both superwords and meaningful phrases. Both the Meaningful Phrase Selector 300 and the Superword Selector 305, are trained using test speech utterances and select meaningful phrases and superwords in accordance with the methods set forth above. Once the meaningful phrases and superwords have been selected, they are input to a Lexicon 310. The Lexicon 310 is essentially a standard database and can be stored in any known internal or external memory device. The Lexicon 310 serves as a source of meaningful phrases and superwords which the Speech Recognizer 315 looks for in input speech. The Speech Recognizer 315 can be any standard speech recognizer known in the art.
WO 99/22363 1 p PCT/US98/22399 Figure 4 shows a structure in which the resulting lexicon of meaningful phrases and superwords may be implemented. As can be seen from the drawing, the structure comprises two related subsystems: The Superword and Meaningful Phrase Generation Subsystem 400 and Input Speech Classification Subsystem 410.
The Superword and Meaningful Phrase Generation Subsystem 400 operates on a database or lexicon of a large number of utterances each of which is related to one of a predetermined set of routing objectives, where each such utterance is labeled with its associated routing objective.
The operation of this subsystem is essentially carried out by Meaningful Phrase Selector 300 and Superword Selector 305 which select as outputs a set of meaningful phrases and superwords having a probabilistic relationship with one or more of the set of predetermined routing objectives with which the input speech utterances are associated. The selected meaningful phrases and superwords are then input into the Lexicon 310 which stores the meaningful phrases and superwords for use by the Input Speech Classification Subsystem 4I0. The operation of Meaningful Phrase Selector 300 and the Superword Selector 305 are generally determined in accordance with any known algorithm for selecting meaningful phrases and superwords including the methods disclosed herein.
Operation of Input Speech Classification Subsystem 410 begins with inputting a user's task objective request, in the caller's natural speech, to Input Speech Recognizer 315.
The Input Speech Recognizer 315 may be of any known design and performs the function of recognizing, or spotting, the existence of one or more meaningful phrase in the input speech. As can be seen in the figure, the meaningful phrases and superwords developed by the Superword and Meaningful Phrase Generation Subsystem 400 are provided as an input to the Input Speech Recognizer 315 to define the routing objectives related to meaningful phrases and superwords and to establish the levels of probability for a relation of such input meaningful phrases and superwords to a particular routing objective.
The output of the Input Speech Recognizer 315, which includes the detected meaningful phrases and superwords appearing in the caller's routing objective request, is provided to Interpretation Module 420. The Interpretation Module 420 compiles a list of probable routing objectives based on the probabilistic relation between the recognized meaningful phrases and superwords and selected routing objectives. The Interpretation Module 420 then inputs a list of possible routing objectives to the Dialog Manager 430. In the Dialog Manager 430, a decision is made either to implement the chosen routing objective with an announcement to the caller that such an objective is being implemented (if only one objective is possible), or to seek additional information and/or confirmation from the caller (if none or more than one objective are possible). Dialog continues until either a clear decision can be made to implement a particular routing objective or a determination is made that no such decision is likely, in which case the caller is defaulted to an operator position.
While this invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations would be apparent to those skilled in the art. Accordingly, preferred embodiments in the invention as set forth herein are intended to be illustrative, not limiting. Various changes may be made without departing from the spirit and scope of the invention as defined in the following claims.
fundamental aspect of this method is a determination of a set of meaningful phrases. :>uch meaningful phrases are determined by a grammatical inference algorithm which operates orr a predeterrn;~rn ed corpus of speech utterances, each such utterance being associated with a specific task objecti~re, and wherein each utterance is marked with its associ~~ted task objective.
The determination of the meaningful phrases used in the above application is founded in the concept of combining a measure of commonality of words and/or structure within the language -- I.e., how often groupings of things co-occur. -- with a measure of significance to a defined task for such a grouping. That commonality measure within the language can be manifested as the mutual information in n-grams derived from a database of training speech utterances and the measure of usefulness to a task is manifested as a salience measure.
Mutual information ("MI"), which measures the likelihood of co-occurrence for two or more words, involves only the language itself. For example, given War and Peace in the original Russian, one could compute the mutual information for all the possible pairings of words in that I0 text without ever understanding a word of the language in which it is written. In contrast, computing salience involves both the language and its extra-linguistic associations to a device's environment. Through the use of such a combination of MI and a salience factor, meaningful 15 phrases are selected which have both a positive MI
(indicating relative strong association among the words comprising the phrase) and a high salience value.
Such methods are based upon the probability that separate sets of salient words occur in the particular input 20 utterance. For example, the salient phrases "made a long distance" would be determined as a meaningful phrases by that grammatical inference algorithm based on their individual mutual information and salience values.
In addition, while the task goal involves 25 recognizing meaningful words and phrases, this is typically accomplished via a large vocabulary recognizer, constrained by stochastic language models, such as an n-gram model. One approach to such modeling to constrain the recognizer is to train a stochastic finite state grammar represented by a 30 Variable Ngram Stochastic Automaton (VNSA). A VNSA is a non-deterministic automaton that allows for parsing any possible sequence of words drawn from a given vocabulary.
Traditionally, such n-gram language models for speech recognition assume words as the basic lexical unit.
The order of a VNSA network is t1:= maximum number of words that can be predicted as occurring a_ter the occurrence of a particular word in an utterance. :'hus, using conditional probabilities, VNSAs have been uses to approximate standard n-gram language models yielding similar performance to standard bigram and trigram models. aowever, when the "n" in the n-gram becomes large, a database for predicting the occurrence of words in response to _he appearance of a word in an utterance, becomes large and unmanageable. In addition, the occurrence of words which are not strongly recurrent in the language may be assigned mistakenly high probabilities, and thus generating _ number of misdetections in recognized speech.
Thus, a method to create _~zger units for language modeling is needed in order to prc:.-.ete the efficient use of n-gram language models for speech =cognition and for using these longer units along with mear._ngful words and phrases for language recognition and understanding.
SUMMARY OF THE I':':~NTION
A method and apparatus ~~= using superwords and meaningful phrases for both =peech recognition and understanding is provided. For superwords, a training corpus is initialized with the _::itial language model corresponding to a stochastic n-gra:~ model and words. Rank candidate symbol pairs are generated based on a correlation coefficient where the probability o-_' a word occurring denotes the probability of further words occurring sequentially. The resulting superword is selected on_y if the probabilities occur sequentially and the traininc set perplexity decreases by the incorporation of the lager lexical unit (or superword) into the model.
In addition, meaningful ~:ords and phrases are selected which have both a mutua_ information value that indicates a relatively strong asstoiation among the words comprising the phrase, and a high salience value. The selected meaningfu_L words and phrases are input along with the superwords ir..to a common lexicon which provides input to a speech recognizes and interpretation module.
The lexicon consists of all of tr:e meaningful words and phrases, existing superwords, and newly generated superwords. The language model is then retrained using this lexicon.
In accordance with one aspect of the present invention there is provided a method for determining superwords and meaningful phrases, comprising the steps of: determining superwords from a set of candidate phrases contained in a database of sequences of words, symbols and/or sounds based on commonality measurements;
determining meaningful phrases based on commonality measurements; and inserting the meaningful phrases into the database.
In accordan<:e with another aspect of the present invention there is provided an apparatus that determines superwords and meaningful phrases, comprising: a database of commonly occurring words, symbols and/or sounds, the database having a perplexity value;
superword selecting means for selecting superwords based on commonality measurement's, the selected superwords being input to the database; and meaningful phrase selecting means for selecting meaningful phrases based on commonality measurements, the selected meaningful phrases being input to the database.
BRIEF DESCRIPTION OF THE DRAWINGS
The invention :..s dc~sc~rbed in detail with reference to the following drawings wherein like numerals reference like elements, and wherein:
4a Fig. 1A and 1B ar:-a f:Lowcharts for generating and selecting superwords and meaningful phrases;
Fig. 2 is a chart sl~owirig the false rejection rate for entropy-based and saL:fence-based phrases;
Fig. 3 is a block ~iag~_~am of a speech recognizer; and Fig. 4 is a functional block diagram of a system in which the method for sel.ec.t.ing superwords may be employed.
DETAILED DESCRIPTION OF PREFERRED EMBC)DIMENT'i N-gram language models for speech recognition are currently implemented by us:ing words as the basic lexical unit. However, there axon several motivations for choosing longer Lznits for language modeling. First, not all languages have a predefined word unit (e. g.
Chinese). Second, many word groups or phrases are strongly recurrent in thf.~ Languages and can be through as a single lexical entry, e.g. 'area code' , '.I would Like to' or 'New Jersey'. Third, for any model of a fixed order, we can selectively enhance the conditional probabilities by using variable l.engt~h units to capture long spanning dependencies. The use of these longer lexical units is t::he basis f~:~r the generation of superwords.
The method for automa=_cally generating and selecting such variable length un'_=~, or superwords is based on minimization of the language perplexity PP(T) on a training corpus T.
Perplexity can be descrit=as follows. Natural language can be viewed as an i-'~r:nation source W whose output is a word sequence. Thus, -.:~ can associate W to the entropy H(W) as:
H(W)= -log P(wi,wZ,...,wn) (1) n where wl,w2,..., w~ is the actual text corpus ) .
Perplexity (PP) is used .., speech recognition to measure the difficulty of a recoc-_tion task relative to a given language model. In partic-~_ar, perplexity measures the "Average Branching Factor" (i.e_., the average number of words, symbols or sounds that ca.-. follow any given word, symbol or sound). The larger the ~~.::nber of words that could possibly follow one or more words, symbols or sounds, the higher the perplexity value. Fo= example, there are many more words that could follow she word "to" (higher perplexity value) than can follow _ze phrase "I would like to" (lower perplexity value).
The Perplexity of a lang~.:age model is defined as follows:
PP=2Lp where LP is the logprob:
LP=- n log P (wI, w2,..., wn) ( 2 ) where n is the size of the cc=bus W=wl,wz,...,w~ used to estimate PP and P ( wi,w2,..., w~) is t::--__ probability estimate of the corpus W. While there has ~~en other research into automatically acquiring entropy-r~~~~cing phrases, this work differs significantly in the language model components and optimization parameters.
On the other hand, the determination of meaningful phrases used is founded in the concept of combining a measure of commonality of words and/or structure within the language -- i.e., how often groupings of things co-occur -- with a measure of significance to a defined task for such a grouping. That commonality measure within the language can be manifested as the mutual information in n-grams derived from a database of training speech utterances and the measure of usefulness to a task is manifested as a salience measure.
As shown in Figures 1A and 1B, the phrase acquisition method is an iterative process which generates a set of Entropy-based superwords (Ephr) and a set of Salience-based meaningful words and phrases (Sphr). In particular, given a fixed model order n and a training corpus T, the method proceeds as follows.
At step S100, the initial parameters are set so that K equals the number of superword candidates generated at each iteration, and M equals the number of iterations, TIi is set as the initial training corpus T, and x,11 is set as the language model of order n corresponding to a stochastic n-gram model on words trained from that corpus. At step S105, a counter m is advanced from 1 to M, M being the number of iterations set in the parameters at step 5100.
At step 5110, for each iteration, candidate symbol-pairs (x, y) are ranked and generated based on a correlation coefficient P (x.YJ - P (x.YJ l l P (xJ + P (yJ 7 ( 3 ) where P(xJ denotes the probability of the event x and P(x,yJ
denotes the probability of the symbols x and y occurring sequentially. At the first iteration, x and y are both WO 99/22363 ~ PCT/US98/22399 words, in subsequent iterations they are potentially larger units. Observe that 0 __<p(x,y) <_0.5. This correlation measure provides advantages with respect to ease of scaling and thresholding.
At step 5115, a counter is advanced for each candidate phrase so that k - 1 to K, where K is the set of superword candidates. At step S120, the current training corpus Tm,k-i is filtered by replacing each occurrence of a superword with the superword unit (x y)k. The new filtered superword set is denoted by Tmx.
At step S125, a new language model (still of order n) is trained from Tmk. The newly trained model is denoted by ~,mk. At step 5130, a test is performed to determine whether adding the superword candidate decreases perplexity ( i . a . , whether PP (.Z.mk , Tmx) < PP ( ~,m,x-i , Tm,x-i) ) . I f perplexity is decreased, the process is continued. If perplexity is not decreased, the superword candidate is rejected at step S135 by setting Tmk = Tm,x-i~ Thus, a phrase x y is selected only i f P (x, y) - P (x) . P (y) ( i . a . , P (y ~ x) _ I) I and the training set perplexity is decreased by incorporating this larger lexical unit into the model.
At step 5140, the method tests whether any additional superword candidates, k, are required to be examined. At step S145, after all of the superword candidates are examined, the next iteration is performed, m, until m = M. The resulting set of superwords generated is designated as Ephr.
Next, at step S150, a set of salience-based phrases Sphr are selected as candidate phrases. At step 5155, the training corpus T1 is initialized and set as TMx. At step S160, a counter m is advanced from 1 to M, M being the number of iterations to be performed for each candidate phrase in Sphr.
At step 5165, the current training corpus Tm is filtered by replacing each occurrence of the phrase with the phrase unit (x y)m. This new filtered set is denoted by Tm+i ~ At step 5170, the next candidate phrase is retrieved until m = M. The resulting set of salience-based meaningful phrases is designated Sphr.
Finally, at step S175, a final language model from the filtered corpus TM plus the original T is trained with the lexicon comprising all original words plus the phrases contained in the sets Ephr and Sphr. This preserves the granularity of the original lexicon, generating alternate paths comprising both the new phrases plus their original word sequences. I.e., if the words "long" and "distance"
only occur together in the corpus leading to the acquisition of the phrase "long distance", this final step preserves the possibility of the recognized words occurring separately in some test utterance.
A decision is then made as to how to classify an utterance in a particularly straightforward manner. A
speech recognizes is applied to an utterance, producing a single best word recognition output. This ASR output is then searched for occurrences of the salient phrase fragments. In case of fragment overlap, some parsing is required. The method for parsing is a simple one in that longer fragments are selected over shorter ones proceeding left to right in the utterance. This yields a transduction from the utterance s to a sequence of associated call-types.
To each of these fragments fi is associated the peak value and location of the a posteriors distribution.
Pi = maxk P (Ck~ f~~ ( 4 ) ki = arg maxk P (Ck[fij ( 5 ) Thus, for each utterance s we have a sequence (fi ki, pij .
The decision rule is to select the call-type of the fragment with maximum pi, i . a . select CKrs~ where i (s) = arg maxi pi ( 6 ) K (sJ = ki rs~ ( ~ ) If this overall peak is less than some threshold, PT, then the utterance is rej ected and classified as other, i . a . if Pi rs! < Pz .
Figure 2 shows the difference in the "False Rejection Rate" (the number of transactions misdirected or misunderstood by a machine based on inputs from a user) for speech recognizers using a lexicon with generated superwords versus a lexicon using generated superwords and meaningful phrases. The lexicon containing the generated superwords plus meaningful phrases produces fewer false rejections than the lexicon with only generated superwords and therefore leads to better understanding of a user's language patterns.
Figure 3 shows a configuration for a Speech Recognizer 315 receiving both superwords and meaningful phrases. Both the Meaningful Phrase Selector 300 and the Superword Selector 305, are trained using test speech utterances and select meaningful phrases and superwords in accordance with the methods set forth above. Once the meaningful phrases and superwords have been selected, they are input to a Lexicon 310. The Lexicon 310 is essentially a standard database and can be stored in any known internal or external memory device. The Lexicon 310 serves as a source of meaningful phrases and superwords which the Speech Recognizer 315 looks for in input speech. The Speech Recognizer 315 can be any standard speech recognizer known in the art.
WO 99/22363 1 p PCT/US98/22399 Figure 4 shows a structure in which the resulting lexicon of meaningful phrases and superwords may be implemented. As can be seen from the drawing, the structure comprises two related subsystems: The Superword and Meaningful Phrase Generation Subsystem 400 and Input Speech Classification Subsystem 410.
The Superword and Meaningful Phrase Generation Subsystem 400 operates on a database or lexicon of a large number of utterances each of which is related to one of a predetermined set of routing objectives, where each such utterance is labeled with its associated routing objective.
The operation of this subsystem is essentially carried out by Meaningful Phrase Selector 300 and Superword Selector 305 which select as outputs a set of meaningful phrases and superwords having a probabilistic relationship with one or more of the set of predetermined routing objectives with which the input speech utterances are associated. The selected meaningful phrases and superwords are then input into the Lexicon 310 which stores the meaningful phrases and superwords for use by the Input Speech Classification Subsystem 4I0. The operation of Meaningful Phrase Selector 300 and the Superword Selector 305 are generally determined in accordance with any known algorithm for selecting meaningful phrases and superwords including the methods disclosed herein.
Operation of Input Speech Classification Subsystem 410 begins with inputting a user's task objective request, in the caller's natural speech, to Input Speech Recognizer 315.
The Input Speech Recognizer 315 may be of any known design and performs the function of recognizing, or spotting, the existence of one or more meaningful phrase in the input speech. As can be seen in the figure, the meaningful phrases and superwords developed by the Superword and Meaningful Phrase Generation Subsystem 400 are provided as an input to the Input Speech Recognizer 315 to define the routing objectives related to meaningful phrases and superwords and to establish the levels of probability for a relation of such input meaningful phrases and superwords to a particular routing objective.
The output of the Input Speech Recognizer 315, which includes the detected meaningful phrases and superwords appearing in the caller's routing objective request, is provided to Interpretation Module 420. The Interpretation Module 420 compiles a list of probable routing objectives based on the probabilistic relation between the recognized meaningful phrases and superwords and selected routing objectives. The Interpretation Module 420 then inputs a list of possible routing objectives to the Dialog Manager 430. In the Dialog Manager 430, a decision is made either to implement the chosen routing objective with an announcement to the caller that such an objective is being implemented (if only one objective is possible), or to seek additional information and/or confirmation from the caller (if none or more than one objective are possible). Dialog continues until either a clear decision can be made to implement a particular routing objective or a determination is made that no such decision is likely, in which case the caller is defaulted to an operator position.
While this invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations would be apparent to those skilled in the art. Accordingly, preferred embodiments in the invention as set forth herein are intended to be illustrative, not limiting. Various changes may be made without departing from the spirit and scope of the invention as defined in the following claims.
Claims (26)
1. A method for determining superwords and meaningful phrases, comprising the steps of:
determining superwords from a set of candidate phrases contained in a database of sequences of words, symbols and/or sounds based on commonality measurements;
determining meaningful phrases based on commonality measurements; and inserting the meaningful phrases into the database.
determining superwords from a set of candidate phrases contained in a database of sequences of words, symbols and/or sounds based on commonality measurements;
determining meaningful phrases based on commonality measurements; and inserting the meaningful phrases into the database.
2. The method of claim 1, wherein the step of determining superwords determines how each candidate phrase affects a perplexity value of the database so that if the perplexity value decreases, the candidate phrase is added to the database as a superword, and if the perplexity value is not decreased, the candidate phrase is not added to the database.
3. The method of claim 1, wherein the step of determining the meaningful phrases determines meaningful phrases based on salience and mutual information measurements
4. The method of claim 1, further comprising the step of:
training a speech recognizes using the database.
training a speech recognizes using the database.
5. The method of claim 1, further comprising the step of:
interpreting a user's input using the database.
interpreting a user's input using the database.
6. The method of claim 5, further comprising the step of:
implementing one or more of a set of task objectives based on the user's input.
implementing one or more of a set of task objectives based on the user's input.
7. An apparatus that determines superwords and meaningful phrases, comprising:
a database of commonly occurring words, symbols and/or sounds, the database having a perplexity value;
superword selecting means for selecting superwords based on commonality measurements, the selected superwords being input to the database; and meaningful phrase selecting means for selecting meaningful phrases based on commonality measurements, the selected meaningful phrases being input to the database.
a database of commonly occurring words, symbols and/or sounds, the database having a perplexity value;
superword selecting means for selecting superwords based on commonality measurements, the selected superwords being input to the database; and meaningful phrase selecting means for selecting meaningful phrases based on commonality measurements, the selected meaningful phrases being input to the database.
8. The apparatus of claim 7, wherein the superword selecting means creates a ranked set of candidate phrases from the database.
9. The apparatus of claim 8, wherein the superword selecting means inputs each of the candidate phrases into the database and determines whether the inserted candidate phrase decreases the perplexity value of the database.
10. The apparatus of claim 9, wherein the superword selecting means determines that if the perplexity value decreases, the inserted candidate phrase is added to the database as a superword, and if the perplexity is not decreased, the inserted candidate phrase is not added to the database.
11. The apparatus of claim 7, wherein the meaningful phrase selecting means selects the meaningful phrases based on salience and mutual information measurements.
12. The apparatus of claim 7, further comprising:
recognizing means for recognizing, based on a user's input, at least one meaningful phrase or superword contained in the database.
recognizing means for recognizing, based on a user's input, at least one meaningful phrase or superword contained in the database.
13. The apparatus of claim 12, wherein the database trains the recognizing means.
14. The apparatus of claim 12, further comprising:
interpretation means for making a classification decision based on the recognized meaningful phrases and superwords.
interpretation means for making a classification decision based on the recognized meaningful phrases and superwords.
15. The apparatus of claim 14, wherein the interpretation means implements one or more task objectives based on the user's input.
16. The apparatus of claim 15, further comprising:
dialog means for obtaining additional input from the user in order to implement one or more of the task objectives.
dialog means for obtaining additional input from the user in order to implement one or more of the task objectives.
17. An apparatus that determines superwords and meaningful phrases, comprising:
a database of words, symbols and sounds, the database having a perplexity value;
a superword selector that selects superwords based on commonality measurements, the superword selector inputting the selected superwords into the database; and a meaningful phrase selector that selects meaningful phrases based on commonality measurements, the meaningful phrase selector inputting the selected meaningful phrases into the database.
a database of words, symbols and sounds, the database having a perplexity value;
a superword selector that selects superwords based on commonality measurements, the superword selector inputting the selected superwords into the database; and a meaningful phrase selector that selects meaningful phrases based on commonality measurements, the meaningful phrase selector inputting the selected meaningful phrases into the database.
18. The apparatus of claim 17, wherein the superword selector creates a ranked set of candidate phrases from the database.
19. The apparatus of claim 18, wherein the superword selector inserts each separate candidate phrase into the database and determines whether the inserted candidate phrase decreases the perplexity value of the database.
20. The apparatus of claim 19, wherein the superword selector determines that if the perplexity value decreases, the candidate phrase is added to the database as a superword, and if the perplexity is not decreased, the candidate phrase is not added to the database.
21. The apparatus of claim 17, wherein the meaningful phrase selector selects a set of meaningful phrases based on salience and mutual information measurements.
22. The apparatus of claim 17, further comprising:
a recognizer that recognizes, from a user's input, at least one of the selected meaningful phrases and superwords contained in the database.
a recognizer that recognizes, from a user's input, at least one of the selected meaningful phrases and superwords contained in the database.
23. The apparatus of claim 22, wherein the database trains the recognizer.
24. The apparatus of claim 22, further comprising:
an interpretation module that makes a classification decision based on the recognized meaningful phrases and superwords.
an interpretation module that makes a classification decision based on the recognized meaningful phrases and superwords.
25. The apparatus of claim 24, wherein the interpretation module implements one or more task objectives based on the user's input.
26. The apparatus of claim 25, further comprising:
a dialog manager that obtains additional input from the user to implement one or more of the task objectives.
a dialog manager that obtains additional input from the user to implement one or more of the task objectives.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US08/960,289 US6044337A (en) | 1997-10-29 | 1997-10-29 | Selection of superwords based on criteria relevant to both speech recognition and understanding |
| US08/960,289 | 1997-10-29 | ||
| PCT/US1998/022399 WO1999022363A1 (en) | 1997-10-29 | 1998-10-23 | Selection of superwords based on criteria relevant to both speech recognition and understanding |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CA2275774A1 CA2275774A1 (en) | 1999-05-06 |
| CA2275774C true CA2275774C (en) | 2003-09-30 |
Family
ID=25503011
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002275774A Expired - Lifetime CA2275774C (en) | 1997-10-29 | 1998-10-23 | Selection of superwords based on criteria relevant to both speech recognition and understanding |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US6044337A (en) |
| EP (1) | EP0950240B1 (en) |
| CA (1) | CA2275774C (en) |
| DE (1) | DE69823644T2 (en) |
| WO (1) | WO1999022363A1 (en) |
Families Citing this family (34)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6173261B1 (en) | 1998-09-30 | 2001-01-09 | At&T Corp | Grammar fragment acquisition using syntactic and semantic clustering |
| US6185592B1 (en) * | 1997-11-18 | 2001-02-06 | Apple Computer, Inc. | Summarizing text documents by resolving co-referentiality among actors or objects around which a story unfolds |
| US6405171B1 (en) * | 1998-02-02 | 2002-06-11 | Unisys Pulsepoint Communications | Dynamically loadable phrase book libraries for spoken language grammars in an interactive system |
| WO2000073936A1 (en) * | 1999-05-28 | 2000-12-07 | Sehda, Inc. | Phrase-based dialogue modeling with particular application to creating recognition grammars for voice-controlled user interfaces |
| US20020032564A1 (en) | 2000-04-19 | 2002-03-14 | Farzad Ehsani | Phrase-based dialogue modeling with particular application to creating a recognition grammar for a voice-controlled user interface |
| US7085720B1 (en) | 1999-11-05 | 2006-08-01 | At & T Corp. | Method for task classification using morphemes |
| US8392188B1 (en) | 1999-11-05 | 2013-03-05 | At&T Intellectual Property Ii, L.P. | Method and system for building a phonotactic model for domain independent speech recognition |
| US7275029B1 (en) * | 1999-11-05 | 2007-09-25 | Microsoft Corporation | System and method for joint optimization of language model performance and size |
| US6681206B1 (en) | 1999-11-05 | 2004-01-20 | At&T Corporation | Method for generating morphemes |
| US20030191625A1 (en) * | 1999-11-05 | 2003-10-09 | Gorin Allen Louis | Method and system for creating a named entity language model |
| US7286984B1 (en) | 1999-11-05 | 2007-10-23 | At&T Corp. | Method and system for automatically detecting morphemes in a task classification system using lattices |
| US7725307B2 (en) * | 1999-11-12 | 2010-05-25 | Phoenix Solutions, Inc. | Query engine for processing voice based queries including semantic decoding |
| US7392185B2 (en) * | 1999-11-12 | 2008-06-24 | Phoenix Solutions, Inc. | Speech based learning/training system using semantic decoding |
| US9076448B2 (en) | 1999-11-12 | 2015-07-07 | Nuance Communications, Inc. | Distributed real time speech recognition system |
| US7050977B1 (en) * | 1999-11-12 | 2006-05-23 | Phoenix Solutions, Inc. | Speech-enabled server for internet website and method |
| US6606597B1 (en) * | 2000-09-08 | 2003-08-12 | Microsoft Corporation | Augmented-word language model |
| JP4517260B2 (en) * | 2000-09-11 | 2010-08-04 | 日本電気株式会社 | Automatic interpretation system, automatic interpretation method, and storage medium recording automatic interpretation program |
| US6941266B1 (en) | 2000-11-15 | 2005-09-06 | At&T Corp. | Method and system for predicting problematic dialog situations in a task classification system |
| US6751591B1 (en) | 2001-01-22 | 2004-06-15 | At&T Corp. | Method and system for predicting understanding errors in a task classification system |
| US7729918B2 (en) * | 2001-03-14 | 2010-06-01 | At&T Intellectual Property Ii, Lp | Trainable sentence planning system |
| WO2002073453A1 (en) * | 2001-03-14 | 2002-09-19 | At & T Corp. | A trainable sentence planning system |
| US7574362B2 (en) * | 2001-03-14 | 2009-08-11 | At&T Intellectual Property Ii, L.P. | Method for automated sentence planning in a task classification system |
| US7058575B2 (en) * | 2001-06-27 | 2006-06-06 | Intel Corporation | Integrating keyword spotting with graph decoder to improve the robustness of speech recognition |
| CA2481080C (en) * | 2002-04-05 | 2010-10-12 | At&T Corp. | Method and system for detecting and extracting named entities from spontaneous communications |
| US20030115062A1 (en) * | 2002-10-29 | 2003-06-19 | Walker Marilyn A. | Method for automated sentence planning |
| US20040193557A1 (en) * | 2003-03-25 | 2004-09-30 | Olsen Jesse Dale | Systems and methods for reducing ambiguity of communications |
| US7925506B2 (en) * | 2004-10-05 | 2011-04-12 | Inago Corporation | Speech recognition accuracy via concept to keyword mapping |
| US8478589B2 (en) | 2005-01-05 | 2013-07-02 | At&T Intellectual Property Ii, L.P. | Library of existing spoken dialog data for use in generating new natural language spoken dialog systems |
| US8185399B2 (en) * | 2005-01-05 | 2012-05-22 | At&T Intellectual Property Ii, L.P. | System and method of providing an automated data-collection in spoken dialog systems |
| US7912707B2 (en) * | 2006-12-19 | 2011-03-22 | Microsoft Corporation | Adapting a language model to accommodate inputs not found in a directory assistance listing |
| US7912503B2 (en) * | 2007-07-16 | 2011-03-22 | Microsoft Corporation | Smart interface system for mobile communications devices |
| US8165633B2 (en) * | 2007-07-16 | 2012-04-24 | Microsoft Corporation | Passive interface and software configuration for portable devices |
| US20110184736A1 (en) * | 2010-01-26 | 2011-07-28 | Benjamin Slotznick | Automated method of recognizing inputted information items and selecting information items |
| US9237224B2 (en) * | 2011-05-03 | 2016-01-12 | Padmanabhan Mahalingam | Text interface device and method in voice communication |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| DE69330427T2 (en) * | 1992-03-06 | 2002-05-23 | Dragon Systems Inc | VOICE RECOGNITION SYSTEM FOR LANGUAGES WITH COMPOSED WORDS |
| US5384892A (en) * | 1992-12-31 | 1995-01-24 | Apple Computer, Inc. | Dynamic language model for speech recognition |
| EP0938076B1 (en) * | 1992-12-31 | 2001-11-28 | Apple Computer, Inc. | A speech recognition system |
| US5794193A (en) * | 1995-09-15 | 1998-08-11 | Lucent Technologies Inc. | Automated phrase generation |
| US5675707A (en) * | 1995-09-15 | 1997-10-07 | At&T | Automated call router system and method |
| US5839106A (en) * | 1996-12-17 | 1998-11-17 | Apple Computer, Inc. | Large-vocabulary speech recognition using an integrated syntactic and semantic statistical language model |
| US5860063A (en) * | 1997-07-11 | 1999-01-12 | At&T Corp | Automated meaningful phrase clustering |
-
1997
- 1997-10-29 US US08/960,289 patent/US6044337A/en not_active Expired - Lifetime
-
1998
- 1998-10-23 DE DE69823644T patent/DE69823644T2/en not_active Expired - Lifetime
- 1998-10-23 WO PCT/US1998/022399 patent/WO1999022363A1/en active IP Right Grant
- 1998-10-23 CA CA002275774A patent/CA2275774C/en not_active Expired - Lifetime
- 1998-10-23 EP EP98953894A patent/EP0950240B1/en not_active Expired - Lifetime
Also Published As
| Publication number | Publication date |
|---|---|
| EP0950240A1 (en) | 1999-10-20 |
| US6044337A (en) | 2000-03-28 |
| CA2275774A1 (en) | 1999-05-06 |
| DE69823644D1 (en) | 2004-06-09 |
| DE69823644T2 (en) | 2005-04-28 |
| EP0950240B1 (en) | 2004-05-06 |
| WO1999022363A1 (en) | 1999-05-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA2275774C (en) | Selection of superwords based on criteria relevant to both speech recognition and understanding | |
| US8200491B2 (en) | Method and system for automatically detecting morphemes in a task classification system using lattices | |
| US5797123A (en) | Method of key-phase detection and verification for flexible speech understanding | |
| US6021384A (en) | Automatic generation of superwords | |
| US7139698B1 (en) | System and method for generating morphemes | |
| US5870706A (en) | Method and apparatus for an improved language recognition system | |
| Jelinek | Statistical methods for speech recognition | |
| US8612212B2 (en) | Method and system for automatically detecting morphemes in a task classification system using lattices | |
| US5953701A (en) | Speech recognition models combining gender-dependent and gender-independent phone states and using phonetic-context-dependence | |
| US6738745B1 (en) | Methods and apparatus for identifying a non-target language in a speech recognition system | |
| EP1593049B1 (en) | System for predicting speech recognition accuracy and development for a dialog system | |
| US20040186714A1 (en) | Speech recognition improvement through post-processsing | |
| EP0570660A1 (en) | Speech recognition system for natural language translation | |
| JPH11352994A (en) | Statistical sequence model generator, statistical language model generator, and speech recognition system | |
| US20050038647A1 (en) | Program product, method and system for detecting reduced speech | |
| US7085720B1 (en) | Method for task classification using morphemes | |
| US20040148169A1 (en) | Speech recognition with shadow modeling | |
| Rose | Word spotting from continuous speech utterances | |
| D'Orta et al. | Phoneme classification for real time speech recognition of Italian | |
| Young | Learning new words from spontaneous speech: A project summary | |
| Castro et al. | MLP emulation of N-gram models as a first step to connectionist language modeling | |
| Kobayashi et al. | A sub-word level matching strategy in a speech understanding system | |
| JPH0916192A (en) | Continuous speech recognition system and reference pattern training system | |
| Paeseler et al. | Continuous-Speech Recognition in the SPICOS-II System | |
| MXPA97002521A (en) | Method and apparatus for a better language recognition system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| EEER | Examination request | ||
| MKEX | Expiry |
Effective date: 20181023 |