WO2000057386A1

WO2000057386A1 - Method and system for computer assisted natural language instruction with adjustable speech recognizer

Info

Publication number: WO2000057386A1
Application number: PCT/US2000/007624
Authority: WO
Inventors: Marvin Shannon
Original assignee: Planetlingo, Inc.
Priority date: 1999-03-25
Filing date: 2000-03-22
Publication date: 2000-09-28
Also published as: AU3910500A

Abstract

A method of computerized language instruction for a student. Based on data regarding past performance of the student, an adjustable speech recognizer is adjusted. An utterance is received from the student, and the utterance is processed using the adjusted adjustable speech recognizer. The adjustable speech recognizer may comprise an Automatic Speech Recognition (ASR) engine. A set of contexts is created. Each context includes a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer. For each context, a set of subcontexts is created. Each subcontext includes the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context. Recursively passing a portion of a received utterance to an ASR engine is described. According to one embodiment, a context includes a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, and a set of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context. Systems based on the described methods are disclosed.

Description

METHOD AND SYSTEM FOR COMPUTER ASSISTED

NATURAL LANGUAGE INSTRUCTION WITH

ADJUSTABLE SPEECH RECOGNIZER

BACKGROUND

Field of the Invention

The invention relates to computer assisted natural language instruction. Description of Related Art Learning a language can be difficult. A foreign language student typically invests many years of study and may still never achieve fluency or reduction of an accent. Many students have invested heavily in language instruction by human teachers. Such training may be expensive, and inconsistent, because of the differences in training and ability of the teachers. Some foreign language instructional software has been developed. For example, systems are described in U.S. Patent Nos. 5,810,599 and 5,697,789. Some prior systems provide non-interactive language instruction. Such a system may not sufficiently provide feedback to a student to allow the student to correct pronunciation problems. The student may continue to make the same errors, without improvement even after diligent training with the system.

In addition to learning foreign languages, it is desirable to improve pronunciation and articulation of one's native language. It would be desirable to obtain a system and method to assist in doing so.

SUMMARY OF THE INVENTION An embodiment of the invention relates to a method of computerized language instruction for a student. Based on data regarding past performance of the student, an adjustable speech recognizer is adjusted. An utterance is received from the student, and the utterance is processed using the adjusted adjustable speech recognizer. The adjustable speech recognizer may comprise an Automatic Speech Recognition (ASR) engine.

The data regarding past performance of the student includes data regarding detected speech pathologies, or may comprise data regarding past performance of the student with a series of contexts, wherein a context comprises a set of words and utterances. The data regarding past performance of the student may comprise a statistical weighted average of data regarding past performance of the student with each of a series of contexts. A context may comprise a set of words and utterances selected to allow recognition of the words and utterances by the adjustable speech recognizer. Processing the utterance using the adjusted adjustable speech recognizer then may include determining whether the utterance matches an utterance in a selected context. According to an embodiment of the invention a set of contexts is created. Each context includes a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer. For each context, a set of subcontexts is created. Each subcontext includes the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context. A subcontext is selected from the set of contexts. An utterance is received from a student. The utterance is processed based on the selected subcontext, and the student is responded to based on the processing.

An embodiment of the invention includes recursively passing a portion of a received utterance to an ASR engine. The words in the utterance are determined based on a first output of the ASR engine and a second output of the

ASR engine resulting from processing the portion.

An embodiment of the invention includes creating a context. The context includes a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, and a set of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context. An audio input is received, and the input is compared with information in the context. If the input more closely matches a distracter than a word or utterance, the input is ignored. Otherwise, the input is processed to determine an associated word or utterance. Embodiments of the invention include systems based on the described methods. Embodiments of the invention include a computer readable medium having computer readable code embodied on the medium that causes a computer to execute the described methods or portions or combinations thereof. The computer readable medium may comprise a floppy disk, CD ROM, memory, or other medium.

One embodiment of the invention is a system for computerized language instruction for a student. The system includes a processor and logic to receive a digital representation of an audio signal. An adjustable speech recognizer is coupled to receive the digital representation. Logic in the system stores data onto a storage device based on performance of the student, and logic in the system adjusts the adjustable speech recognizer based on the stored data based on performance of the student.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention is illustrated by way of example, and not limitation in the drawings.

FIG. 1 is a block diagram of a system for computerized language instruction, according to an embodiment of the invention.

FIG. 2 is a schematic of a context and sets of pathologies, according to an embodiment of the invention.

DETAILED DESCRIPTION

The following is a description of embodiments of the invention. The embodiments shown help to illustrate the invention. However, it is not intended that the invention be limited to the precise embodiments shown.

One embodiment of the invention is a system for computerized language instruction. The system includes an ASR engine, which performs initial processing on an utterance by the student. An NLP engine performs further processing on the data yielded by the ASR engine. The ASR engine processes based on a context, which includes various words and utterances expected from the student. The context also includes various mispronunciations or misarticulations of the word or utterances. The ASR determines whether the incoming audio matches items in the context. Thus, incoming speech is recognized by finding a match with an item in the context. According to one embodiment of the invention, multiple subcontexts exist, which can be varied according to the past performance of the student in the system. The subcontexts include different mispronunciations or misarticulations of the words in the context and are selected so as to match different levels of development of the student. The selection is based on a grading of a student, and the grading is based on the performance of the student with the system. Optional subparsing provides additional processing of portions of the utterance in order to provide better resolution of individual words or portions of the utterance. Settings of the ASR engine, according to an embodiment of the invention, may be varied depending on the student's past performance with the system.

The system teaches the language through a series of lessons which may be taken over a period of sessions over many days. Student performance includes the correctness with which the student pronounces words in the language in the various lessons. Student performance may be tracked within the various lessons and over a period of time including various sessions with various lessons over many days.

FIG. 1 is a block diagram of a system for computerized language instruction, according to an embodiment of the invention. The following is a legend concerning FIG. 1 : ASR Engine: Automatic Speech Recognition

NLP Engine: Natural Language Processing Utterance Constraint Analysis Object: Determines subfragment to analyze

Sub-Parsing Request: Request to a fragment of the original Audio Data Grading/Level Control Object: Maintains the current grade level information based upon student history

Grade Constraint Data Object: Variable Pronunciation Constraints loaded from a database

Variable Context Data Object: Variable Mispronunciation Contexts selected from a family of contexts based upon grade level and success rate of student Utterance Data Object: Information output by ASR (N-best, timing, positional, etc.), and Candidate Utterance as determined by the NLP Token/Info: Results of ASR/NLP engine Audio Data: Wave format data buffer Sub-Sampled Audio Data: Fragment of original Audio Data.

The ASR / NLP engine 140 may be implemented in a computer system, such as a multimedia desktop computer including a processor, monitor, storage, microphone, and speakers. Alternatively, ASR / NLP engine may be implemented in another computer system, such a server on a network. The ASR / NLP engine 140 includes an Automatic Speech Recognition (ASR) engine 110 and Natural Language Processing (NLP) engine 112, which process input, Audio Data with Context ID 142 to provide token / info 144. ASR engine 110 receives the Audio Data w/ Context ID 142. Audio Data w/ Context ID 142 represents audio data, such as from a wave format data buffer, received by the computer, for example by a microphone coupled to a local computer.

The audio received may include an utterance by a student and other audio input, such as noise.

ASR Engine 110 processes Audio Data w/ Context ID 142 based on Variable Context Data Object 116. Variable Context Data Object 116 includes variable mispronunciation or misarticulation subcontexts selected from a family of subcontexts selected based on grade level and success rate of the student. Grading/Level Control Object 118 maintains the current grade level information based on student history and is used for selecting the subcontext by Variable Context Data Object 116. Grade Constraint Data Object 120 includes variable pronunciation constraints loaded from a database. Variable Context Data

Object 122 also includes variable mispronunciation or misarticulation subcontexts selected from a family of subcontexts selected based on grade level and success rate of the student and is an input to Grading/Level Control Object 118. Grade Constraint Data Object 124 includes variable pronunciation constraints loaded from a database and is an input to NLP Engine 112.

Utterance Data Object 114 is output by ASR Engine 110 and includes such information as an N-best list, timing, and positional information. Utterance Data Object 114 is passed to NLP 112 to determine a candidate utterance, which is output in the form of Token / Info 144.

ASR / NLP Engine 140 recursively passes a portion or portions of the Utterance Data Object 126 into the ASR / NLP Engine. This helps to more accurately determine the words contained within the utterance. As shown,

Utterance Data Object 126 is passed into block Sub-Parsing Request 132. Utterance Constraint Analysis Object 128 receives Utterance Data Object 126 and determines the portion, or subfragment, of Utterance Data Object 126 to analyze. The Sub-Sampled Audio Data with Sub-Context ID 146 is passed from Utterance Constraint Analysis Object 128 to ASR / NLP Engine 130 for processing. The resulting Token/Info 148 is passed to Utterance Constraint Analysis Object 128, which then yields a result as Token / Info 144.

ASR / NLP engine is described with many of the components shown as objects, which may be implemented in an object oriented paradigm. Such functionality may also be implemented in other programming paradigms, such as in a structured programming architecture using function calls, or other software architecture. Alternatively, ASR / NLP engine functionality may be implemented in hardware.

One example of an ASR takes input from a microphone or wavebuffer and makes a determination based on a statistical analysis to determine what words of series of words are associated with incoming speech. An N-best list is built, which includes a set of N words for each position in the utterance and respective probabilities for each word. An ASR is adjustable in that it has a number of settings which determine how it processes incoming data. The settings may include voice model (male / female / child), number of states

(number of states in the Hidden Markov Model, which relate to granularity), rejection penalties, and garbage penalties. According to one example of an ASR, adjustments to the ASR include (1) number of processing states, (2) amount of time the ASR spends processing the utterance, (3) garbage penalty, (4) rejection penalty, (5) minimum duration, and (6) sensitivity. For discussion of computer speech recognition, see Fundamentals of Speech Recognition, by Lawrence Rabiner and Biing-Hwang Juang, published by Prentice Hall, Englewood Cliffs, NJ, which is incorporated herein by reference in its entirety. One embodiment of the invention is a system with an adjustable speech recognizer, such as an ASR, where the adjustable speech recognizer is adjusted based on the student performance. Accordingly, according to an embodiment of the invention, the Grading Level Control Object 118 allows the system or a user, such as the student or teacher, to set a series of variables to adjust the ASR to accept or reject utterances unconditionally or to accept conditionally. The variables to be set include states used with the Hidden Markov Model, rejection penalties, garbage penalties, voice models and other items. When under the control of the program, the variables can be adjusted so that the ASR is more accepting or less accepting.

Past student performance is used to adjust an ASR, according to one embodiment of the invention, by statistically combining data corresponding to student performance in various aspects of the system. An ASR's settings may be selected based on statistical cumulative percentages of correct performance by the student in the past. If the student has performed well in a large number of past sessions, then ASR settings of the ASR are not varied as greatly as in a case where the student has not performed well in a large number of past sessions. An advantage of setting the ASR in this manner is that the student's current performance may be low because of an aberration that day, e.g., the student is tired, and may not warrant a large change in the settings of the ASR. According to one embodiment of the invention, selected ones or all of the following parameters of an ASR are adjusted based on past student performance: (1) number of processing states, (2) amount of time the ASR spends processing the utterance, (3) garbage penalty, (4) rejection penalty, (5) minimum duration, and (6) sensitivity.

The parameters are adjusted based on the grade level of the student, and the grade level of the student is determined based on the past performance of the student with the system. For each grade level, a particular set of ASR parameters are selected. The set of ASR parameters associated with a grade level are determined empirically for the particular ASR and hardware, such as microphone, sound cards, etc. For the most challenging grade level, settings are chosen based on the settings that yield the best recognition by the ASR of voice input from a number of excellent speakers of the language which for which instruction is given by the system (the target language). For other grade levels, sets of speakers are chosen with other respective levels of ability in the target language. The other sets of speakers are typically native speakers of another language, other than the target language. This other language is the language of the student for whom the system is designed. For example, if the system is designed to teach English to Japanese speakers, the first set of speakers for the most advanced grade level may be perfect speakers of English. The speakers used to develop the settings for the other grade levels would be native speakers of Japanese with respectively varying abilities in English. The system may be used for teaching a first language, or accent reduction for speakers of one dialect who wish to be understood in another dialect. For example, the system may be used to teach British English to persons who first learned Singapore

English. In such cases, appropriate groups of speakers are chosen to develop the respective grade levels.

According to one embodiment of the invention, statistics are stored based on the student's past pathologies on a per pathology basis. Statistics are also stored for the student for each context. A student progress is tracked as a grade level (Gw) and a pathology level (Pw). A Word Error Rate (WER) and Word Pathology Rate (WPR) are tracked. For a new student (user), Pw = Pmax and Gw = Gmid. For a returning student (user) Pw = Phistory and Gw = Ghistory. The following sequence is followed to determine ASR settings, according to an embodiment of the invention,

if WER decreasing if WPR decreasing decrease Pn -> (Pn-1 || P0) if WPR increasing increase Pn -> (Pn+1 || Pmax) increase Gn -> (Gn+1 || Gmax) clear loop counter else if WER increasing if at Pmax decrease Gn -> (Gn-1 || GO) clear loop counter increment loop counter at low loop counter increase Pn ->(Pn+j || Pmax) [Pn+j≡last active pathology] at high loop counter increase Pn -> (Pn+k || Pmax) where K is some number of pathology levels as determined by a fixed percentage of total pathology levels In the code shown above, a loop may comprise completion of some portion of language instruction, such as a unit of 4 or 5 utterances.

Thus, a grade level is determined and changed dynamically for a student as the student uses the system. The grade level is one example of data regarding past performance of the student and can be used to adjust the ASR, based on a set of ASR settings associated with the grade level. In addition to adjusting the ASR settings, a subcontext may be chosen based on the student's past performance. According to one embodiment of the invention, the ASR settings are changed to make the ASR more sensitive when the student gets a large number of words correct. The context is changed to a lower level if the student tends to fail lessons. This adjustment is made because the failure of lessons may be an indication of a specific pathology, which can be addressed by a subcontext which includes mispronunciations or misarticulations based on that pathology.

According to one embodiment of the invention, the student is prompted to speak about a one particular topic at a time. For each topic, the system stores a context, which includes words and utterances related to the topic. The words and utterances in the context are chosen so as to help the ASR to distinguish the words in the context. The context may be stored as text or binary representation, BNF grammar, or other format.

According to one embodiment of the invention, the system presents to the student graphics of various physical settings. For example, the settings may include a restaurant, a store, a home, or an airport. In the various physical settings or in portions of the various physical settings, the student may speak about relevant topics. The physical setting or portion of the physical setting may correspond to a context.

A context is used both by an ASR and an NLP. The context is used both for the purposes of comparison and utterance strength.

For each expected utterance the program receives, it has pre-defined context that includes the anticipated utterances. Included in this context are expected mispronunciations or misarticulations according to an embodiment of the invention. For example, the program expects "I would like some rice." In addition to the example being in the context, mispronunciations such as "I would rike some lice" or "I would rike some rice" would also be included. The inclusion of specific items in the context will depend to some degree on the grading level used at a particular time.

A set of contexts is created. A context would typically be created by a linguist. Each context includes a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, such as an ASR. Each context includes mispronunciations or misarticulations of the words in the context. Each context is divided into subcontexts, with each subcontext including various mispronunciations or misarticulations of the words in the context. The subcontexts may be created so that they reflect mispronunciations or misarticulations encountered in various stages of learning the language, or may target a particular pathology.

According to an embodiment of the invention, the subcontext is selected based on past performance of the student. If past performance of the student indicates a particular pathology that can be effectively addressed with a particular subcontext, that particular subcontext is chosen. The first subcontext chosen for a beginning student would include mispronunciations or misarticulations typical of a beginning student of a particular native language learning the language. Therefore, different families of subcontexts are designed for native speakers of different languages learning the same language. When the ASR / NLP or other speech processing system detects mispronunciations or misarticulations in the received utterance, the system provides the student feedback regarding the mispronunciation or misarticulations.

FIG. 2 is a schematic of a context and sets of pathologies, according to an embodiment of the invention. Context 210 includes a number of utterances and words, such as the word "love" 222 and the word "umbrella" 224. Set of pathologies 220 includes subset 226, which includes mispronunciation "rove"

234 and "umbrerra" 236. Set of pathologies 220 also includes subsets 228, 230, and 232. Context 210 is a dictionary of words and phrases. Set of pathologies 220 is a dictionary of all of the most frequent mispronunciations or misarticulations that occur when learning one language from another (for example - when a Japanese speaker is learning English). According to an embodiment of the invention, the pathologies are sorted from those that occur most to those that occur least. For example, the English words that are typically the most difficult for Japanese speakers to pronounce will have a higher acceptance level in the pathology box. Words that are typically not as difficult to pronounce will have a lower threshold for mispronunciation. The system uses the subcontexts in the sorted order.

Thus, for a particular context "C," Co = the entire context (the entire dictionary of words) Ci = the entire context plus Pathology group A (The mispronunciations which occur the most)

C₂ = the entire context plus Pathology group B (The mispronunciations which occur second most)

OR C₂ = Ci plus Pathology group B — C_N = the entire context plus the entire pathology set. (All mispronunciations or misarticulations)

This would be for very beginners for example. They would be allowed the lowest threshold and many of their pronunciation mistakes would be accepted at first.

An advantage of development of subcontexts is flexibility. Given a set of subcontexts, an appropriate subcontext can be chosen to correspond to the specific speech problems of the student. Further, the subcontext helps to identify specific linguistic problems matching the mispronunciations or misarticulations in the subcontext. The approximate severity of the student's problem can also be identified. Computational burden on an ASR is reduced by separating mispronunciations into various subcontexts.

According to one embodiment of the invention, a context includes a set of distracters, which comprise sets of syllables used less frequently in the context. The sets of syllables in the distracters do not form words or utterances in the context. The distractors, according to one embodiment, have lengths in the range of 1 to n+2 syllables, wherein n comprises the number of syllables the word having the largest number of syllables in the context. A distracter does not have the same syllable in a row. An example of syllables used to form distractors would be nonsense syllables such as "um" or "er." The context does not include more that approximately 10 or 12 distractors of a given length.

An advantage of inclusion of distractors in a context is that noise may be mapped to the distractors. This occurs because, given the random pattern of the distractors, they tend to be more similar to noise than actual words and utterances in the context.

If the system detects noise, it ignores the noise. However, if the system detects an utterance for which it does not know the meaning, it responds. For example, in a context related to computers, if the student asks for a potato, the system may respond, "Sorry, I don't know what that means here." Various embodiments of the invention have been illustrated in the figures and have been described in the corresponding text of this application. This foregoing description is not intended to limit the invention to the precise forms disclosed. Rather, the invention is to be construed to the full extent allowed by the following claims and their equivalents. What is claimed is:

Claims

CLAIMS 1. A method of computerized language instruction for a student, the method comprising: based on data regarding past performance of the student, adjusting an adjustable speech recognizer; receiving an utterance from the student; and processing the utterance using the adjusted adjustable speech recognizer.

2. The method of claim 1 , wherein the adjustable speech recognizer comprises an Automatic Speech Recognition (ASR) engine.

3. The method of claim 1 , wherein the data regarding past performance of the student comprises data regarding past performance of the student with the computerized language instruction.

4. The method of claim 1 , wherein the data regarding past performance of the student includes data regarding detected speech pathologies.

5. The method of claim 1 , wherein the data regarding past performance of the student comprises data regarding past performance of the student with a series of contexts, wherein a context comprises a set of words and utterances.

6. The method of claim 1 , wherein the data regarding past performance of the student comprises a statistical weighted average of data regarding past performance of the student with each of a series of contexts, wherein context comprises a set of words and utterances.

7. The method of claim 1 , including creating a set of contexts, each context including a set of words and utterances selected to allow recognition of the words and utterances by the adjustable speech recognizer, and wherein processing the utterance using the adjusted adjustable speech recognizer includes determining whether the utterance matches an utterance in a selected context.

8. The method of claim 6, including, for each context, creating a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and wherein processing the utterance using the adjustable speech recognizer includes determining whether the utterance matches an utterance in a selected subcontext.

9. The method of claim 8, including, based on data regarding past performance of the student, selecting the subcontext.

10. The method of claim 9, including increasing sensitivity of the adjustable speech recognizer primarily based the student pronouncing a set of words correctly.

11. The method of claim 9, including decreasing difficulty the subcontext primarily based on the student not pronouncing a specific word correctly.

12. The method of claim 10, including decreasing difficulty of the subcontext primarily based on the student not pronouncing a specific word correctly.

13. The method of claim 1 , wherein the adjustable speech recognizer has a voice model adjustment including setting of male, female, and child and wherein adjusting the adjustable speech recognizer comprises selecting a voice model.

14. The method of claim 1 , wherein the voice model has a number of states in a Hidden Markov Model, and wherein adjusting the adjustable speech recognizer comprises adjusting the states.

15. The method of claim 1 , including, after processing the utterance using the adjusted adjustable speech recognizer, receiving an output from the adjusted adjustable speech recognizer and processing the output using a Natural Language Processing (NLP) engine.

16. A system for computerized language instruction for a student, the method comprising: an adjustable speech recognizer; logic that, based on data regarding past performance of the student, adjusts the adjustable speech recognizer; logic that receives an utterance from the student; and logic that processes the utterance using the adjusted adjustable speech recognizer.

17. The system of claim 16, wherein the adjustable speech recognizer comprises an Automatic Speech Recognition (ASR) engine.

18. The system of claim 16, wherein the data regarding past performance of the student comprises data regarding past performance of the student with a series of contexts, wherein a context comprises a set of words and utterances.

19. The system of claim 16, including logic to process the utterance using the adjusted adjustable speech recognizer by determining whether the utterance matches an utterance in a selected context, wherein the context includes a set of words and utterances selected to allow recognition of the words and utterances by the adjustable speech recognizer.

20. The system of claim 19, wherein the context includes a subcontext, the subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and wherein the adjustable speech recognizer determines whether the utterance matches an utterance in a selected subcontext.

21. The system of claim 20, including, logic that, based on data regarding past performance of the student, selects the subcontext.

22. A system for computerized language instruction for a student, the system including: a processor; logic to receive a digital representation of an audio signal; an adjustable speech recognizer coupled to receive the digital representation; a storage device; logic that stores data onto the storage device based on performance of the student; and logic that adjusts the adjustable speech recognizer based on the stored data based on performance of the student.

23. The system of claim 22, wherein the adjustable speech recognizer comprises an Automatic Speech Recognition (ASR) engine.

24. The system of claim 22, including logic that selects a subcontext stored on the storage device and causes the adjustable speech recognizer to process the audio signal based on the subcontext, wherein the subcontext includes selected mispronunciations or misarticulations of words or utterances.

25. The system of claim 22, including logic that causes the adjustable speech recognizer to first process the entire audio signal and then process a selected portion of the audio signal, the portion selected based on uncertainty as to the words corresponding to the portion.

26. The system of claim 22, including a context stored on the storage device, the context including a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, and a set of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context; and logic to cause the adjustable speech recognizer to process the audio signal based on the context.

27. A method of computerized language instruction for a student, the method comprising: creating a set of contexts, each context including a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, for each context, creating a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, selecting a subcontext from the set of contexts; receiving an utterance from a student; processing the utterance based on the selected subcontext; and responding to the student based on the processing.

28. The method of claim 27, wherein subcontexts in the set of subcontexts include mispronunciations or misarticulations based on linguistic pathologies.

29. The method of claim 27, wherein the student has a native language, and wherein subcontexts in the set of subcontexts include mispronunciations or misarticulations based on linguistic pathologies typical in persons having the native language trying to learn the language.

30. The method of claim 27, wherein the set of contexts includes contexts differing based on mispronunciations or misarticulations typically made in different stages of language learning.

31. The method of claim 27, including selecting the subcontext based on data regarding past performance of the student.

32. The method of claim 31 , including, based on the data regarding past performance of the student, adjusting an adjustable speech recognizer that processes the utterance.

33. The method of claim 32, including selecting the subcontext based primarily on data regarding past performance with a relatively small set of words and adjusting the adjustable speech recognizer based on data regarding past performance with a relatively large set of words.

34. The method of claim 27, wherein contexts in the set of contexts include sets of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context, and the method including ignoring input if the input matches a distracter in the set of distracters.

35. The method of claim 34, wherein the distracters of a particular context have lengths in the range of 1 to n+2 syllables, wherein n comprises the number of syllables the word having the largest number of syllables in the particular context.

36. The method of claim 27, including storing the contexts as text.

37. The method of claim 27, including storing the contexts in BNF grammar form.

38. The method of claim 27, wherein words and utterances in each context correspond to a topic of conversation.

39. A system for computerized language instruction for a student, the system comprising: a memory device including a set of contexts, each context including a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, for each context, a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, logic that selects a subcontext from the set of contexts; logic that receives an utterance from a student; logic that processes the utterance based on the selected subcontext; and logic that responds to the student based on the processing.

40. The system of claim 39, wherein subcontexts in the set of subcontexts include mispronunciations or misarticulations based on linguistic pathologies.

41. The system of claim 39, wherein the student has a native language, and wherein subcontexts in the set of subcontexts include mispronunciations or misarticulations based on linguistic pathologies typical in persons having the native language trying to learn the language.

42. The system of claim 39, wherein the set of contexts includes contexts differing based on mispronunciations or misarticulations typically made in different stages of language learning.

43. The system of claim 39, including logic that selects the subcontext based on data regarding past performance of the student.

44. The system of claim 43, including: an adjustable speech recognizer for processing the utterance; and logic that, based on the data regarding past performance of the student, adjusts the adjustable speech recognizer.

45. The system of claim 43, wherein the logic that selects a subcontext from the set of contexts comprises software code.

46. The system of claim 43, wherein the logic that receives an utterance from a student comprises hardware.

47. A method of computerized language instruction for a student, the method comprising: receiving an utterance from the student; processing the utterance using an Automatic Speech Recognition (ASR) engine; receiving a first output from the ASR engine; processing the output using a Natural Language Processing (NLP) engine; receiving an output from the NLP engine; based on the output from the NLP engine, selecting a portion of the utterance; processing the portion of the utterance using the ASR engine; receiving a second output from the ASR engine; and based on the first output and the second output of the ASR engine, determining words in the utterance.

48. The method of claim 47, including selecting the portion of the utterance based on a confidence in the processing of the portion of the utterance.

49. The method of claim 47 wherein the portion of the utterance is selected to be likely a single word.

50. The method of claim 47, including adjusting the ASR engine based on data regarding past performance of the student.

51. The method of claim 47, including creating a set of contexts, each context including a set of words and utterances selected to allow recognition of the words and utterances by the ASR engine, and wherein the first output from the ASR engine corresponds to an utterance in the context and the second output from the ASR engine comprises a word in the context.

52. The method of claim 51 , including, for each context, creating a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and wherein the ASR engine determines whether the utterance matches an utterance in a selected subcontext.

53. The method of claim 52, including, based on data regarding past performance of the student, selecting the subcontext.

54. A system for computerized language instruction for a student, the system comprising: logic that receives an utterance from the student; an Automatic Speech Recognition (ASR) engine; logic that processes the utterance using the ASR engine; logic that receives a first output from the ASR engine; a Natural Language Processing (NLP) engine; logic that processes the output using the NLP engine; logic that receives an output from the NLP engine; logic that, based on the output from the NLP engine, selects a portion of the utterance; logic that processes the portion of the utterance using the ASR engine; logic that receives a second output from the ASR engine; and logic that, based on the first output and the second output of the ASR engine, determines words in the utterance.

55. The system of claim 54, including logic that selects the portion of the utterance based on a confidence in the processing of the portion of the utterance.

56. The system of claim 54 wherein the portion of the utterance is selected to be likely a single word.

57. The system of claim 54, including logic that adjusts the ASR engine based on data regarding past performance of the student.

58. The system of claim 54, including a memory including a set of contexts, each context including a set of words and utterances selected to allow recognition of the words and utterances by the ASR engine, and wherein the first output from the ASR engine corresponds to an utterance in the context and the second output from the ASR engine comprises a word in the context.

59. The system of claim 58, including, a memory including for each context, a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and ASR engine determines whether the utterance matches an utterance in a selected subcontext.

60. The system of claim 59, including, logic that, based on data regarding past performance of the student, selects the subcontext.

61. A computer program product operating a computer for computerized language instruction for a student, the computer program product comprising: a computer usable medium having computer readable program code embodied in the medium, the computer usable code including: code that causes the computer to receive an utterance from the student; code that causes the computer to process the utterance using an Automatic Speech Recognition (ASR) engine; code that causes the computer to receive a first output from the ASR engine; code that causes the computer to process the output using a Natural Language Processing (NLP) engine; code that causes the computer to receive an output from the NLP engine; code that causes the computer to, based on the output from the NLP engine, select a portion of the utterance; code that causes the computer to process the portion of the utterance using the ASR engine; code that causes the computer to receive a second output from the ASR engine; and code that, based on the first output and the second output of the ASR engine, determines words in the utterance.

62. A method of computerized language instruction for a student, the method comprising: creating a context including a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, and a set of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context; receiving an audio input; comparing the input with information in the context; if the input more closely matches a distracter than a word or utterance, ignoring the input; and otherwise, processing the input to determine an associated word or utterance.

63. The method of claim 62, wherein the distracters have lengths in the range of 1 to n+2 syllables, wherein n comprises the number of syllables the word having the largest number of syllables in the context.

64. The method of claim 62, wherein a distracter does not include the same syllable next to the same syllable.

65. The method of claim 62, including processing the input with a speech recognizer and adjusting the speech recognizer based on data regarding past performance of the student.

66. The method of claim 62, including creating a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and wherein processing the input to determine an associated word or utterance includes determining whether the input matches a word or utterance or selected mispronunciation or misarticulation thereof in a selected subcontext.

67. The method of claim 66, including, based on data regarding past performance of the student, selecting the subcontext.

68. A system for computerized language instruction for a student, the system comprising: memory including a context having a set of words and utterances selected to allow recognition of the words and utterances by a speech recognizer, and a set of distracters, which comprise sets of syllables used less frequently in the context, the sets of syllables not forming words or utterances in the context; logic that receives an audio input; logic that compares the input with information in the context; logic that, if the input more closely matches a distracter than a word or utterance, ignores the input; and logic that otherwise processes the input to determine an associated word or utterance.

69. The system of claim 68, wherein the distracters have lengths in the range of 1 to n+2 syllables, wherein n comprises the number of syllables the word having the largest number of syllables in the context.

70. The system of claim 68, wherein a distracter does not include the same syllable next to the same syllable.

71. The system of claim 68, including logic that processes the input with a speech recognizer and adjusts the speech recognizer based on data regarding past performance of the student.

72. The system of claim 68, the memory including a set of subcontexts, each subcontext including the words and utterances of the context and selected mispronunciations or misarticulations of the words and utterances of the context, and wherein the system includes logic that processes the input to determine whether the input matches a word, utterance, or mispronunciation or misarticulation thereof in a selected subcontext.

73. The system of claim 72, including, logic that, based on data regarding past performance of the student, selects the subcontext.