Title: Method for robust voice recognition by analyzing redundant features of source signal
Abstract: A method for processing digitized speech signals by analyzing redundant features to provide more robust voice recognition. A primary transformation is applied to a source speech signal to extract primary features therefrom. Each of at least one secondary transformation is applied to the source speech signal or extracted primary features to yield at least one set of secondary features statistically dependant on the primary features. At least one predetermined function is then applied to combine the primary features with the secondary features. A recognition answer is generated by pattern matching this combination against predetermined voice recognition templates.
Patent Number: 6,957,183 Issued on 10/18/2005 to Malayath, et al.
| Inventors:
|
Malayath; Narendranath (San Diego, CA);
Garudadri; Harinath (San Diego, CA)
|
| Assignee:
|
Qualcomm Inc. (San Diego, CA)
|
| Appl. No.:
|
104178 |
| Filed:
|
March 20, 2002 |
| Current U.S. Class: |
704/246; 704/236; 704/243 |
| Intern'l Class: |
G10L 015/00 |
| Field of Search: |
704/236,243,246
|
References Cited [Referenced By]
U.S. Patent Documents
| 5381512 | Jan., 1995 | Holton et al.
| |
| 5414796 | May., 1995 | Jacobs et al.
| |
| 5450522 | Sep., 1995 | Hermansky et al.
| |
| 5517558 | May., 1996 | Schalk.
| |
| 5655057 | Aug., 1997 | Takagi.
| |
| 5764853 | Jun., 1998 | Watari et al.
| |
| 6671404 | Dec., 2003 | Kawatani et al.
| |
| Foreign Patent Documents |
| 0192898 | Sep., 1986 | EP.
| |
Other References
Gemello, R. et al: "Multi-Source Nerual Networks for Speec Recognition", Neural
Networks, 1999. IEEE, Jul. 10, 1999, pp. 2946-1949.
Finke, M. et al: The Karlsruhe-Verbmobil Speech Recognition Engine:. Acoustics,
Speech, and Signal Processing, 1997. IEEE International Conference on Munich, Germany,
Apr. 21, 1997, pp. 83-86.
Kuo-Hwei You, et al: "Robust Features Derived from Temporal Trajectory Filtering
for Speech Recognition Under The Corruption of Additive and Convolutional Noises",
Acoustics, Speech and Signal Processing, 1998. IEEE International Conference On
Seattle, WA, May 12, 1998. pp. 577-580.
Sharma, S. et al: "Feature Extraction Using Non-linear Transformation for Robust
Speech Recognition On the Aurora Database", ICASSP 2000, vol. 2, Jun. 5, 2000,
pp. 1117-1120.
|
Primary Examiner: Abebe; Daniel
Attorney, Agent or Firm: Wadsworth; Philip, Brown; Charles D, Pappas; George C.
Claims
1. A method of processing speech signals comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
2. The method of claim 1, where the primary transformation comprises a spectral transformation.
3. The method of claim 1, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
4. The method of claim 1, where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
5. The method of claim 1, further comprising:
forming voice recognition templates by performing each of the applying and generating
operations to predetermined training signals.
6. The method of claim 5, further comprising:
forming the voice recognition templates by performing each of the applying and
generating operations to predetermined training signals.
7. A method of processing speech signals comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
8. The method of claim 7, where each of the separately modifying operations comprises
at least one of the following:
scaling, power change, self-multiplying, exponentiation.
9. The method of claim 7, where the primary transformation comprises a spectral transformation.
10. The method of claim 7, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
11. The method of claim 7, where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
12. A signal-bearing medium tangibly embodying a program of machine-readable
instructions executable by a digital processing apparatus to perform operations
for processing speech signals, the operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependent
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
13. The medium of claim 12, where the primary transformation comprises a spectral transformation.
14. The medium of claim 12, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
15. The medium of claim 12 where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
16. The medium of claim 12, further comprising:
forming the voice recognition templates by performing each of the applying and
generating operations to predetermined training signals.
17. A signal-bearing medium tangibly embodying a program of machine-readable
instructions executable by a digital processing apparatus to perform operations
for processing speech signals, the operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependent
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates;
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
18. The medium of claim 17, where each of the separately modifying operations
comprises at least one of the following:
scaling, power change, self-multiplication, exponentiation.
19. The medium of claim 17, where the primary transformation comprises a spectral transformation.
20. The medium of claim 17, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
21. The medium of claim 17 where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
22. The medium of claim 17, further comprising forming the voice recognition
templates by performing each of the applying and generating operations to predetermined
training signals.
23. Circuitry of multiple interconnected electrically conductive elements configured
to perform operations to process speech signals, the operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
24. The circuitry of claim 23, where the primary transformation comprises a spectral transformation.
25. The circuitry of claim 23, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
26. The circuitry of claim 23, where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
27. The circuitry of claim 23, further comprising:
forming the voice recognition templates by performing each of the applying and
generating operations to predetermined training signals.
28. Circuitry of multiple interconnected electrically conductive elements configured
to perform operations to process speech signals, the operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features;
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates; and
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
29. The circuitry of claim 28, where each of the separately modifying operations
comprises at least one of the following:
scaling, power change, self-multiplication, exponentiation.
30. The circuitry of claim 28, where the primary transformation comprises a spectral transformation.
31. The circuitry of claim 28, where the primary transformation comprises production
of a time-frequency representation of the input speech signal.
32. The circuitry of claim 28, where the primary transformation comprises a spectral
transformation and the secondary transformation comprises a cepstral transformation.
33. The circuitry of claim 28, further comprising forming tire voice recognition
templates by performing each of the applying and generating operations to predetermined
training signals.
34. A voice recognition system comprising:
a primary feature extractor applying a primary function to extract primary features
from a digital input speech signal;
at least one secondary transformation module each producing secondary features
statistically dependent on the primary features by applying a secondary function
to an input comprising one of the following: the input speech signal, the primary
features;
a feature combination module coupled to the primary feature extractor and each
of the secondary transformation modules to apply one or more predetermined functions
to combine the primary features with the secondary features forming a combined
signal; and
a statistical modeling engine, coupled to the feature combination module to generate
a recognition answer by pattern matching the combined signal against predetermined
voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
35. A voice recognition system comprising:
primary feature extractor means for applying a primary function to extract primary
features from a digital input speech signal;
secondary transformation means for producing secondary features statistically
dependent on the primary features by applying at least one secondary function to
an input comprising one of the following: the input speech signal, the primary
features;
feature combination means for applying one or more predetermined functions to
combine the primary features with the secondary features forming a combined signal;
and
statistical modeling means for generating a recognition answer by pattern matching
the combined features against predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear disriminant
analysis, principal component transfer, and concatenation.
36. A wireless communications device having:
a transceiver coupled to an antenna;
a speaker;
a microphone;
a user interface;
a manager coupled to components including the transceiver, speaker, microphone,
and user interface to manage operation of the components, the manager including
a voice recognition system configured to perform operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
37. A wireless communications device having:
a transceiver coupled to an antenna,
a speaker;
a microphone;
a user interface;
means for managing operation of the transceiver, speaker, microphone, and user
interface, and for performing voice recognition by:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
the at least one predetermined function utilizing at least one of linear discriminant
analysis, principal component transfer, and concatenation.
38. A voice recognition system comprising:
a primary feature extractor applying a primary function to extract primary features
from a digital input speech signal;
at least one secondary transformation module each producing secondary features
statistically dependent on the primary features by applying a secondary function
to an input comprising one of the following: the input speech signal, the primary
features;
a feature combination module coupled to the primary feature extractor and each
of the secondary transformation modules to apply one or more predetermined functions
to combine the primary features with the secondary features fanning a combined
signal;
a statistical modeling engine, coupled to the feature combination module to generate
a recognition answer by pattern matching the combined signal against predetermined
voice recognition templates;
the predetermined function utilizing at least one of the following to combine
the primary features and the secondary features: linear disriminant analysis, principal
component transfer, concatenation;
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
39. A voice recognition system comprising:
primary feature extractor means for applying a primary function to extract primary
features front a digital input speech signal;
secondary transformation means for producing secondary features statistically
dependent on the primary features by applying at least one secondary function to
an input comprising one of the following: the input speech signal, the primary
features;
feature combination means for applying one or more predetermined functions to
combine the primary features with the secondary features forming a combined signal;
statistical modeling means for generating a recognition answer by pattern matching
the combined features against predetermined voice recognition templates;
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
40. A wireless communications device having:
a transceiver coupled to an antenna;
a speaker;
a microphone;
a user interface;
a manager coupled to components including the transceiver, speaker, microphone,
and the user interface to manage operation of the components, and a voice recognition
system configured to perform operations comprising:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features;
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates;
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates,
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
41. A wireless communications device having:
a transceiver coupled to an antenna;
a speaker;
a microphone;
a user interface;
means for managing operation of the transceiver, speaker, microphone, and user
interface, and for performing voice recognition by:
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features;
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates;
applying a primary transformation to a digital input speech signal to extract
primary features therefrom;
applying each of at least one secondary transformation to one of the input speech
signal and the primary features to yield secondary features statistically dependant
on the primary features;
applying at least one predetermined function to form a combined signal comprising
a combination of the primary features with the secondary features; and
generating a recognition answer by pattern matching the combined signal against
predetermined voice recognition templates;
where the at least one predetermined function
utilizes at least one of linear discriminant analysis, principal component transfer,
and concatenation; and
separately modifies at least one of the primary features and the secondary features,
the at least one predetermined function being used to form a combined signal comprising
a combination of the primary features including any modifications with the secondary
features including any modifications.
Description
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates to voice recognition techniques and circuits, and
more particularly, to a system for more accurate and noise-tolerant robust voice
recognition by analyzing redundant features of a source signal.
2. Description of the Related Art
Various signal processing techniques have been developed for analyzing digitized
speech signals in order to recognize the underlying content of such speech. Once
recognized, this content can then be used to control a handheld telephone, computer,
household appliance, or other device. Some such known techniques employ short-time
Fourier spectra or "spectrogram" of a speech signal, which are computed using windowed
Fourier transforms as explained more fully in Rabiner et al.,
Fundamentals of
Speech Recognition, the entirety of which is incorporated herein by reference.
FIG. 1 shows one known spectral feature extractor
100 for spectral analysis,
which includes stages of windowing
102, FFT
104, MEL/BARK filtering
106, Log
108, and RASTA filtering
110. A digitized input speech
signal
101 is fed into the windowing stage
102, which divides the
input signal into smaller sized segments of appropriate duration, such as 20 milliseconds.
The FFT stage
104 performs a Fast Fourier Transform to windowed segments
output by the stage
102. The MEL/BARK stage
106 performs warping
of the linear frequency scale to a different scale, so that the resolution for
lower frequencies is greater than that for higher frequencies. The resolution on
the frequency scale becomes progressively coarser from low frequencies to high
frequencies in the hearing range. MEL scale and BARK scale are two known transformations
that result in the above frequency warping. These two (and some variations) are
commonly used in speech recognition. The Log stage
108 takes the Logarithm
of the input number, and more particularly, the log of each MEL/BARK transformed
spectral value that has been computed. The foregoing stages
102-
108
are described in various known publications, with one example being the above-cited
text Rabiner et al.,
Fundamentals of Speech Recognition.
The RASTA stage
110 serves to filter the output of the Log stage
108
by a predefined bandpass filter. For example, if there are sixteen BARK numbers,
there will be sixteen filters operating on each of the bark bands. The RASTA stage
110 may be implemented by any known RASTA processing technique, with one
example being described in U.S. Pat. No. 5,450,522 entitled "Auditory Model for
Parameterization of Speech" to Hermansky et al., the entirety of which is incorporated herein.
The output of the spectral feature extractor
100 comprises spectral output
signals
111, which are thereafter processed by various subsequent techniques
(not shown) to yield a "recognition answer" that gives the predicted content of
the input speech signal. Recognition answers based on such spectral output signals
111 provide decent accuracy in low noise environments. Advantageously, degradation
of their accuracy occurs slowly with decreasing signal-to-noise ratios. Spectral
output signals can be further processed in various ways. For instance, one approach
further processes the spectral output signals
111 by a cepstral transformation
112 to yield cepstral output signals
114. One type of cepstral transformation
112, for example, utilizes a discrete cosine transform (DCT) followed by
a dimensionality reduction. Broadly, "cepstrum" is explained as the inverse Fourier
transform of the logarithm of the power spectrum of a signal, as further discussed
in the following references, hereby incorporated by reference in their entirety:
A. V. Oppenheim and R. W. Schafer,
Discrete-
Time Signal Processing,
J. R. Deller, Jr., J. G. Proakis and J. H. L. Hansen,
Discrete-
Time Processing
of Speech Signals, and L. R. Rabiner and R. W. Schafer,
Digital Processing
of Speech Signals.
In systems where the cepstrum
114 is calculated, the cepstrum (rather
than
the spectrum
111) is processed by statistical modeling techniques to yield
a recognition answer. One benefit of basing recognition answers upon cepstral output
signals
114 is that they provide more accurate voice recognition at low
levels of noise. However, as noise increases, the error rate increases rapidly
for these systems. Therefore, neither spectral nor cepstral voice recognition systems
are entirely adequate for applications that could potentially encounter a wide
range of noise levels.
SUMMARY OF THE INVENTION
Broadly, the present invention concerns a method for processing digitized
speech signals for voice recognition. Unlike conventional approaches, which seek
compactness and simplicity of operation by removing redundant features of input
speech prior to recognition, the present invention purposefully retains and analyzes
redundant features of a source signal in order to perform voice recognition accurately
in a variety of acoustic environments. A primary transformation is applied to a
digitized source speech signal to extract primary features therefrom. One example
is a spectral transformation applied to extract spectral features. Each of at least
one secondary transformation is applied to the source speech signal or extracted
spectral features to yield at least one set of secondary features. Each secondary
transformation is designed to yield data containing some information that is already
present in the extracted primary features. At least one predetermined function
is then utilized to combine the primary features with the secondary features. A
recognition answer is generated by pattern matching this combination against a
predetermined set of voice recognition templates.
The invention affords its users with a number of distinct advantages. Chiefly,
the invention provides accurate voice recognition with increasing levels of noise
without sacrificing performance in low noise environments. This differs from conventional
cepstral voice recognition systems, where performance rapidly drops with increasing
noise, and also differs from conventional spectral voice recognition systems, where
performance degrades more slowly with increasing noise with the tradeoff of some
performance in low noise environments. The invention also provides a number of
other advantages and benefits, which should be apparent from the following description
of the invention.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a block diagram of the hardware components and interconnections of
known spectral and cepstral processing systems.
FIG. 2 is a block diagram of the hardware components and interconnections of
one embodiment of voice recognition system.
FIG. 3 is a block diagram of a digital data processing machine.
FIG. 4 shows an exemplary signal-bearing medium.
FIG. 5 shows an exemplary wireless telephone.
FIG. 6 is a flowchart of an operational sequence for voice recognition utilizing
redundant features of speech.
DETAILED DESCRIPTION
The nature, objectives, and advantages of the invention will become more apparent
to those skilled in the art after considering the following detailed description
in connection with the accompanying drawings.
Hardware Components & Interconnections
Overall Structure
One aspect of the invention concerns a voice recognition system, which may be
embodied by various hardware components and interconnections, with one example
being described by the voice recognition system 200 as shown in FIG. 2.
The system 200 serves to process an input speech signal 201, which
comprises a digitized speech signal. Optionally, the system 200 may include
a digital-to-analog converter (DAC) to digitize analog input speech signals. The
input speech signal 201 is fed to a spectral feature extractor 202,
which comprises circuitry for extracting spectral features from the input speech
signal 201. As one example, the extractor 202 may be embodied by
components similar to the spectral feature extractor 100 shown in FIG. 1.
Output 202
a of the extractor 202 comprises one or more
spectral output signals, such as a spectral feature vector. The output 202
a
is directed to a secondary transformation module 204. The secondary
transformation module 204 applies a predetermined function in order to provide
a secondary output 204
a. In the illustrated example, the module 204
applies a cepstral transformation, in which case the secondary output 204
a
comprises a set of cepstral vectors. The module 204's secondary feature
signals 204
a are directed to a feature combination module 206.
As an alternative to the foregoing embodiment, where the secondary transformation
module 204 receives output 202
a of the extractor 202,
the module 204 may instead receive the original input speech signal 201.
In still another embodiment, there may be multiple secondary transformation modules
204, each applying a different function to the input speech signal 201
or spectral output 202
a.
In any case, the foregoing description of spectral and cepstral features is merely
exemplary, and the scope of the present disclosure nonetheless contemplates a variety
of different primary and secondary transformations (not necessarily spectrum and
cepstrum). As an example, the spectral feature extractor 202 and its spectral
features 202
a may be represented by any appropriate primary feature
extractor 202 performing a first or higher order transformation to create
a time-frequency representation of the input speech signal. Some examples include
a spectral transformation, wavelet transformation, modulation spectrum, cumulants, etc.
Each secondary transformation may be implemented by a discrete cosine transform
(producing cepstrum), principal component transform, or other projection of the
primary features into another known or user-defined space. The secondary features
may be produced mathematically, empirically, or by another means. In any case,
each set of secondary features is "statistically dependent" upon the primary features,
meaning that the secondary features are related to the primary features by mathematical
function. In other words, it is possible to derive the secondary features by applying
a mathematical function to the primary features. The secondary features, for instance,
may be related from the primary features by correlation (including but not limited
to addition, subtraction, multiplication by coefficients and adding, or another
combination), nonlinear processing, or another technique. The primary features
may, or may not, be mathematically reproducible from the secondary features. As
one example, then, ceptstral features of the module 204 are statistically
dependent on the spectral features of the extractor 202, and in fact, the
illustrated cepstral vectors are derived by mathematically transforming the spectral
vectors using cepstrum. Hence, outputs 202
a, 204
a contain
redundant information regarding the input speech signal 201.
Unlike conventional voice recognition systems, where spectral or other primary
output signals are not used except for further downstream (serial) processing,
the spectral output signals 202
a are separately analyzed apart from
their use in the secondary transformation 204. Namely, the output signals
202
a are fed directly to the feature combination module 206
for analysis.
The module 206 combines the spectral output signals 202
a and
the secondary output signals 204
a. This combination may occur in
various ways. As one example, the signals 202
a, 204
a may
be combined by concatenation, linear discriminate analysis (LDA), principal component
transform (PCT), or another function applied to both outputs 202
a,
204
a together. In a different embodiment, the module 206 may
include processing subcomponents 206
a, 206
b that separately
modify one or both of the signals 202
a, 204
a after
which the module 206 combines the signals 202
a, 204
a
as modified. If more than one secondary transformation 204 is performed,
the module 206 may include an equal number of components 206
b.
Each subcomponent 206
a, 206
b may perform various functions,
such as scaling (i.e., multiplying) its input signal by a fixed or variable coefficient,
changing the exponent of the input signal, multiplying the input signal by itself
one or more times (self-multiplying), or another one or more linear or nonlinear
processes. In this embodiment, the modified signals are then combined using one
of the foregoing techniques, e.g., concatenation, LDA, PCT, etc.
As mentioned above, there may be multiple secondary transformation modules 204,
each of which receives an input signal such as the input signal 201 or the
output 202
a of the extractor 202. In such embodiment, the
feature combination module 206 still operates to combine spectral feature
signals with secondary transformation output signals, although the secondary transformation
output signals come from multiple different modules 204.
The feature combination module 206 feeds its output to a statistical modeling
engine 208, also referred to as a "recognizer" or a "pattern matching" unit.
The engine 208, which may be implemented by a number of known techniques,
produces an output comprising a recognition answer 210. The recognition
answer 210 constitutes the system 200's estimation of the meaning
of the input speech signal 201. The engine 208 includes a training
input 208
a for receiving input that trains the engine 208
to recognize certain model or sample speech patterns. As one particular example,
the engine 208 may cross-reference the output of the module 206 in
a lookup table to obtain the represented meaning of the input speech signal 201,
i.e., its "answer."
Exemplary Digital Data Processing Apparatus
As mentioned above, data processing entities such as the extractor 202,
secondary transformation module(s) 204, feature combination module 206,
statistical modeling engine 208, and the like may be implemented in various
forms. As one example, each of these components (or two or more components collectively)
may be implemented by a digital data processing apparatus, as exemplified by the
hardware components and interconnections of the digital data processing apparatus
300 of FIG. 3.
The apparatus 300 includes a processor 302, such as a microprocessor,
personal computer, workstation, or other processing machine, coupled to storage
304. In the present example, the storage 304 includes a fast-access
storage 306, as well as nonvolatile storage 308. One example of the
fast-access storage 306 is random access memory ("RAM"), used to store the
programming instructions executed by the processor 302. The nonvolatile
storage 308 may comprise, for example, battery backup RAM, EEPROM, one or
more magnetic data storage disks such as a "hard drive", a tape drive, or any other
suitable storage device. The apparatus 300 also includes an input/output
310, such as a line, bus, cable, electromagnetic link, or other means for
the processor 302 to exchange data with other hardware external to the apparatus 300.
Despite the specific foregoing description, ordinarily skilled artisans (having
the benefit of this disclosure) will recognize that the apparatus discussed above
may be implemented in a machine of different construction, without departing from
the scope of the invention. As a specific example, one of the components 306,
308 may be eliminated; furthermore, the storage 304, 306,
and/or 308 may be provided on-board the processor 302, or even provided
externally to the apparatus 300.
Logic Circuitry
In contrast to the digital data processing apparatus discussed above, a different
embodiment of the invention uses logic circuitry instead of computer-executed instructions
to implement one or all components of the system 200. Depending upon the
particular requirements of the application in the areas of speed, expense, tooling
costs, and the like, this logic may be implemented by constructing an application-specific
integrated circuit ("ASIC") having thousands of tiny integrated transistors. Such
an ASIC may be implemented with CMOS, TTL, VLSI, or another suitable construction.
Other alternatives include a digital signal processing chip ("DSP"), discrete circuitry
(such as resistors, capacitors, diodes, inductors, and transistors), field programmable
gate array ("FPGA"), programmable logic array ("PLA"), and the like.
Wireless Telephone Application
In one exemplary application, the voice recognition system 200 may be implemented
in a wireless telephone 500 (FIG. 5), along with other circuitry
known in the art of wireless telephony. The telephone 500 includes a speaker
508, user interface 510, microphone 514, transceiver 504,
antenna 506 , and manager 502. The manger 502, which may be
implemented by circuitry such as that discussed above in conjunction with FIGS.
3-4, manages operation and signal routing between the components 504, 508,
510, and 514. The manager 502 includes a voice recognition
module 502
a, embodied by the system 200, and serving to perform
a function such a decoding speech commands of a human operator of the telephone
500 regarding dialing, call management, etc.
Operation
Having described the structural features of the present invention, the operational
aspect of the present invention will now be described. Unlike conventional approaches,
which seek compactness and simplicity of operation by removing redundant features
of input speech prior to analysis, the present invention purposefully analyzes
redundant features of a source signal in order to perform voice recognition accurately
in a variety of acoustic environments.
Signal-Bearing Media
Wherever the functionality of the invention is implemented using one or
more machine-executed program sequences, these sequences may be embodied in various
forms of signal-bearing media. In the context of FIG. 3, such a signal-bearing
media may comprise, for example, the storage 304 or another signal-bearing
media, such as a magnetic data storage diskette 400 (FIG. 4), directly
or indirectly accessible by a processor 302. Whether contained in the storage
306, diskette 400, or elsewhere, the instructions may be stored on
a variety of machine-readable data storage media. Some examples include direct
access storage (e.g., a conventional "hard drive", redundant array of inexpensive
disks ("RAID"), or another direct access storage device ("DASD")), serial-access
storage such as magnetic or optical tape, electronic non-volatile memory (e.g.,
ROM, EPROM, or EEPROM), battery backup RAM, optical storage (e.g., CD-ROM, WORM,
DVD, digital optical tape), paper "punch" cards, or other suitable signal-bearing
media including analog or digital transmission media and analog and communication
links and wireless communications. In an illustrative embodiment of the invention,
the machine-readable instructions may comprise software object code, compiled from
a language such as assembly language, C, etc.
Logic Circuitry
In contrast to the signal-bearing medium discussed above, some or all of the
invention's
functionality may be implemented using logic circuitry, instead of using instruction
processing machines. Such logic circuitry is therefore configured to perform operations
to carry out the method of the invention. The logic circuitry may be implemented
using many different types of circuitry, as discussed above.
Overall Sequence of Operation
FIG. 6 shows a sequence 600 to illustrate an exemplary voice recognition
process. Broadly, this sequence serves to process digitized speech signals by analyzing
redundant features to provide more noise robust voice recognition. For ease of
explanation, but without any intended limitation, the example of FIG. 6 is described
in the context of the voice recognition system 200 described above.
The sequence is initiated in step 602, where training of the system 200
is conducted. In the training step 602, an input speech signal 201
("training speech") of desired characteristics is fed to the spectral feature extractor
202, and steps 604, 606, 608 (discussed below) are
performed utilizing this input speech signal. For instance, the training speech
may comprise model speech (for a speaker independent system) or a particular person's
sample speech (for a speaker dependent system). Output of the feature combination
module 206 forms a voice recognition template, which is associated with
the training speech by preparing a training input 208
a (signifying
the content of the training speech), supplying the input 208
a to
the statistical modeling engine 208, and instructing the engine 208
to generate the desired recognition answer in the future whenever the system 200
encounters speech similar to the training speech. For instance, if the training
speech signal 201 comprises the word "lesson," then the text "lesson" is
fed to the engine 208 in association with the training speech signal 201.
Further discussion of training is available from many sources, as many different
techniques for training voice recognition systems are well known in the art. Some
exemplary training schemes are discussed in the following materials, each incorporated
herein by reference in its entirety: (1) U.S. patent application Ser. No. 09/248,513
entitled "Voice Recognition Rejection Scheme," filed Feb. 8, 1999, (2) U.S. patent
application Ser. No. 09/255,891 entitled "System and Method for Segmentation and
Recognition of Speech Signals," filed Jan. 4, 1999, and (3) U.S. patent application
Ser. No. 09/615,572 entitled "Method and Apparatus for Constructing Voice Templates
for a Speaker-independent Voice Recognition System," filed Jul. 13, 2000. Additional
training may be conducted at later times (not shown), as required or desired.
After some training 602 is concluded, a non-training input speech signal
201 is input to the spectral feature extractor 202 (step 604).
This signal 201 is that for which voice recognition is desired. Also in
step 604, the extractor 202 performs a spectral transformation upon
the input speech signal 201, which extracts spectral features from the input
speech signal 201 and provides the extracted spectral features as the output
202
a. One exemplary spectral extraction technique includes linear
predictive coding (LPC), which is described in U.S. Pat. No. 5,414,796, entitled
"Variable Rate Encoder," fully incorporated herein by reference, and the above-cited
reference of Rabiner et al.,
Digital Processing of Speech Signals.
In step 606, the secondary transformation module 204 applies its
secondary transformation to the spectral output 202
a. In the illustrated
example, the module 204 applies a cepstral transformation to the output
202
a. Optionally, step 606 may also perform one or more additional
secondary transformations (ceptstral or other types) in parallel with the illustrated
transformation 204. As mentioned above, all secondary transformations are
statistically dependent to the spectral transformation of step 604, meaning
that that there is some common information carried by outputs 202
a and
204
b.
In step 608, the module 206 combines the spectral output 202
a
with the secondary output(s) 204
a from the module 204
and any other modules of secondary transformation, if implemented. For ease of
discussion, the present example is limited to the case where a single secondary
transformation is used. As mentioned above, the combination of step 608
may occur in various ways. For instance, the module 206 may combine the
signals 202
a, 204
a by concatenation, LDA, PCT, or another
function. In a different embodiment, the processing subcomponents 206
a,
206
b first modify one or more of the signals 202
a,
204
a separately, after which the module 206 combines the modified
signals. Each subcomponent 206
a, 206
b may perform a
function such as scaling (i.e., multiplying) the input signal by a fixed or variable
coefficient, changing the exponent of the input signal, multiplying the input signal
by itself one or more times, or another one or more nonlinear or linear processes.
In this embodiment, the signals as modified by 206
a, 206
b
are then combined by one of the foregoing techniques, e.g., concatenation,
LDA, PCT, etc.
In step 610, the statistical modeling engine 208 generates a recognition
answer, representing the estimated content of the input speech signal 201.
Step 610 is performed by pattern matching the signal from the module 206
against the predetermined set of voice recognition templates that were prepared
in step 602.
Other Embodiments
While the foregoing disclosure shows a number of illustrative embodiments of
the invention, it will be apparent to those skilled in the art that various changes
and modifications can be made herein without departing from the scope of the invention
as defined by the appended claims. Furthermore, although elements of the invention
may be described or claimed in the singular, the plural is contemplated unless
limitation to the singular is explicitly stated. Additionally, ordinarily skilled
artisans will recognize that operational sequences must be set forth in some specific
order for the purpose of explanation and claiming, but the present invention contemplates
various changes beyond such specific order.
*
