`
`[19]
`
`[11] Patent Number:
`
`5,956,683
`
`Jacobs et al.
`
`[45] Date of Patent:
`
`*Sep. 21, 1999
`
`US005956683A
`
`................................ .. 381/42
`4/1991 Liu et al.
`8/1991 Bethards . . . . . .
`. . . . .. 395/2.1
`9/1991 Smith et al.
`395/2.84
`
`
`
`5,012,518
`5,040,212
`5,045,082
`
`[54] DISTRIBUTED VOICE RECOGNITION
`SYSTEM
`
`4/1992 Katayama et al.
`.................... .. 395/600
`(S}(:)l1):11h:tr aelt. a.1........................... 39358/$223
`1/1994 K n"
`1
`1.
`395/275
`5:280:585
`6/1994 Aillcrnlseteaf ..
`395/700
`5,321,840
`6/1994 Black et al.
`.
`395/600
`5,325,524
`12/1994 Reed et al.
`............................. 395/2.84
`5,371,901
`FOREIGN PATENT DOCUMENTS
`0108354
`5/1984 European Pat on
`0177405
`4/1986 European Pat. Off.
`0534410
`3/1993 European Pat. Off.
`
`
`
`G06F 15/16
`H04M 1/27
`
`........ .. G10L 5/00
`
`5,109,509
`
`[75]
`
`Inventors: Paul E_ Jacobs; Chienchung Chang,
`both of San Diego, Calif.
`.
`[73] Aeelgneee QUALCOMM Ineenpennten’ San
`D1eg°> Cahf
`This patent issued on a Continued pr0S_
`ecution application filed under 37 CFR
`1'5e(e)’ and is Snnjeet te tne twenty year
`Patent
`term P‘°"i5i°“5 Of 35 U~S~C~
`154(a)(2)~
`
`[*1 Notice:
`
`[21] Appl. No.: 08/627,333
`
`[22]
`
`Filed:
`
`Apt 4, 1996
`
`Related U_S_ Application Data
`[63] Continuation of application No. 08/534,080, Sep. 21, 1995,
`which is a continuation of application No. 08/173,247, Dec.
`22, 1993.
`Int. Cl.5 ........................................................ G10L 3/00
`[51]
`[52] U.S. Cl.
`........................ .. 704/275; 704/201; 704/251;
`704/270
`
`[58] Field of Search .................................. .. 395/209, 704,
`395/275’ 270’ 201’ 251’ 455’ 411’ 412’
`413
`
`[56]
`
`References Cited
`
`U.S. PATENT DOCUMENTS
`1/1986 Vensko et al.
`.......................... .. 381/43
`10/1990 Takahashi
`
`.......................... .. 381/43
`2/1991 Garrett et al.
`
`4,567,606
`4,961,229
`4,991,217
`
`OTHER PUBLICATIONS
`
`International Search Report dated Apr. 7, 1995.
`Farvardin, et al. “Efficient Encoding of Speech LSP Param-
`estgegrs Us1p§8T'l11:1D1screte Cosine Transformation”, ICASSP
`7
`_
`‘
`.
`.
`.
`Primary Exam”:‘er—DaY1d R‘ Hndspeth .
`Assistant Exam;/1er—M1chael N. Opsasnick
`‘(‘f;f£::¥S§‘af1e’g;1g1ei:hF‘””_Rneeen B" Mlnen Llnn L"
`’
`[57]
`ABSTRACT
`
`a feature extraction
`A Voice reco nmon S Stem havin
`apparatus is logcated in 21’ remote statigon The feature extrac
`tion apparatus extracts features from an input speech frame
`and ‘then provides the extracted features to a central pro-
`cessing station. In the central processing station, the features
`are provided to a Word decoder which determines the syntax
`ef tne Input epeeen name‘
`
`30 Claims, 3 Drawing Sheets
`
`24
`
`FEATURE
`
`TRANSMITTER
`
`EXTRACTION
`RECEIVER
`
`38
`
`CONTROL
`
`30
`
`26
`
`
`
`ESTIMATED
`WORDS
`
`TRANSMITTED
`FEATURES
`
`1111107 1 1 1 1 1 1 1 1 1 ' 1 1 1'
`49
`
`CONTROL
`
`48
`
`46
`
`I
`
`WORD
`DECODER
`
`RECEIVER
`
`50
`
`TRANSMITTER
`
`Cardiocom Ex. 1005
`
`Cardiocom Ex. 1005
`
`
`
`U.S. Patent
`
`Sep.21, 1999
`
`Sheet 1 013
`
`5,956,683
`
`PRIOR ART
`
`(2
`
`SPEECH
`SIGNALS
`
`PREEMPHASIS
`
`K-4
`
`
`
`
`/- 8
`
`LPC ANALYSIS
`
`10
`
` ACOUSTIC
`
`PATTERN
`MATCHING
`
`ESTIMATED
`WORD
`STRING
`
`FIG. 1
`
`SSIE]i‘]Ef5S
`
`52
`
`POTENTIAL
`
`I
`FEATURES
`I
`
`
`L
`
`FEATURE EXTRACTION
`
`I
`
`ESTIMATED
`WORD
`STRING
`
`64 ACOUSTIC
`FEATURES
`
`
`WORD ESTIMATION
`
`,
`
`FIG. 3
`
`
`
`U.S. Patent
`
`Sep.21, 1999
`
`Sheet 2 of3
`
`5,956,683
`
`3»aF
`
`QmF<_>:emm
`
`mmmoz,
`
`Qm:%S2mZ<MH
`
`mm~5e<mE
`
`8mme:2mz<~:L’E
`
`mmE<mE
`
`ZO:LU<~C.Xm_
`
`om
`
`mm
`
`mmfimomm
`
`AOMHZOU
`
`8
`
`wv
`
`mmfimumm
`
`DMO3
`
`MMQOUMD
`
`om
`
`mmE=2mz<me
`
`N.wE
`
`
`
`
`
`
`U.S. Patent
`
`Sep.21, 1999
`
`Sheet 3 of3
`
`5,956,683
`
`r 72
`
`LSP
`
`74
`
`
`
`LPC
`COEFFICIENTS
`
`I — — — — — — — — _ _ _ _ _ _ _ |
`
`TRANSFORM
`
`CEPSTRUM I
`COEFFICIENTS I
`
`(VECTOR)
`LPC TO
`
`CEPSTRUM
`
`INVERSE TRANSFORM
`'
`L. _ _ _ _ _ _ _ _ _ _ _ _ _ _ _]
`
`FIG. 4
`
`100
`_ _ _ _ _ _ _ I _ _ _ _ _ _
`
`I
`
`"I
`
`I
`
`I
`
`.
`
`SIGNALS
`
`
`
`
`
`‘%%1%%'%€%
`
`I
`
`
`ACOUSTIC
`PROCESSOR
`
`
`
`
`
`LOCAL
`WORD
`
`
`ESTIMATED
`DECODER
`COMMUNICATION
`'
`LOCAL
`
`CHANNEL
`WORD I
`
`STRING . — — _ _ _ _ _ _ _ _ _ _ _ _
`
`108
`
`'—_——_TACT)U_STTC—__—__I
`REMOTE
`FEATURES
`INVERSE
`
`WORD
`TRANSFORM
`DECODER
`
`
`
`
`
`
`
`
`ESTIMATED,
`REMOTE
`WORD I
`STRING .
`
`_1
`
`
`
`5,956,683
`
`1
`DISTRIBUTED VOICE RECOGNITION
`SYSTEM
`
`This application is a continuation of application Ser. No.
`08/534,080, filed Sep. 21, 1995 which is a continuation of
`Ser. No. 08/173,247, filed Dec. 22, 1993.
`
`5
`
`2
`removed and a digital windowing operation is performed on
`each frame to lessen the blocking effects due to the discon-
`tinuity at frame boundaries. A most commonly used window
`function in LPC analysis is the Hamming window, w(n)
`defined as:
`
`FIELD OF THE INVENTION
`
`w(n)=0.54—O.46-cos[
`
`N—l
`" ),0sns1v—1
`
`The present invention relates to speech signal processing.
`More particularly, the present invention relates to a novel
`method and apparatus for realizing a distributed implemen-
`tation of a standard voice recognition system.
`
`DESCRIPTION OF THE RELATED ART
`
`Voice recognition represents one of the most important
`techniques to endow a machine with simulated intelligence
`to recognize user voiced commands and to facilitate human
`interface with the machine. It also represents a key technique
`for human speech understanding. Systems that employ tech-
`niques to recover a linguistic message from an acoustic
`speech signal are called voice recognizers (VR). A voice
`recognizer is composed of an acoustic processor, which
`extracts a sequence of information-bearing features
`(vectors) necessary for VR from the incoming raw speech,
`and a word decoder, which decodes this sequence of features
`(vectors) to yield the meaningful and desired format of
`output, such as a sequence of linguistic words corresponding
`to the input utterance. To increase the performance of a
`given system, training is required to equip the system with
`valid parameters. In other words, the system needs to learn
`before it can function optimally.
`The acoustic processor represents a front end speech
`analysis subsystem in a voice recognizer. In response to an
`input speech signal, it provides an appropriate representation
`to characterize the time-varying speech signal. It should
`discard irrelevant information such as background noise,
`channel distortion, speaker characteristics and manner of
`speaking. Efficient acoustic feature extraction systems will
`furnish voice recognizers with higher acoustic discrimina-
`tion power. The most useful characteristic is the short time
`spectral envelope. In characterizing the short time spectral
`envelope, the two most commonly used spectral analysis
`techniques are linear predictive coding (LPC) and filter-bank
`based spectral analysis models. However, it is readily shown
`(as discussed in Rabiner, L. R and Schafer, R. W., Digital
`Processing of Speech Signals, Prentice Hall, 1978) that LPC
`not only provides a good approximation to the vocal tract
`spectral envelope, but
`is considerably less expensive in
`computation than the filter-bank model in all digital imple-
`mentations. Experience has also demonstrated that the per-
`formance of LPC based voice recognizers is comparable to
`or better than that of filter-bank based recognizers (Rabiner,
`L. R. and Juang, B. H., Fundamentals of Speech
`Recognition, Prentice Hall, 1993).
`Referring to FIG. 1, in an LPC based acoustic processor,
`the input speech is provided to a microphone (not shown)
`and converted to an analog electrical signal. This electrical
`signal is then digitized by an A/D converter (not shown). The
`digitized speech signals are passed through preemphasis
`filter 2 in order to spectrally flatten the signal and to make
`it less susceptible to finite precision effects in subsequent
`signal processing. The preemphasis filtered speech is then
`provided to segmentation element 4 where it is segmented or
`blocked into either temporally overlapped or nonoverlapped
`blocks. The frames of speech data are then provided to
`windowing element 6 where framed DC components are
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`The windowed speech is provided to LPC analysis element
`8. In LPC analysis element 8 autocorrelation functions are
`calculated based on the windowed samples and correspond-
`ing LPC parameters are obtained directly from autocorrela-
`tion functions.
`the word decoder translates the
`Generally speaking,
`acoustic feature sequence produced by the acoustic proces-
`sor into an estimate of the speaker’s original word string.
`This is accomplished in two steps: acoustic pattern matching
`and language modeling. Language modeling can be avoided
`in the applications of isolated word recognition. The LPC
`parameters from LPC analysis element 8 are provided to
`acoustic pattern matching element 10 to detect and classify
`possible acoustic patterns, such as phonemes, syllables,
`words, etc. The candidate patterns are provided to language
`modeling element 12, which models the rules of syntactic
`constraints that determine what sequences of words are
`grammatically well formed and meaningful. Syntactic infor-
`mation can be a valuable guide to voice recognition when
`acoustic information alone is ambiguous. Based on language
`modeling, the VR sequentially interprets the acoustic feature
`matching results and provides the estimated word string.
`Both the acoustic pattern matching and language model-
`ing in the word decoder requires a mathematical model,
`either deterministic or stochastic, to describe the speaker’s
`phonological and acoustic-phonetic variations. The perfor-
`mance of a speech recognition system is directly related to
`the quality of these two modelings. Among the various
`classes of models for acoustic pattern matching, template-
`based dynamic time warping (DTW) and stochastic hidden
`Markov modeling (HMM) are the two most commonly used.
`However, it has been shown that DTW based approach can
`be viewed as a special case of an HMM based one, which is
`a parametric, doubly stochastic model. HMM systems are
`currently the most successful speech recognition algorithms.
`The doubly stochastic property in HMM provides better
`flexibility in absorbing acoustic as well as temporal varia-
`tions associated with speech signals. This usually results in
`improved recognition accuracy. Concerning the language
`model, a stochastic model, called k-gram language model
`which is detailed in F. Jelink, “The Development of an
`Experimental Discrete Dictation Recognizer”, Proc. IEEE,
`vol. 73, pp. 1616-1624, 1985, has been successfully applied
`in practical
`large vocabulary voice recognition systems.
`While in the small vocabulary case, a deterministic grammar
`has been formulated as a finite state network (FSN) in the
`application of airline and reservation and information sys-
`tem (see Rabiner, L. R and Levinson, S. Z., A Speaker-
`Independent, Syntax-Directed, Connected Word Recogni-
`tion System Based on Hidden Markov Model and Level
`Building, IEEE Trans. on IASSP, Vol. 33, No. 3, June 1985
`Statistically, in order to minimize the probability of rec-
`ognition error, the voice recognition problem can be formal-
`ized as follows: with acoustic evidence observation O, the
`operations of voice recognition are to find the most likely
`word string W* such that
`
`W*=arg max P(W|O)
`
`(1)
`
`
`
`5,956,683
`
`3
`where the maximization is over all possible word strings W.
`In accordance with Bayes rule,
`the posteriori probability
`P(W|O) in the above equation can be rewritten as:
`
`P(W)P(0| W)
`P(O)
`
`PW I 0) =
`
`(2)
`
`4
`phonemes, syllables or demisyllables may be used as the
`basic speech unit. The word decoder will be modified
`accordingly.
`Conventional voice recognition systems integrate acous-
`tic processor and word decoders without taking into account
`their separability,
`the limitations of application systems
`(such as power consumption, memory availability, etc.) and
`communication channel characteristics. This motivates the
`
`to recognition,
`Since P(O) is irrelevant
`estimate can be obtained alternatively as:
`
`the word string
`
`10
`
`interest in devising a distributed voice recognition system
`with these two components appropriately separated.
`
`W*=arg max P(W)P(O|W)
`
`(3)
`
`Here P(W) represents the a priori probability that the word
`string W will be uttered, and P(O|W) is the probability that
`the acoustic evidence O will be observed given that the
`speaker uttered the word sequence W. P(O|W) is determined
`by acoustic pattern matching, while the a priori probability
`P(W) is defined by language model utilized.
`In connected word recognition, if the vocabulary is small
`(less than 100), a deterministic grammar can be used to
`rigidly govern which words can logically follow other words
`to form legal sentences in the language. The deterministic
`grammar can be incorporated in the acoustic matching
`algorithm implicitly to constrain the search space of poten-
`tial words and to reduce the computation dramatically.
`However, when the vocabulary size is either medium
`(greater than 100 but less than 1000) or large (greater than
`1000), the probability of the word sequence, W=(w1,w2, .
`.
`.
`, wn), can be obtained by stochastic language modeling.
`From simple probability theory, the prior probability, P(W)
`can be decomposed as
`
`15
`
`20
`
`25
`
`30
`
`P(W) = P<w1. W2,
`
`n
`[:1
`Wit) = H PM I Wla W2,
`
`Wi—1)
`
`(4)
`
`35
`
`,wl-_1) is the probability that wi will
`.
`.
`.
`where P(wl-|w1,w2,
`be spoken given that the word sequence (w1,w2, .
`.
`.
`,wl-_1)
`precedes it. The choice of wt. depends on the entire past
`history of input words. For a vocabulary of size V, it requires
`Vi values to specify P(wl.|w1,w2, .
`.
`. ,wi_1) completely. Even
`for the mid vocabulary size,
`this requires a formidable
`number of samples to train the language model. An inaccu-
`rate estimate of P(wl.|w1,w2,
`.
`.
`.
`,wl._1) due to insufficient
`training data will depreciate the results of original acoustic
`matching.
`A practical solution to the above problems is to assume
`that wt. only depends on (k-1) preceding words, wl._1,wl._2,
`.
`.
`,w,-_k+1. A stochastic language model can be completely
`described in terms of P(wi|w1,w2, .
`.
`.
`,wi_k+1) from which
`k-gram language model is derived. Since most of the word
`strings will never occur in the language if k>3, unigram
`(k=1), bigram (k=2) and trigram (k=3) are the most powerful
`stochastic language models that take grammar into consid-
`eration statistically. Language modeling contains both syn-
`tactic and semantic information which is valuable in
`
`recognition, but these probabilities must be trained from a
`large collection of speech data. When the available training
`data are relatively limited, such as K-grams may never occur
`in the data, P(wl.|wl._2,wl._1) can be estimated directly from
`bigram probability P(w,-|wl-_1. Details of this process can be
`found in F. Jelink, “The Development of An Experimental
`Discrete Dictation Recognizer”, Proc. IEEE, vol. 73, pp.
`1616-1624, 1985. In connected word recognition, whole
`word model
`is used as the basic speech unit, while in
`continuous voice recognition, subband units, such as
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`SUMMARY OF THE INVENTION
`
`The present invention is a novel and improved distributed
`voice recognition system, in which
`the front end acoustic
`processor can be LPC based or filter bank based; (ii) the
`acoustic pattern matching in the word decoder can be based
`on hidden Markov model (HMM), dynamic time warping
`(DTW) or even neural networks (NN); and (iii) for the
`connected or continuous word recognition purpose,
`the
`language model can be based on deterministic or stochastic
`grammars. The present
`invention differs from the usual
`voice recognizer
`in improving system performance by
`appropriately separating the components: feature extraction
`and word decoding. As demonstrated in next examples, if
`LPC based features, such as cepstrum coefficients, are to be
`sent over communication channel, a transformation between
`LPC and LSP can be used to alleviate the noise effects on
`
`feature sequence.
`
`BRIEF DESCRIPTION OF THE DRAWINGS
`
`The features, objects, and advantages of the present
`invention will become more apparent from the detailed
`description set forth below when taken in conjunction with
`the drawings in which like reference characters identify
`correspondingly throughout and wherein:
`FIG. 1 is a block diagram of a traditional speech recog-
`nition system;
`FIG. 2 is a block diagram of an exemplary implementa-
`tion of the present invention in a wireless communication
`environment;
`FIG. 3 is a general block diagram of the present invention;
`FIG. 4 is a block diagram of an exemplary embodiment of
`the transform element and inverse transform element of the
`
`present invention; and
`FIG. 5 is a block diagram of a preferred embodiment of
`the present invention comprising a local word detector in
`addition to a remote word detector.
`
`DETAILED DESCRIPTION OF THE
`PREFERRED EMBODIMENTS
`
`In a standard voice recognizer, either in recognition or in
`training, most of the computational complexity is concen-
`trated in the word decoder subsystem of the voice recog-
`nizer.
`In the implementation of voice recognizers with
`distributed system architecture, it is often desirable to place
`the word decoding task at the subsystem which can absorb
`the computational load appropriately. Whereas the acoustic
`processor should reside as close to the speech source as
`possible to reduce the effects of quantization errors intro-
`duced by signal processing and/or channel induced errors.
`An exemplary implementation of the present invention is
`illustrated in FIG. 2. In the exemplary embodiment, the
`environment is a wireless communication system compris-
`ing a portable cellular telephone or personal communica-
`tions device 40 and a central communications center referred
`
`
`
`5,956,683
`
`5
`to as a cell base station 42. In the exemplary embodiment the
`distributed VR system is presented. In the distributed VR the
`acoustic processor or feature extraction element 22 resides
`in personal communication device 40 and word decoder 48
`resides in the central communications center. If, instead of
`distributed VR, VR is implemented solely in portable cel-
`lular phone it would be highly infeasible even for medium
`size vocabulary, connected word recognition due to high
`computation cost. On the other hand, if VR resides merely
`at the base station, the accuracy can be decreased dramati-
`cally due to the degradation of speech signals associated
`with speech codec and channel effects. Evidently, there are
`three advantages to the proposed distributed system design.
`The first is the reduction in cost of the cellular telephone due
`to the word decoder hardware that is no longer resident in
`the telephone 40. The second is a reduction of the drain on
`the battery (not shown) of portable telephone 40 that would
`result from locally performing the computationally intensive
`word decoder operation. The third is the expected improve-
`ment in recognition accuracy in addition to the flexibility
`and extendibility of the distributed system.
`The speech is provided to microphone 20 which converts
`the speech signal into electrical signals which are provided
`to feature extraction element 22. The signals from micro-
`phone 20 may be analog or digital. If the signals are analog,
`then an analog to digital converter (not shown) may be
`needed to be interposed between microphone 20 and feature
`extraction element 22. The speech signals are provided to
`feature extraction element 22. Feature extraction element 22
`
`extracts relevant characteristics of the input speech that will
`be used to decode the linguistic interpretation of the input
`speech. One example of characteristics that can be used to
`estimate speech is the frequency characteristics of an input
`speech frame. This is frequently provided as linear predic-
`tive coding parameters of the input frame of speech. The
`extracted features of the speech are then provided to trans-
`mitter 24 which codes, modulates and amplifies the
`extracted feature signal and provides the features through
`duplexer 26 to antenna 28, where the speech features are
`transmitted to cellular base station or central communica-
`
`tions center 42. Various types of digital coding, modulation,
`and transmission schemes well known in the art may be
`employed.
`the transmitted
`At central communications center 42,
`features are received at antenna 44 and provided to receiver
`46. Receiver 46 may perform the functions of demodulating
`and decoding the received transmitted features which it
`provides to word decoder 48. Word decoder 48 determines,
`from the speech features, a linguistic estimate of the speech
`and provides an action signal to transmitter 50. Transmitter
`50 performs the functions of amplification, modulation and
`coding of the action signal, and provides the amplified signal
`to antenna 52, which transmits the estimated words or a
`command signal to portable phone 40. Transmitter 50 may
`also employ well known digital coding, modulation or
`transmission techniques.
`At portable phone 40, the estimated words or command
`signals are received at antenna 28, which provides the
`received signal through duplexer 26 to receiver 30 which
`demodulates, decodes the signal and then provides the
`command signal or estimated words to control element 38.
`In response to the received command signal or estimated
`words, control element 38 provides the intended response
`(e.g., dialing a phone number, providing information to
`display screen on the portable phone, etc.).
`The same system represented in FIG. 2 could also serve
`in a slightly different way in that the information sent back
`
`10
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`55
`
`60
`
`65
`
`6
`from central communications center 42 need not be an
`interpretation of the transmitted speech, rather the informa-
`tion sent back from central communications center 42 may
`be a response to the decoded message sent by the portable
`phone. For example, one may inquire of messages on a
`remote answering machine (not shown) coupled via a com-
`munications network to central communications center 42,
`in which case the signal transmitted from central commu-
`nications center 42 to portable telephone 40 may be the
`messages from the answering machine in this implementa-
`tion. Asecond control element 49 would be collocated in the
`central communications center.
`
`The significance of placing the feature extraction element
`22 in portable phone 40 instead of at central communica-
`tions center 42 is as follows. If the acoustic processor is
`placed at central communications center 42, as opposed to
`distributed VR, a low bandwidth digital radio channel may
`require a vocoder (at
`the first subsystem) which limits
`resolution of features vectors due to quantization distortion.
`However, by putting the acoustic processor in the portable or
`cellular phone, one can dedicate entire channel bandwidth to
`feature transmission. Usually, the extracted acoustic feature
`vector requires less bandwidth than the speech signal for
`transmission. Since recognition accuracy is highly depen-
`dent on degradation of input speech signal, one should
`provide feature extraction element 22 as close to the user as
`possible so that feature extraction element 22 extracts fea-
`ture vectors based on microphone speech,
`instead of
`(vocoded) telephone speech which may be additionally
`corrupted in transmission.
`In real applications, voice recognizers are designed to
`operate under ambient conditions, such as background noise.
`Thus,
`it
`is important
`to consider the problem of voice
`recognition in the presence of noise. It has been shown that,
`if the training of vocabulary (reference patterns) is per-
`formed in the exact (or approximate) same environment as
`the test condition, voice recognizers not only can provide
`good performance even in very noisy environments, but can
`reduce the degradation in recognition accuracy due to noise
`significantly. The mismatch between training and test con-
`ditions accounts for one of the major degradation factors in
`recognition performance. With the assumption that acoustic
`features can traverse communication channels more reliably
`than speech signals (since acoustic features require less
`bandwidth than speech signals for transmission as men-
`tioned previously), the proposed distributed voice recogni-
`tion system is advantageous in providing matching condi-
`tions. If a voice recognizer is implemented remotely, the
`matching conditions can be badly broken due mainly to
`channel variations such as fading encountered in wireless
`communications. Implementing VR locally may avoid these
`effects if the huge training computations can be absorbed
`locally. Unfortunately,
`in many applications,
`this is not
`possible. Obviously, distributed voice recognition imple-
`mentation can avoid mismatch conditions induced by chan-
`nel perplexity and compensate for the shortcomings of
`centralized implementations.
`Referring to FIG. 3, the digital speech samples are pro-
`vided to feature extraction element 51 which provides the
`features over communication channel 56 to word estimation
`
`element 62 where an estimated word string is determined.
`The speech signals are provided to acoustic processor 52
`which determines potential features for each speech frame.
`Since word decoder requires acoustic feature sequence as
`input for both recognition and training tasks, these acoustic
`features need to be transmitted across the communication
`
`channel 56. However, not all potential features used in
`
`
`
`5,956,683
`
`7
`typical voice recognition systems are suitable for transmis-
`sion through noisy channels.
`In some cases,
`transform
`element 54 is required to facilitate source encoding and to
`reduce the effects of channel noise. One example of LPC
`based acoustic features which are widely used in voice
`recognizers is cepstrum coefficients, {cl-}. They can be
`obtained directly from LPC coefficients {at} as follows:
`
`8
`
`P/2
`[:1
`P(z)= <1+ z‘1)]_[<1 — 2c0S(W2z—1)z’1 +52)
`
`P/2
`[:1
`Q(z)=(1—z’1)1_[(1- 2c0S(w2.-)z’1 + 52)
`P
`
`P(z) + Q(z)
`A(z)=1-2&1-z"'= f[:1
`
`(12)
`
`(13)
`
`(14)
`
`The LPC coefficients are then provided to LPC to cepstrum
`element 84 which provides the cepstrum coefficients to word
`decoder 64 in accordance with equations 5 and 6.
`Since the word decoder relies solely on an acoustic
`feature sequence, which can be prone to noise if transmitted
`directly through the communication channel, a potential
`acoustic feature sequence is derived and transformed in the
`subsystem 51 as depicted in FIG. 3 into an alternative
`representation that facilitates transmission. The acoustic
`feature sequence for use in word decoder is obtained after-
`wards through inverse transformation. Hence, in distributed
`implementation of VR, the feature sequence sent through the
`air (channel) can be different from the one really used in
`word decoder. It is anticipated that the output from transform
`element 70 can be further encoded by any error protection
`schemes that are known in the art.
`
`In FIG. 5, an improved embodiment of the present inven-
`tion is illustrated. In wireless communication applications,
`users may desire not to occupy the communication channel
`for a small number of simple, but commonly used voiced
`commands, in part due to expensive channel access. This can
`be achieved by further distributing the word decoding func-
`tion between handset 100 and base station 110 in the sense
`
`that a voice recognition with a relatively small vocabulary
`size is implemented locally at handset while a second voice
`recognition system with larger vocabulary size resides
`remotely at base station. They both share the same acoustic
`processor at handset. The vocabulary table in local word
`decoder contains most widely used words or word strings.
`The vocabulary table in remote word decoder, on the other
`hand, contains regular words or word strings. Based on this
`infrastructure, as illustrated in FIG. 5, the average time that
`channel is busy may be lessened and the average recognition
`accuracy increased.
`Moreover, there will exist two groups of voiced com-
`mands available, one, called special voiced command, cor-
`responds to the commands recognizable by local VR and the
`other, called regular voiced command, corresponds to those
`not recognized by the local VR. Whenever a special voiced
`command is issued, the real acoustic features are extracted
`for local word decoder and voice recognition function is
`performed locally without accessing communication chan-
`nel. When a regular voiced command is issued, transformed
`acoustic feature vectors are transmitted through channel and
`word decoding is done remotely at base station.
`Since the acoustic features need not be transformed, nor
`be coded, for any special voiced command and vocabulary
`size is small for local VR, the required computation will be
`much less than that of the remote one (the computation
`associated with the search for correct word string over
`possible vocabularies is proportional to vocabulary size).
`Additionally, the local voice recognizer may be modeled
`with a simplified version of HMM (such as with a lower
`number of states, a lower number of mixture components for
`state output probabilities, etc.) compared to remote VR,
`since the acoustic feature will be fed directly to local VR
`
`m-1
`k
`cm=am+Z[—]ckam_k, m=l,...,P
`/(:1 m
`m—l
`
`cm:
`
`k:l
`
`/C
`m
`[—]ckam_k, m=P+l,...,Q
`
`(5)
`
`10
`
`(6)
`
`15
`
`20
`
`25
`
`30
`
`35
`
`40
`
`45
`
`50
`
`where P is the order of LPC filter used and Q is the size of
`cepstrum feature vector. Since cepstrum feature vectors
`change rapidly, it is not easy to compress a sequence of
`frames of cepstrum coefficients. However, there exists a
`transformation between LPCs and line spectrum pair (LSP)
`frequencies which changes slowly and can be efficiently
`encoded by a delta pulse coded modulation (DPCM)
`scheme. Since cepstrum coefficients can be derived directly
`from LPC coefficients, LPCs are transformed into LSPs by
`transform element 54 which are then encoded to traverse the
`communication channel 56. At remote word estimation
`
`element 62, the transformed potential features are inverse
`transformed by inverse transform element 60 to provide
`acoustic features to word decoder 64 which in response
`provides an estimated word string.
`An exemplary embodiment of the transform element 54 is
`illustrated in FIG. 4 as transform subsystem 70. In FIG. 4,
`the LPC coefficients from acoustic processor 52 are pro-
`vided to LPC to LSP transformation element 72. Within LPC
`to LSP element 72 the LSP coefficients can be determined as
`
`follows. For Pth order LPC coefficients, the corresponding
`LSP frequencies are obtained as the P roots which reside
`between 0 and at of the following equations:
`
`P(w)=cos 5W+p1 cos 4w+ .
`
`Q(w)=cos 5w+q1 cos 4w+ .
`
`.
`
`.
`
`. +p5/2
`
`. +q5/2
`
`(7)
`
`(8)
`
`where pi and qt. can be computed recursively as follows:
`
`P0=‘l0=1
`
`Pi=‘“i‘“P—i‘Pi—1:1 éiép/2
`
`qt-=—a,-+aP_,-—q,-_1,1 §i§P/2
`
`(9)
`
`(10)
`
`(11)
`
`55
`
`The LSP frequencies are provided to DPCM element 74
`where they are encoded for transmission over communica-
`tion channel 76.
`
`At inverse transformation element 78, the received signal
`from the channel is passed through inverse DPCM element
`80 and LSP to LPC element 82 to recover the LSP frequen-
`cies of the speech signal. The inverse process of LPC to LSP
`element 72 is performed by LSP to LPC element 82 which
`converts the LSP frequencies back into LPC coefficients for
`use in deriving the cepstrum coefficients. LSP to LPC
`element 82 performs the conversion as follows:
`
`60
`
`65
`
`
`
`5,956,683
`
`9
`without potential corruption in the channel. This will enable
`a local, though, limited vocabulary, implementation of VR at
`the handset where the computational load is limited. It is
`envisioned that the distributed VR structure can also be used
`in other target applications different than wireless commu-
`nication systems.
`Referring to FIG. 5, speech signals are provide to acoustic
`processor 102, which then extracts features, for example
`LPC based feature parameters,
`from the speech signal.
`These features are then provided to local word decoder 106
`which searches to recognize the input speech signal from its
`small vocabulary. If it fails to decode the input word string
`and determines that remote VR should decode it, it signals
`transform element 104 which prepares the features for
`transmission. The transformed features are then transmitted
`over communication channel 108 to remote word decoder
`110. The transformed features are received at inverse trans-
`form element 112, which performs the inverse operation of
`transform element 104 and provides the acoustic features to
`remote word decoder element 114 which in response pro-
`vides the estimate remote word string.
`The previous description of the preferred embodiments is
`provided to enable any person skilled in the art to make or
`use the present invention. The various modifications to these
`embodiments will be readily apparent to those skilled in the
`art, and the generic principles defined herein may be applied
`to other embodiments without
`the use of the inventive
`faculty. Thus, the present invention is not intend