NZ705071B2 - Method and system for predicting speech recognition performance using accuracy scores - Google Patents
Method and system for predicting speech recognition performance using accuracy scores Download PDFInfo
- Publication number
- NZ705071B2 NZ705071B2 NZ705071A NZ70507112A NZ705071B2 NZ 705071 B2 NZ705071 B2 NZ 705071B2 NZ 705071 A NZ705071 A NZ 705071A NZ 70507112 A NZ70507112 A NZ 70507112A NZ 705071 B2 NZ705071 B2 NZ 705071B2
- Authority
- NZ
- New Zealand
- Prior art keywords
- speech recognition
- prediction
- performance
- merit
- keyword
- Prior art date
Links
- 239000011159 matrix material Substances 0.000 claims description 15
- 230000000877 morphologic Effects 0.000 claims description 11
- 238000001514 detection method Methods 0.000 claims description 9
- 230000006399 behavior Effects 0.000 abstract description 7
- 238000010801 machine learning Methods 0.000 abstract description 6
- 238000000034 method Methods 0.000 description 20
- 230000002452 interceptive Effects 0.000 description 11
- 241000282326 Felis catus Species 0.000 description 6
- 238000010586 diagram Methods 0.000 description 6
- 150000002500 ions Chemical class 0.000 description 6
- 230000036461 convulsion Effects 0.000 description 4
- 239000000470 constituent Substances 0.000 description 3
- 230000001419 dependent Effects 0.000 description 3
- 238000003780 insertion Methods 0.000 description 3
- 230000003595 spectral Effects 0.000 description 3
- 230000004048 modification Effects 0.000 description 2
- 238000006011 modification reaction Methods 0.000 description 2
- 230000005236 sound signal Effects 0.000 description 2
- 238000006467 substitution reaction Methods 0.000 description 2
- 241000229754 Iva xanthiifolia Species 0.000 description 1
- 241000227653 Lycopersicon Species 0.000 description 1
- 235000007688 Lycopersicon esculentum Nutrition 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 150000001768 cations Chemical class 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 230000001276 controlling effect Effects 0.000 description 1
- 230000000875 corresponding Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000001537 neural Effects 0.000 description 1
- 230000001737 promoting Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Abstract
system and method are presented for predicting speech recognition performance using accuracy scores in speech recognition systems within the speech analytics field. A keyword set is selected. Figure of Merit (FOM) is computed for the keyword set. Relevant features that describe the word individually and in relation to other words in the language are computed. A mapping from these features to FOM is learned. This mapping can be generalized via a suitable machine learning algorithm and be used to predict FOM for a new keyword. In at least embodiment, the predicted FOM may be used to adjust internals of speech recognition engine to achieve a consistent behavior for all inputs for various settings of confidence values. ly and in relation to other words in the language are computed. A mapping from these features to FOM is learned. This mapping can be generalized via a suitable machine learning algorithm and be used to predict FOM for a new keyword. In at least embodiment, the predicted FOM may be used to adjust internals of speech recognition engine to achieve a consistent behavior for all inputs for various settings of confidence values.
Description
TITLE
METHOD AND SYSTEM FOR PREDICTING SPEECH ITION PERFORMANCE USING ACCURACY
SCORES
OUND
The present invention lly relates to telecommunication systems and methods, as well as
automatic speech recognition s. More particularly, the t invention pertains to machine
learning within automatic speech recognition systems.
It is known in the art that speech recognition may be performed by measuring a system’s ability
to recognize a target word by analyzing its audio file with reference to r audio file(s) of a set of
words. The target word may then be separated from the set of words if it does not meet a certain
recognition threshold. By separating below-threshold target words from the set of words, the set may
be restricted to readily-identified words. The words can thus be used in a speech recognition
application with a certain degree of confidence. However this process can be time-consuming, and
impractical in many applications. Having a system that can predict recognition accuracy of a target word,
without the need for processing a large set of audio files to measure ition rate, enables a user to
tand how the system will perform in the real world without having to wait for a full deployment,
thus saving money, effort, and resources.
SUMMARY
A system and method are presented for predicting speech recognition performance using
accuracy scores in speech recognition s within the speech analytics field. The same keyword set
is used hout. Figure of Merit (FOM) is a measure used to describe accuracy of speech recognition
systems and keyword spotting systems in particular. It is defined as the detection rate for an average of
false alarms per keyword per hour (FA/KW/Hr). In at least one embodiment, FOM is predicted
through an algorithm which is discussed in greater detail below. The FOM uses several features of a
keyword in order to predict the accuracy with which a system can determine a word match. For each
keyword within the set, the keyword r is run on a large body of recorded speech to determine the
FOM. nt features that describe the word individually and in relation to other words in the
language are computed. A mapping from these features to FOM is learned. This g can then be
generalized via a suitable machine learning algorithm and be used to predict FOM for a new keyword.
The predicted FOM may be used to adjust internals of speech recognition engine to achieve a consistent
behavior for all inputs for s settings of ence values.
In one embodiment, a computer-implemented method for predicting speech recognition
performance is disclosed, comprising the steps of: ing an input; ing at least one feature
vector for said input; ing said at least one feature vector into a prediction model; and obtaining a
prediction for the input from the prediction model.
In another embodiment a system for predicting speech recognition performance is disclosed,
sing: means for accepting an input; means for computing at least one feature vector for said user
input; means for inputting said at least one feature vector into a prediction model; and means for
obtaining a prediction of figure of merit for the input from the prediction model.
In another embodiment a computer-implemented method for using predicted speech
recognition performance to adjust internal scores of a speech recognition engine is disclosed, the
method comprising the steps of: accepting an input; computing at least one feature vector for said
input; inputting said at least one feature vector into a prediction model; obtaining a prediction for figure
of merit for the keyword; and adjusting a mapping of said internal scores to confidence values based on
said prediction.
BRIEF DESCRIPTION OF THE DRAWINGS
Figure 1 is a diagram illustrating an exemplary system for keyword spotting.
Figure 2 is a flowchart illustrating a s for FOM prediction.
Figure 3 is an illustration of a user interface.
Figure 4 is a flowchart illustrating a system for model learning.
Figure 5 is a flowchart illustrating a process for choosing the training keyword set.
Figure 6 is a diagram illustrating the relation between the internal match ”Score” and external
”Confidence” values.
Figure 7 is an ration of FOM in relation to detection accuracy and false alarms per hour of
speech
Figure 8 is a table illustrating keyword examples.
Figure 9 is a table illustrating FOM model expression.
DETAILED DESCRIPTION
For the purposes of promoting an understanding of the principles of the invention, reference
will now be made to the embodiment illustrated in the drawings and specific ge will be used to
describe the same. It will nevertheless be understood that no limitation of the scope of the invention is
thereby intended. Any alterations and further modifications in the described embodiments, and any
further ations of the principles of the invention as described herein are contemplated as would
normally occur to one skilled in the art to which the invention s.
tic speech recognition (ASR) systems analyze human speech and translate the speech
into text or words. Performance of these systems is commonly evaluated based on the accuracy,
reliability, language support, and the speed with which speech can be recognized. The performance of
the system is ed to be very high. Superior performance is often fied by a high detection
rate and a low false alarm rate. Industry standard is considered to be around a 70% detection rate at 5
false alarms per keyword per hour of speech, or 5 FA/KW/Hr. This may be read as an FOM of 70.
Factors such as accent, articulation, speech rate, pronunciation, ound noise, etc., can have a
negative effect on the accuracy of the system. Processing speed is ary to analyze hundreds of
telephone conversations at once and in real-time. The system is also expected to m consistently
and reliably irrespective of channel conditions and various artifacts introduced by modern telephony
channels, especially VolP. Keywords from multiple ges also need to be d on the same audio
source.
e learning may be used to predict the performance of an engine on a ular keyword.
Supervised ng may be ed to as the machine learning task of inferring a function from
supervised, or labeled, training data. Such training data may consist of a set of training examples, which
represent accuracy values for a large set of keywords. |n supervised learning, each training example is a
pair consisting of an input e vector and a desired output accuracy value. A supervised learning
algorithm analyzes the training data and produces an inferred function, or regression function. Such
function should predict the correct output value for any valid input object. This requires the learning
algorithm to generalize from the training data to unseen ions in a ”reasonable” way. The
regression function may be modeled using a variety of forms such as a simple straight line to a complex
neural network.
Those skilled in the art will recognize from the present disclosure that the various
methodologies disclosed herein may be computer implemented using a great many different forms of
data processing equipment, such as digital microprocessors and ated memory executing
appropriate software program(s), to name just one non-limiting example. The specific form of the
hardware, firmware and software used to implement the presently disclosed embodiments is not critical
to the present invention.
A method and system is defined for predicting speech recognition mance using cy
scores. The same d set is used throughout. FOM is computed for each keyword in the keyword
set. A FOM is determined through an algorithm which is discussed in greater detail below. The FOM
uses several features in order to predict the accuracy with within a system can determine a word match.
For each keyword within the set, the keyword spotter is run on a large body of recorded speech to
determine the FOM. Relevant features that be the word individually and in relation to other
words in the language are computed. A g from these features to FOM is learned. This mapping
can then be generalized via a suitable machine learning algorithm and be used to predict FOM for a new
Figure 1 is a diagram illustrating an exemplary system for keyword spotting, 100. The basic
components of a keyword spotter 100 may include: User Data/Keywords 105; d Model 110;
Knowledge Sources 115, which may include an Acoustic Model 120 and a Lexicon/Pronunciation
Predictor 125; an Audio Stream 130; a Front End e Calculator 135; a Recognition Engine (Pattern
Matching) 140; and the Report of Found Keywords in Real-Time 145.
Keywords 105 may be defined by the user of the system ing to user preference. The
Keyword Model 110 may be formed by concatenating phoneme hidden Markov models (HMMs) or any
other statistical representation of lexical units that comprise a word. The Keyword Model 110 may be
composed based on the keywords that are defined by the user and the input to the Keyword Model 110
based on dge Sources 115. Such Knowledge Sources may include an Acoustic Model 120 and a
n/Pronunciation Predictor 125.
The Knowledge Sources 115 may store probabilistic models of ons between pronunciations
and acoustic events. The Knowledge Sources 115 may be developed by analyzing large quantities of
audio data. The Acoustic Model 120 and the Lexicon/Pronunciation Predictor 125 are made, for
example, by looking at a word like ”hello” and examining the phonemes that comprise the word. Every
keyword in the system is represented by a statistical model of its constituent sub-word units called the
es. The phonemes for ”hello” as defined in a standard phoneme dictionary are: ”hh”, ”eh”, ”I”,
and ”ow”. Models of the four phonemes are then strung er into one composite model which then
becomes the keyword model for the world ”hello”. These models are language dependent. In order to
also provide multi-lingual support, multiple knowledge sources may be provided.
The Acoustic Model 120 may be formed by statistically modeling the various sounds that occur
in a particular language. A phoneme is assumed to be the basic unit of sound. A predefined set of such
phonemes is assumed to completely describe all sounds of a particular language. An HMM, which
encodes the relationship of the ed audio signal and the unobserved phonemes, forms the
fundamental theory for most modern speech recognition systems. A e is considered to be
ed of three states, representing the beginning, central, and trailing portions of the sound. An
HMM is ucted by concatenating these three states. A training process studies the statistical
properties of each of these states for all of the phonemes over a large collection of transcribed audio. A
relation between the textual ties and the spoken properties is thus formed. Typically, the
statistics of states may be encoded using a Gaussian mixture model (GMM). A set of these GMMs is
termed as an acoustic model. Specifically, the one bed in this ation is referred to as a
t-independent, or monophone, model. Many other model types may also be used. For example,
many modern speech recognition systems may utilize a more advanced acoustic model, which may be
context-dependent and capture the complex variations created due to the position of phonemes in
conversational speech. Each state of a phoneme is specialized to its left and right neighboring
phonemes.
The Lexicon/Pronunciation Predictor, 125, may be responsible for decomposing a word into a
sequence of phonemes. Keywords presented from the user may be in human readable form, such as
grapheme/alphabets of a particular language. However, the pattern matching algorithm may rely on a
sequence of phonemes which represent the pronunciation of the keyword. A Pronunciation Predictor
may store a mapping between commonly spoken words and their pronunciations. Once the sequence of
es is obtained, the corresponding statistical model for each of the phonemes in the ic
Model 120 may be examined. A concatenation of these statistical models may be used to perform
keyword spotting for the word of interest.
The Audio Stream (i.e., what is spoken into the system by the user) 130 may be fed into the
Front End Feature Calculator, 135, which may convert the Audio Stream 130 into a entation of the
audio stream, or a sequence of spectral features. Audio analysis may be med by segmenting the
audio signal as a sequence of short (typically 10 ms) windows and extracting spectral domain features.
The Keyword Model, 110, which may be formed by concatenating phoneme HMMs, and the
es extracted from the Audio Stream, 135, may both then be fed into a Recognition Engine for
pattern matching, 140. The task of the ition Engine 140 may be to take a set of keyword models
and search through presented audio stream to find if the words were spoken. In the multidimensional
space constructed by the feature calculator, a spoken word may become a sequence of spectral domain
feature vectors forming a trajectory in the acoustic space. Keyword spotting may now simply become a
m of computing the probability of generating the trajectory given the keyword model. This
operation may be achieved by using the well-known principle of dynamic programming, specifically the
Viterbi algorithm, which aligns the keyword model to the best segment of the audio , and results in
a match score. If the match score is significant, the keyword ng algorithm infers that the keyword
was spoken and reports a keyword spotted event.
The resulting keywords may then be reported in real-time, 145. The Report may be ted as
a start and end time of the keyword in the Audio Stream 130 with a confidence value that the keyword
was found. The primary ence value may be a function of how the keyword is spoken. For example,
in the case of multiple pronunciations of a single word, the keyword ”tomato” may be spoken as ”tuh-
mah-tow” and ”tuh-may-tow”. The primary confidence value may be lower when the word is spoken in
a less common pronunciation or when the word is not well enunciated. The specific variant of the
pronunciation that is part of a particular recognition is also displayed in the report.
As illustrated in Figure 2, a process 200 for FOM prediction is provided. The process 200 may be
ive on any or all elements of the system 100 (Figure 1).
Input is entered into a User Interface in step 205. User input may be in the form of words or
phonetic pronunciation. A User Interface is described in r detail in Figure 3 as follows. Control is
passed to operation 210 and the process 200 ues.
In step 210, the feature vector is computed for the user input. The feature vector may include
such features as the number of phonemes, the number of syllables and the number of ed vowels.
Control is passed to operation 215 and the process 200 continues.
In operation 215, the e vector is passed through the learned prediction model. A learned
tion model for FOM may be created using a phoneme recognizer, a lexicon, a morphological
analyzer, on statistics, and a keyword set containing, for example, 500 keywords. The lexicon may
be a lookup or predictive module that can t input words into a sequence of constituent
phonemes. The morphological analyzer may be another lookup or predictive module that contains
entries for ng the rules of morphology in a language. Common affixes in a language are used. For
example, common affixes in the English language may include: ”ment", ”ing”, ”tion”, and ”non”. The
phoneme confusion matrix may be computed through the creation of a phoneme recognizer. The
matrix quantitatively describes how the speech engine typically confuses sounds in the ge. This
matrix may later be used as the source for computing distances between words. With the creation of
the matrix, it is possible to determine how the speech engine sees the phonetic space, but not
necessarily what the theory of phonology expects the confusability to be. Duration statistics for
phonemes are based on the is of phonemes on a large speech corpus using the phoneme
recognizer created. The 500 word keyword set is carefully chosen to span the range of values that the
modeling features can take. Figure 5 below describes the process for choosing the keyword set in
greater detail.
In operation 220, the predicted FOM is obtained and the process ends. For example, the result
may be a FOM number output with a range of 0-100. A value approaching or equal to 0 may indicate
low accuracy or high false alarm rate while a value approaching or equal to 100 may indicate high
accuracy or ence.
Operations 205, 210 and 215 may be performed interactively in real-time as a user adds more
input.
Figure 3 is an illustration of an example Guided User ace 300 that may be used for data
input in Process 200. The User Interface 300 may contain: a Keyword field 305 and a FOM field 310,
FOM bars 315, and Keyword examples 320. Short words may have lower FOM and possibly higher false
alarm rates. Longer words may have higher FOM. A keyword such as ”Jerk” 320c may be more prone
to error than the keyword ”Screw You” 320b because the keyword ”Jerk” is used in many other ts
and has a short acoustic context to help disambiguation. For example, ”jerk” may sound similar to the
parts of ”manager”, ”integer”, or ”German”. Conversely, ”screw” is a pretty distinctive sound and is
easily recognized. The length of the bar 315 is indicative of the degree of FOM for each keyword 305.
For example, the keyword ”We Appreciate Your Business” 320d has a bar length of 98, 315d. This may
indicate that there is a higher predicted FOM for ”We Appreciate Your Business” than a word such as
”Jerk” 320c with a FOM bar length of 20, 315c. In at least one embodiment, the color of the bar may
change based on the predicted FOM in order to provide more visual feedback.
As illustrated in Figure 4, one embodiment of a system for model learning is provided and
indicated generally at 400. The system 400 may be operative on any or all ts of the system 100
(Figure 1). The basic components of the system 400 may e: a Keyword Set from the database
405; a Feature Vector Computation Module 410; izer Data 415, which may consist of a Phoneme
Confusion Matrix 420, and Duration Statistics 425; a Lexicon 430 and a Morphological er 435; a
Model Learning Module 440; and an FOM Model 445.
The Keyword Set from the database 405 may be comprised of words for which sufficient audio
recordings exist to compute tically icant accuracy numbers. The Keyword Set may be
comprised of 500 keywords, for example, that are fed into the Feature Vector Computation Module 410.
The Feature Vector Computation Module 410 may utilize data input from the Recognizer, which
may consist of a Phoneme Confusion Matrix 420 and Duration Statistics 425, and from the Lexicon 430
and Morphological Analyzer 435, to determine the feature vector of the each keyword.
The Recognizer Data 415 is provided by the ition Engine 140 (Fig 1) and is from the
recognizer output. These data may include a e Confusion Matrix 420 and Duration Statistics
425. The Phoneme Confusion Matrix 420 is ed through the on of a e recognizer.
The matrix quantitatively describes how the speech engine typically confuses sounds in the language.
The Duration Statistics 425 may be based on the analysis of es on a large speech corpus using
the phoneme izer created.
The n 430 and Morphological Analyzer 435 are language dependent. The Lexicon 430
may comprise a lookup or predictive module that can convert input words into a sequence of
constituent phonemes. The Morphological Analyzer 435 is also another lookup or predictive module
that may contain entries for the most common prefixes and suffixes in a language.
The Model Learning Module 440 may use the output from the Feature Vector Computation
Module 410 to infer a regression function from the data. The module may also adjust parameters to
optimize a cost function, which in at least one embodiment is the minimization of the absolute value of
the prediction error.
The FOM Model 445 may comprise the result of the model learning module 440 output which is
saved by the system for use at runtime to predict the FOM on user input words. This is described in
greater detail in Fig. 7 as follows.
Referring now to Figure 5, one embodiment of a process 500 for choosing the training keyword
set as used in step 405 of Figure 4 is illustrated. In at least one embodiment, this forms a key part of the
learning process as a well-chosen keyword set helps the d model generalize well to words not
seen during the supervised learning process.
A large keyword set is selected in step 505. For example, a keyword set containing a large
number of words (e.g., 200 words in one embodiment) is chosen by examining the values of similar
dictionary words and ascertaining that words cover the range of acceptable values for this feature. For
example low, medium, and high values should be represented in this keyword set. Control is passed to
operation 510 and the process 500 continues.
In operation 510, a feature is ted. As previously described, feature s may include
such features as the number of phonemes, number of syllables, number of stressed vowels, etc. This
information may be derived from the Lexicon, Morphological Analyzer, Duration Statistics and Confusion
Matrix 515. Control is passed to operation 520 and the process 500 continues.
In ion 520, the range of the feature value is d. The range of values for each feature
may vary and thus, values are examined to determine if they are lower, medium, or higher values.
As stated above, the keyword set is lly chosen to span the range of values that the
modeling features can take. Therefore, in operation 525, it is determined r or not the feature is
well represented within the key word set. If it is determined that the feature is well represented, then
control is passed to step 510 and process 500 continues. If it is determined that the feature is not well
represented, then the system control is passed to step 530 and process 500 ues.
The determination in operation 525 may be made based on any suitable criteria. For example, if
the range of the feature value is too high or too low, able words may have been chosen in the
keyword set. A keyword set with words that are too similar will have a skewed range. Where control is
passed to step 510, in the FOM algorithm later described herein, the value of i is set equal to i + 1,
which is indicative of the next feature.
In ion 530, the number of keywords may be adjusted by adding more keywords to the set.
In the FOM algorithm later described herein, the value of i is set equal to 0 which is tive of the first
feature. Control is passed operation 510 and the process 500 continues.
In at least one embodiment, this e is used to guide users in determining a good set of
keywords. Other uses may include feedback to the recognition engine and controlling the false alarm
rate. The diagram in Figure 6 exhibits the relationship n the match probability, or the ”score”, as
determined by the ition engine and the confidence values as reported by the system. By default,
the curve 605 may be used if no information about the keyword is known. If FOM is known, the
relationship may be modified by changing the operating score range of the keyword, as illustrated by
lines 610 and 615. The line 610 exhibits a low FOM keyword while the line 615 exhibits a high FOM
keyword. As the value of the score increases, so does the confidence in the match where 0.0 may be
indicative of highly confident match and a large negative value could indicate very low confidence in the
match, for example. As the Score becomes more negative, likelihood of a mismatch ses. For
example, as the Score approaches 0.0, there is a greater likelihood of a match. Thus, a Score of 0 and a
ence of 1.0 would indicate a perfect match in this illustration. In at least one ment, it is
desired to change the score range such that a chosen confidence value represents a similar score value
for words with either low or high FOM.
Figure 7 is a diagram illustrating the system behavior with varied confidence settings. The result
of changing the operating range based on FOM may be a more controlled behavior of the system. For
example, when a user registers a keyword to be spotted, an associated FOM measure is presented, such
as 70. By definition, this means the system results in 70% cy with a false alarm rate of 5 per hour.
To obtain this behavior from the system, the internal score range is modified as shown in Figure 7, such
2012/053061
that at the default confidence setting (0.5) the system produces 5 false alarms per hour and a detection
rate of 70%. If the user wishes a higher accuracy, the confidence setting may be lowered, which in turn
could possibly create a higher false alarm rate. If the user wishes lower false alarm rate, confidence
setting may be increased, thus possibly resulting in lower detection rate. By changing the internal score
range based on FOM, this behavior s consistent for all words irrespective of their FOMs.
The diagram 700 illustrates the or of the system as the confidence gs are altered.
For example, as the Confidence setting approaches 0.0, the rate of False Alarms (FA/Hr) increases and
rate of detection increases as well. sely, as the Confidence setting approaches 1.0, the rate of
false alarms decreases until it reaches a value 0.0 while the rate of detections also decreases and
approaches 0.0.
Figure 8 is a table illustrating keyword examples 800. Records 800a and 800b may contain a
d field 805, Predicted FOM field 810, Number of Phonemes field 815, Number of Stressed Vowels
field 820, Number of Syllables field 825, Duration Mean field 830, Duration Standard Deviation field 835,
Partial Dictionary Words field 840, r Dictionary Words field 845, Similar Prefix field 850, Similar
Suffix field 855, and Confusion Index field 860.
The keyword field 805 may contain the keyword example. For example, Record 800a contains
the word ”debug” and Record 800b contains the word ”interactive”.
In at least one embodiment, the Predicted FOM field 810 contains the value predicted by the
FOM expression equation:
fom : Edi-(x;- —bi)2
where i represents the index of features, x represents the i-th feature, and the equation
parameters a and b are learned values, the values of which are exhibited in Figure 9. N ents an
upper limit on a number of features used to learn the prediction. For example, N=10 may be used.
For example, Record 800a contains a FOM value of 29.6971 for keyword ’ while Record
800b contains a FOM value of 78.5823 for keyword 'interactive’.
The Number of Stressed Vowels field 820 may exhibit the number of vowels in each keyword
that emphasis is put on when the word is spoken. For example, the more vowels in a word that are
stressed, the better enunciated they are and have higher accuracies in general. The keyword 'debug’
contains 1 stressed vowel while 'interactive’ contains 2 stressed vowels as illustrated in Figure 8.
The Number of Syllables field 825 may contain the number of les within each keyword.
For example, the keyword 'debug’ has 2 syllables while the keyword 'interactive’ contains 4 syllables.
The Duration Mean field 830 may contain the duration mean value from the feature. For
example, the d 'debug’ has a duration mean of 36.6276 while the keyword 'interactive’ has a
on mean of 61.9474.
The Duration Standard ion field 835 may contain the standard deviation of the duration of
the keyword. For example, the keyword 'debug’ has a duration standard deviation value of 8.96752
while the d 'interactive’ has a duration standard deviation value of 10.5293.
The Partial Dictionary Words field 840 may n a measure of how many words in the typical
vocabulary of the language of interest that the keyword is a part of. The higher this number, the fewer
the number of words that the keyword is a part of which lly results in higher accuracy of a match.
For example, the d ’ has a value of 33.3333 in the partial dictionary words field 840 while
the keyword 'interactive’ has a value of 50.
The Similar nary Words field 845 may contain a measure of how many words in the typical
vocabulary of the language of interest that the keyword is similar to. For example, the lower this
number is, the more the number of words that the keyword is similar to and thus confusable with. A
lower accuracy in general may result. Similarity is measured using a distance metric. An example of a
distance metric can be seen with the words ”cat” and ”bat”, which have the same number of es.
Broken down into phonemes, ”cat” and ”bat” become the following:
CAT -> k ae t
BAT -> b ae t
A comparison of the words shows that they have one phoneme that differs. A simple edit
distance score of 1 results. The Confusion Matrix based edit distance could be 0.2 if the recognizer is
confused between the sounds ”k” and ”b”.
The words ”cat” and ”vacate” can be used as an example of words ning different numbers
of phonemes. The words ”cat” and ”vacate” become:
CAT -> * * k ae t
VACATE -> v ey k ey t
If it is assumed that the insertion of a phoneme costs 1 and the distance between ”ae” and ”ey”
is 0.3, then the total distance between the words is 2.3.
In another example, the distance between words that have errors can be shown below with the
words ”cat” and ”aft”:
CAT -> k ae t *
AFT -> * aef t
Errors may include insertions, deletions, and substitutions of phonemes. If it is assumed that
the insertion of a phoneme costs 1, on costs 2, and distance between phonemes ”t” and ”f” is 0.7,
then the total distance from ”cat” to ”aft” is 3.7. This ts for one ion, one deletion, and one
substitution of the phonemes.
In Figure 8, for example, keyword 'debug’ contains a value of 5 in the Similar Dictionary Words
field 845 while the keyword 'interactive’ contains a value of 33.3333.
The Similar Prefix field 850 may n a measure of how many typical prefixes the d is
confusable with. This number is provided by the morphological analyzer. A higher value in this field
indicates less similarity with common prefixes and therefore higher typical accuracy for the word. A
prefix is an affix which is placed before the root of a word. Examples are ”pre” and ”non”. For e,
the keyword 'debug’ contains a Similar Prefix value of 20 while the keyword 'interactive’ contains a
Similar Prefix value of 100.
The Similar Suffix field 855 may contain a measure of how many typical suffixes the keyword
may be confusable with. This number is provided by the morphological analyzer. A higher value in this
field indicates less rity with common prefixes and therefore higher typical accuracy for the word.
In linguistics, a suffix (which also may be referred to as a postfix or ending) is an affix which is placed
after the stem of a word. Two examples of suffixes are ”tion” and ”ous”. As illustrated in Figure 8, the
keyword 'debug’ contains a Similar Suffix value of 25 while the d 'interactive’ contains a Similar
Suffix value of 100.
The Confusion Index field 860 may contain a measure of the total sum of the ability of the
es comprising the word. If a word is comprised of several often confused phonemes such as
plosives and , it is susceptible to having a lower accuracy. For example, the d 'debug’
contains a Confusion Index value of 38.85 while the keyword 'interactive’ contains a Confusion Index
value of 61.65.
Figure 9 is a table illustrating FOM model expression. This table illustrates examples of learned
values through the previously described processes that may be used to compute the values described
for records 800a and 800b in Figure 8. These values are input into the FOM algorithm from above:
fom : Edi-(x;- —bi)2
Figure 9 may contain the following fields: Feature Name 905, Number of Phonemes polynomial
910, Number of ed Vowels mial 915, Number of Syllables polynomial 920, Duration Mean
polynomial 925, Duration Standard ion polynomial 930, Partial Dictionary Words polynomial 935,
Similar Dictionary Words polynomial 940, Similar Prefix polynomial 945, Similar Suffix polynomial 950,
and Confusion Index mial 955.
The feature name field contains the index of features heading as represented by i, and the
equation parameters a and b. Fields 910 through 955 show the learned values for each field. For
example, the Number of Phonemes polynomial 910 has values of i = 1, a = 0.1499, and b = -32.2629.
While the invention has been illustrated and bed in detail in the drawings and foregoing
description, the same is to be considered as illustrative and not restrictive in character, it being
understood that only the preferred embodiment has been shown and described and that all equivalents,
changes, and modifications that come within the spirit of the inventions as described herein and/or by
the following claims are desired to be protected.
Hence, the proper scope of the present invention should be ined only by the broadest
interpretation of the appended claims so as to encompass all such cations as well as all
relationships equivalent to those illustrated in the drawings and bed in the specification.
Claims (17)
1. A computer-implemented method for predicting speech recognition performance in a speech recognition , the system comprising a recognition engine, a database, a model learning module, and a performance prediction module, the method comprising: a. using the performance prediction module to determine at least one e vector for an input into the speech recognition system; b. creating a prediction model by: i. selecting a set of keywords; ii. computing an other feature vector of d features for each of the keywords; iii. inputting the other e vector into the model learning module, wherein the model learning module adjusts parameters to minimize a cost function; iv. saving the results from the model learning module as the prediction model for prediction of a figure of merit of the input; c. passing the at least one feature vector into the tion model; d. using the performance prediction module to apply the prediction model to predict a figure of merit for the speech recognition system, wherein the figure of merit is indicative of the accuracy of performance of the speech recognition system; e. using the performance prediction module to report the ted figure of merit for the speech recognition system performance; and f. adjusting the recognition engine based on the predicted figure of merit.
2. The method of claim 1, n the figure of merit prediction has a detection rate averaging 5 FA/Kw/Hr.
3. The method of claim 1, wherein the input comprises at least one word.
4. The method of claim 1, wherein the input comprises a phonetic pronunciation.
5. The method of claim 1, wherein the method is med in real-time as additional input is provided.
6. The method of claim 1, wherein the at least one feature vector is determined sing: a. converting the input into a sequence of phonemes; and b. performing morphological analysis of words in a language.
7. The method of claim 6, wherein the converting is performed using statistics for es and phoneme confusion matrix.
8. The method of claim 7, further comprising the step of computing the phoneme confusion matrix using a phoneme recognizer.
9. The method of claim 1, further comprising the step of automatically adjusting internal scores of the ition engine based on the prediction reported by the performance prediction module.
10. A computer system with a digital microprocessor and associated memory configured for executing software programs, the system being configured for predicting speech recognition performance, sing: means configured for using a performance tion module to determine at least one feature vector for an input into the speech recognition system; means configured to create a prediction model by: selecting a set of keywords; computing an other feature vector of desired features for each of the keywords; inputting the other feature vector into the model learning module, wherein the model learning module adjusts parameters to ze a cost function; and saving the results from the model learning module as the prediction model for prediction of a figure of merit of the input; means configured for passing the at least one e vector into the prediction model; means configured for using the performance prediction module to apply the prediction model to t a figure of merit for the speech recognition system, wherein the figure of merit is tive of the accuracy of performance of the speech recognition system; means configured for using the performance prediction module to report the predicted figure of merit for the speech recognition system performance; and means ured for adjusting the recognition engine based on the ted figure of merit.
11. The system of claim 10, wherein the figure of merit has a detection rate averaging 5 FA/Kw/Hr.
12. The system of claim 10, wherein the input comprises at least one word.
13. The system of claim 10, wherein the input comprises a phonetic pronunciation.
14. The system of claim 10, wherein the at least one feature vector is determined comprising: a. converting the input into a sequence of phonemes; and b. performing morphological analysis of words in a language.
15. The system of claim 14, wherein the converting is performed using statistics for phonemes and phoneme confusion matrix.
16. The system of claim 15, further comprising: means for computing the phoneme confusion matrix using a phoneme recognizer.
17. The system of claim 10, further sing: means for automatically adjusting internal scores of the ition engine based on the prediction reported by the performance prediction module. I\11236045.1
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/US2012/053061 WO2014035394A1 (en) | 2012-08-30 | 2012-08-30 | Method and system for predicting speech recognition performance using accuracy scores |
Publications (2)
Publication Number | Publication Date |
---|---|
NZ705071A NZ705071A (en) | 2017-01-27 |
NZ705071B2 true NZ705071B2 (en) | 2017-04-28 |
Family
ID=
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10360898B2 (en) | Method and system for predicting speech recognition performance using accuracy scores | |
KR101056080B1 (en) | Phoneme-based speech recognition system and method | |
US5623609A (en) | Computer system and computer-implemented process for phonology-based automatic speech recognition | |
US20180114525A1 (en) | Method and system for acoustic data selection for training the parameters of an acoustic model | |
AU2012388796B2 (en) | Method and system for predicting speech recognition performance using accuracy scores | |
US20180137109A1 (en) | Methodology for automatic multilingual speech recognition | |
CA2882569C (en) | Method and system for selectively biased linear discriminant analysis in automatic speech recognition systems | |
US20140025379A1 (en) | Method and System for Real-Time Keyword Spotting for Speech Analytics | |
Sonderegger et al. | Automatic measurement of voice onset time using discriminative structured prediction | |
US20140067394A1 (en) | System and method for decoding speech | |
US20130289987A1 (en) | Negative Example (Anti-Word) Based Performance Improvement For Speech Recognition | |
Metze | Articulatory features for conversational speech recognition | |
Azim et al. | Large vocabulary Arabic continuous speech recognition using tied states acoustic models | |
JP4764203B2 (en) | Speech recognition apparatus and speech recognition program | |
Rouas et al. | Language and variety verification on broadcast news for Portuguese | |
NZ705071B2 (en) | Method and system for predicting speech recognition performance using accuracy scores | |
JP6199994B2 (en) | False alarm reduction in speech recognition systems using contextual information | |
Wang | A study of meta-linguistic features in spontaneous speech processing | |
Xie | Speech Analyzer in an ICAI System for TESOL | |
NZ719961B2 (en) | Method and system for real-time keyword spotting for speech analytics | |
Cernak | DASR: A Diagnostic Tool For Automatic Speech Recognition | |
Mitter | A Linguistic Feature Representation of the Speech Waveform | |
NZ704832B2 (en) | Method and system for real-time keyword spotting for speech analytics |