GB1603927A - Continuous speech recognition method - Google Patents

Description

PATENT SPECIFICATION ( 11) 1 603 927

r E ( 21) Application No 19336/80 ( 22) Filed 31 May 1978 ( 19) 1 ( 62) Divided Out of No 1603926 ( 31) Convention Application No's 901005 ( 32) Filed 27 Apr 1978 in 901006 : ( 33) United States of America (US) ( 44) Complete Specification Published 2 Dec 1981 ( 51) INT CL 3 G 1 OL 1/00 ( 52) Index at Acceptance G 4 R 11 A 11 C 11 D 11 E 11 F 12 F 12 G 1 F 3 B 3 E 6 B 6 D 7 F 8 F 8 X 9 B 9 C ( 54) IMPROVED CONTINUOUS SPEECH RECOGNITION METHOD ( 71) We, DIALOG SYSTEMS, INC, a corporation organised and existing under the laws of the State of Massachusetts, United States of America, of 32 Locust Street, Belmont, Massachusetts, United States of America, do hereby declare the invention for which we pray that a patent may be granted to us, and the method by which it is to be performed, to

be particularly described in and by the following statement: 5

The present invention relates to a speech recognition method and more particularly to a method for recognizing, in real time, one or more keywords in a continuous audio signal.

Various speech recognition systems have been proposed herebefore to recognize isolated utterances by comparing an unknown isolated audio signal, suitably processed, with one or more previously prepared representations of the known keywords In this context, 10 "keyword" is used to mean a connected group of phonemes and sounds and may be, for example, a portion of a syllable, a word, a phrase, etc While many systems have met with limited success, one system, in particular, has been employed successfully, in commercial applications, to recognize isolated keywords That system operates substantially in accordance with the method described in U S Patent No 4,038,503, granted July 26, 1977, 15 assigned to the assignee of this application, and provides a successful method for recognizing one of a restricted vocabulary of keywords provided that the boundaries of the unknown audio signal data are either silence or background noise as measured by the recognition system That system relies upon the presumption that the interval, during which the unknown audio signal occurs, is well defined and contains a single utterance 20 In a continuous audio signal, (the isolated word is one aspect of the continuous speech signal), such as continuous conversational speech, wherein the keyword boundaries are not a priori known or marked, several methods have been devised to segment the incoming audio data, that is, to determine the boundaries of linguistic units, such as phonemes, syllables, words, sentences, etc, prior to initiation of a keyword recognition process These 25 nrior continuous speech systems, however, have achieved only a limited success in part because a satisfactory segmenting process has not been found Other substantial problems still exist; for example, only limited vocabularies can be consistently recognized with a low false alarm rate, the recognition accuracy is highly sensitive to the differences between voice characteristics of different talkers, and the systems are highly sensitive to distortion in 30 the audio signals being analyzed, such as typically occurs, for example, in audio signals transmitted over ordinary telephone communications apparatus Thus, even though continuous speech is easily discernible and understood by the human observer, machine recognition of even a limited vocabulary of keywords in a continuous audio signal has yet to achieve major success 35 A speech analysis system which is effective in recognizing keywords in continuous speech is described and claimed in copending application No 25735/78 (Serial No 1603925).

That system employs a method in which each keyword is characterized by a template consisting of an ordered sequence of one or more target patterns and each target pattern represents a plurality of short-term keyword power spectra spaced apart in time Together, 40 the target patterns cover all important acoustical events in the keyword The invention claimed in said copending application, features a frequency analysis method comprising the steps of evaluating a set of parameters determining a short-term power spectrum of the audio signal at each of a plurality of equal duration sampling intervals, thereby generating a continuous time-ordered sequence of short-term, audio power spectrum frames; and 45 1 603 927 repeatedly selecting from the sequence of short-term power spectrum frames, one first frame and at least one later occurring frame to form a multi-frame spectral pattern The method further features the steps of comparing, preferably using a likelihood statistic, each thus formed multi-frame pattern, with the first target pattern of each keyword template; and deciding whether each multi-frame pattern corresponds to one of the first target 5 patterns of the keyword templates For each multi-frame pattern which, according to the deciding step, corresponds to a first target pattern of a potential candidate keyword, the method features selecting later occurring frames to form later occurring multi-frame patterns The method then features the steps of deciding in a similar manner whether the later multi-frame patterns correspond respectively to successive target patterns of the 10 potential candidate keyword, and identifying a keyword when a selected sequence of multi-frame patterns corresponds respectively to the target patterns of a keyword template, designated the selected keyword template.

Even though the method claimed in copending application No 25735/78 (Serial No.

1603925) is significantly more effective in recognizing keywords in continuous speech than 15 the prior art systems, even that method falls short of the desired goals.

A principal object of the present invention is therefore a speech recognition method having improved effectiveness in recognizing keywords in a continuous, unmarked audio signal.

The present invention is, in speech analysis apparatus for recognizing at least one 20 predetermined keyword in an audio signal, each said predetermined keyword being represented by a template having at least one target pattern and each target pattern representing at least one short-term power spectrum, an analysis method comprising the steps of forming a sequence of selected patterns from the audio signal, identifying a candidate keyword when said sequence of selected patterns corresponds respectively to the 25 sequence of target patterns of a said keyword template, normalizing the time duration spacings between selected patterns corresponding to said candidate keyword, and applying a prosodic test to said normalized time duration spacings, wherein said normalized time duration spacings for a candidate keyword must meet the timing criteria imposed by said prosodic test before a said candidate keyword is accepted as a recognized keyword 30 An embodiment of the present invention will now be described, by way of example, with reference to the accompanying drawings, in which:Figure 1 is a flow chart illustrating in general terms the sequence of operations performed in accordance with the practice of the present invention.

Figure 2 is a schematic block diagram of electronic apparatus for performing certain 35 preprocessing operations in the overall process illustrated in Figure 1; Figure 3 is a flow diagram of a digital computer program performing certain procedures in the process of Figure 1; and Figure 4 is a graphic tabulation of classification accuracy using different transformation procedures 40 Corresponding reference characters indicate corresponding parts throughout the several views of the drawings.

In the particular preferred embodiment which is described herein, speech recognition is performed by an overall apparatus which involves both a specially constructed electronic system for effecting certain analog and digital processing of incoming audio data signals, 45 generally speech, and a general purpose digital computer which is programmed to effect certain other data reduction steps and numerical evaluations The division of tasks between the hardware portion and the software portion of this system has been made so as to obtain an overall system which can accomplish speech recognition in real time at moderate cost.

However, it should be understood that some of the tasks being performed in hardware in 50 this particular system could well be performed in software and that some of the tasks being performed by software programming in this example might also be performed by special purpose circuitry in a different embodiment of the invention.

As indicated previously, one aspect of the present invention is the provision of apparatus which will recognize keywords in continuous speech signals even though those signals are 55 distorted, for example, by a telephone line Thus, referring in particular to Figure 1, the voice input signal, indicated at 10, may be considered a voice signal, produced by a carbon element telephone transmitter and received over a telephone line encompassing any arbitrary distance or number of switching interchanges A typical application of the invention is therefore recognizing keywords in audio data from an unknown source received 60 over the telephone system On the other hand, the input signal may also be any audio data signal, for example, a voice input signal, taken from a radio telecommunications link, for example, from a commercial broadcast station or from a private dedicated communications link.

As will become apparent from the description, the present method and apparatus are 65

3 1 603 927 3 concerned with the recognition of speech signals containing a sequence of sounds or phonemes, or other recognizable indicia In the description herein, and in the claims, reference is made to either "a keyword," "a sequence of target patterns," "a template pattern," or "a keyword template " the four terms being considered as generic and equivalent This is a convenient way of expressing a recognizable sequence of audio sounds, 5 or representations thereof which the method and apparatus can detect The terms should be broadly and generically construed to encompass anything from a single phoneme, syllable, or sound to a series of words (in the grammatical sense) as well as a single word.

An analog-to-digital (A/D) converter 13 receives the incoming analog audio signal data on line 10 and converts the signal amplitude of the incoming data to a digital form The 10 illustrated A/D converter is designed to convert the input signal data to a twelve-bit binary representation, the conversions occurring at the rate of 8,000 conversions per second The A/D converter 13 applies its output over lines 15 to an autocorrelator 17 The autocorrelator 17 processes the digital input signals to generate a shortterm autocorrelation function 100 times per second and applies its output, as indicated, over lines 19 Each 15 autocorrelation function comprises 32 values or channels, each value being calculated to a 30-bit resolution The autocorrelator is described in greater detail hereinafter with reference to Figure 2.

The autocorrelation functions over lines 19 are Fourier tranformed by Fourier transformation apparatus 21 to obtain the corresponding short-term windowed power 20 spectra over lines 23 The spectra are generated at the same repetition rate as the autocorrelation functions, that is, 100 per second, and each short-term power spectrum has thirty-one numerical terms having a resolution of 16 bits each As will be understood, each of the thirty-one terms in the spectrum represents the signal power within a frequency band.

The Fourier transformation apparatus also preferably includes a Hamming or similar 25 window function to reduce spurious adjacent-band responses.

In the illustrated embodiment, the Fourier transformation as well as subsequent processing steps are performed under the control of a general purpose digital computer, appropriately programmed, utilizing a peripheral array processor for speeding the arithmetic operations required repetitively according to the present method The particular 30 computer employed is a model PDP-11 manufactured by the Digital Equipment Corporation of Maynard, Massachusetts The particular array processor employed is described in copending U K patent application 23910/78 (Serial No 1516000) The programming described hereinafter with reference to Figure 3 is substantially predicated upon the capabilities and characteristics of these commercially available digital processing 35 units.

The short-term windowed power spectra are frequency-response equalized, as indicated at 25, equalization being performed as a function of the peak amplitudes occurring in each frequency band or channel as described in greater detail hereinafter The frequencyresponse equalized spectra, over lines 26, are generated at the rate of 100 per second and 40 each spectrum has thirty-one numerical terms evaluated to 16 bit accuracy To facilitate the final evaluation of the incoming audio data, the frequency-response equalized and windowed spectra over lines 26 are subjected to an amplitude tranformation, as indicated at 35, which imposes a non-linear amplitude transformation on the incoming spectra This transformation is described in greater detail hereinafter, but it may be noted at this point 45 that it improves the accuracy with which the unknown incoming audio signal may be matched with keywords in a reference vocabulary In the illustrated embodiment, this transformation is performed on all of the frequency-response equalized and windowed spectra at a time prior to the comparison of the spectra with keyword templates representing the keywords in the reference vocabulary 50 The amplitude tranformed and equalized short-term spectra over lines 38 are then compared against the keyword templates at 40 The keyword templates, designated at 42, represent the keywords of the reference vocabulary in a spectral pattern with which the transformed and equalized spectra can be compared Candidate words are thus selected according to the closeness of the comparison; and in the illustrated embodiment, the 55 selection process is designed to minimize the likelihood of a missed keyword while rejecting grossly inapplicable pattern sequences The candidate words (and accumulated statistics relating to the corresponding incoming data) are applied over lines 44 for post-decision processing at 46 to reduce the false alarm rate The final decision is indicated at 48 The post-decision processing, which includes the use of a prosodic mask and/or an acoustic-level 60 likelihood ratio test, improves the discrimination between correct detections and false alarms as described in more detail below.

1 603 927 1 603 927 Preprocessor In the apparatus illustrated in Figure 2, an autocorrelation function with its intrinsic averaging is performed digitally on the digital data stream generated by the analog-todigital converter 13 from the incoming analog audio data over line 10, generally a voice signal The converter 13 provides a digital input signal over lines 15 The digital processing 5 functions, as well as the input analog-to-digital conversion, are timed under the control of a clock oscillator 51 The clock oscillator provides a basic timing signal at 256,000 pulses per second, and this signal is applied to a frequency divider 52 to obtain a second timing signal at 8,000 pulses per second The slower timing signal controls the analogto-digital converter 13 together with a latch register 53 which holds the twelve-bit results of the last conversion 10 until the next conversion is completed.

The autocorrelation products are generated by a digital multiplier 56 which multiplies the number contained in a register 53 by the output of a thirty-two word shift register 58 Shift register 58 is operated in a recirculating mode and is driven by the faster clock frequency, so that one complete circulation of the shift register data is accomplished for each 15 analog-to-digital conversion An input to shift register 58 is taken from register 53 once during each complete circulation cycle One input to the digital multiplier 56 is taken directly from the latch register 53 while the other input to the multiplier is taken (with one exception described below) from the current output of the shift register through a multiplexer 59 The multiplications are performed at the higher clock frequency 20 Thus, each value obtained from the A/D conversion is multiplied with each of the preceding 31 conversion values As will be understood by those skilled in the art, the signals thereby generated are equivalent to multiplying the input signal by itself, delayed in time by 32 different time increments (one of which is the zero delay) To produce the zero delay correlation, that is, the power of the signal, multiplexer 59 causes the current value of the 25 latch register 53 to be multiplied by itself at the time each new value is being introduced into the shift register This timing function is indicated at 60.

As will also be understood by those skilled in the art, the products from a single conversion, together with its 31 predecessors, will not be fairly representative of the energy distribution or spectrum over a reasonable sampling interval Accordingly, the apparatus of 30 Figure 2 provides for averaging of these sets of products.

An accumulation process, which effects averaging, is provided by a thirtytwo word shift register 63 which is inter-connected with an adder 65 to form a set of thirty-two accumulators Thus, each word can be recirculated after having been added to the corresponding increment from the digital multiplier The circulation loop passes through a 35 gate 67 which is controlled by a divide-by-N divider circuit 69 driven by the low frequency clock signal The divider 69 divides the lower frequency clock by a factor which determines the number of instantaneous autocorrelation functions which are accumulated, and thus averaged, before the shift register 63 is read out.

In the illustrated example, eighty samples are accumulated before being read out In 40 other words, N for the divide-by-N divider circuit 69 is equal to eighty After eighty conversion samples have thus been correlated and accumulated, the divider circuit 69 triggers a computer interrupt circuit 71 over a line 72 At this time, the contents of the shift register 63 are successively read into the computer memory through a suitable interface circuitry 73, the thirty-two successive words in the register being presented in ordered 45 sequence to the computer through the interface 73 As will be understood by those skilled in the art, this data transfer from a peripheral unit, the autocorrelator preprocessor, to the computer may be typically performed by a direct memory procedure Predicated on an averaging of eighty samples, at an initial sampling rate of 8,000 samples per second, it will be seen that 100 averaged autocorrelation functions are provided to the computer every 50 second.

While the shift register contents are being read out to the computer, the gate 67 is closed so that each of the words in the shift register is effectively reset to zero to permit the accumulation process to begin again.

Expressed in mathematical terms, the operation of the apparatus shown in Figure 2 can 55 be described as follows Assuming that the analog-to-digital converter generates the time series S(t), where t = 0, T 0, 2 TO,, and To is the sampling interval (parsec in the illustrated embodiment), the illustrated digital correlation circuitry of Figure 2 may be considered, ignoring start-up ambiguities, to compute the autocorrelation function 1 603 927 5 ( (jit) = LI S(t-k TO 0) S(t-(k j)To) (Equation 1) k=l 5 where j = 0, 1, 2,, 31; t = 80 To, 160 To,, 80 N To, These autocorrelation functions correspond to the correlation output on lines 19 of Figure 1.

Referring now to Figure 3, the digital correlator operates continuously to transmit to the computer a series of data blocks at the rate of one complete autocorrelation function every 10 ten milliseconds This is indicated at 77 (Figure 3) Each block of data represents the autocorrelation function derived from a corresponding subinterval of time As noted above, the illustrated autocorrelation functions are provided to the computer at the rate of one hundred, 32-word functions per second.

In the illustrated embodiment, the processing of the autocorrelation function data is 15 performed by an appropriately programmed, special purpose digital computer The flow chart, which includes the function provided by the computer program is given in Figure 3.

Again, however, it should be pointed out that various of the steps could also be performed by hardware rather than software and that likewise certain of the functions performed by the apparatus of Figure 2 could additionally be performed in the software by a 20 corresponding revision of the flow chart of Figure 3.

Although the digital correlator of Figure 2 performs some time-averaging of the autocorrelation functions generated on an instantaneous basis, the average autocorrelation functions read out to the computer may still contain some anomalous discontinuities or unevenness which might interfere with the orderly processing and evaluation of the 25 samples Accordingly, each block of data, that is, each autocorrelation function V (j,t), is first smoothed with respect to time This is indicated in the flow chart of Figure 3 at 79 The preferred smoothing process is one in which the smoothed autocorrelation output P s(j,t) is given by 30 ip s(j,t) = Cojp (j,t) + Clip (j,t T) + C 2 V (j,t + T) (Equation 2) where p (j,t) is the unsmoothed input autocorrelation defined in Equation 1, p s(j,t) is the 35 smoothed autocorrelation output, j denotes the delay time, t denotes real time, and T denotes the time interval between consecutively generated autocorrelation functions (equal to 01 second in the preferred embodiment) The weighting functions Co, C 1, C 2, are preferably chosen to be 1/2, 1/4, 1/4 in the illustrated embodiment, although other values could be chosen For example, a smoothing function approximating a Gaussian impulse 40 response with a frequency cutoff of, say, 20 Hertz could have been implemented in the computer software However, experiments indicate that the illustrated, easier to implement, smoothing function provides satisfactory results As indicated, the smoothing function is applied separately for each value j of delay.

As indicated at 81, a cosine Fourier transform is then applied to each time smoothed 45 autocorrelation function, ps (j-t), to generate a 31 point power spectrum The power spectrum is defined as 31 50 S(f,t) = s) ( 0,Ot)W( 0) + 2 '-s(j t)W(j)cos 2 nf j J.= 1 8000 (Equation 3) 55 where S(f,t) is the spectral energy in a band centered at f Hz, at time t; W (j) = ( 1 + cos 2 Jrj/63) is the Hamming window function to reduce side lobes; P s(j,t) is the smoothed autocorrelation function at delay j and time t; and 60 f = 30 + 1000 ( 0 0552 m + 0 438)1/' 63 Hz; m=l, 2,, 31 (Equation 4) 1 603 927 which are frequencies equally spaced on the "mel" scale of pitch As will be understood, this corresponds to a subjective pitch (mel scale) frequency-axis spacing for frequencies in the bandwidth of a typical communication channel of about 300-3500 Hertz As will also be understood, each point or value within each spectrum represents a corresponding band of frequencies While this Fourier transform can be performed completely within the 5 conventional computer hardware, the process may be speeded considerably if an external hardware multiplier or Fast Fourier Transform (FFT) peripheral device is utilized The construction and operation of such modules are well known in the art, however, and are not described in detail herein Advantageously built into the hardware Fast Fourier Transform peripheral device is a frequency smoothing function wherein each of the spectra are 10 smoothed in frequency according to the preferred Hamming window weighting function W (j) defined above This is indicated at 83 of the block 85 which corresponds to the hardware Fourier transform implementation.

As successive smoothed power spectra are received from the Fast Fourier Transform peripheral 85, a communication channel equalization function is obtained by determining a 15 (generally different) peak power spectrum for each incoming windowed power spectrum from peripheral 85, and modifying the output of the Fast Fourier Transform apparatus accordingly, as described below Each newly generated peak amplitude spectrum y (f,t), corresponding to an incoming windowed power spectrum S (f,t), where f is indexed over the plural frequency bands of the spectrum, is the result of a fast attack, slow decay, peak 20 detecting function for each of the spectrum channels of bands The windowed power spectra are normalized with respect to the respective terms of the corresponding peak amplitude spectrum This is indicated at 87.

According to the illustrated embodiment, the values of the "old" peak amplitude spectrum y(f,t-T), determined prior to receiving a new windowed spectrum, are compared 25 on a frequency band by frequency band basis with the new incoming spectrum S(f,t) The new peak spectrum y(f,t) is then generated according to the following rules The power amplitude in each band of the "old" peak amplitude spectrum is multiplied by a fixed fraction, for example, 511/512, in the illustrated example This corresponds to the slow decay portion of the peak detecting function If the power amplitude in a frequency band f 30 of the incoming spectrum S(f,t) is greater than the power amplitude in the corresponding frequency band of the decayed peak amplitude spectrum, then the decayed peak amplitude spectrum value for that (those) frequency bandfs) is replaced by the spectrum value of the corresponding band of the incoming windowed spectrum This corresponds to the fast attack portion of the peak detecting function Mathematically, the peak detecting function 35 can be expressed as y(f,t) = max {y(f,t-T) ( 1-E), S(ft)} (Equation 5) 40 where f is indexed over each of the frequency bands, y(f,t) is the resulting peak spectrum, y(f,t-T) is the "old" or previous peak spectrum, S(f,t) is the new incoming power spectrum, and E is the decay parameter After the peak spectrum is generated the resulting peak amplitude spectrum is frequency smoothed at 89 by averaging each frequency band 45 peak value with peak values corresponding to adjacent frequencies of the newly generated peak spectra, the width of the overall band of frequencies contributing to the average value being approximately equal to the typical frequency separation between formant frequencies As will be understood by those skilled in the speech recognition art, this separation is in the order of 1000 Hz By averaging in this particular way, the useful information in the 50 spectra, that is, the local variations revealing formant resonances are retained whereas overall or gross emphasis in the frequency spectrum is suppressed The resulting smoothed peak amplitude spectrum y(f,t) is then employed to normalize and frequency equalize the just received power spectrum, S(ft), by dividing the amplitude value of each frequency band of the incoming smoothed spectrum S(f,t) by the corresponding frequency band value 55 in the smoothed peak spectrum y(f,t) Mathematically, this corresponds to Sn, (f,t) = S(f,t) / y(f,t) (Equation 6) 60 where Sf(f,t) is the peak normalized smoothed power spectrum and f is indexed over each of the frequency bands This step is indicated at 91 There results a sequence of frequency equalized and normalized short-term power spectra which emphasizes changes in the frequency content of the incoming audio signals while suppressing any generalized 65 1 603 927 long-term frequency emphasis or distortion This method of frequency compensation has been found to be highly advantageous in the recognition of speech signals transmitted over frequency distorting communication links such as telephone lines, in comparison to the more usual systems of frequency compensation in which the basis for compensation is the average power level, either in the whole signal or in each respective frequency band 5 It is useful to point out that, while successive spectra have been variously processed andequalized, the data representing the incoming audio signals still comprises spectra occurring at a rate of 100 per second.

The normalized and frequency equalized spectra, indicated at 91, are subjected to an amplitude transformation, indicated at 93, which effects a non-linear scaling of the 10 spectrum amplitude values Designating the individual equalized and normalized spectra as Sn(f,t) (from Equation 6) where f indexes the different frequency bands of the spectrum and t denotes real time, the non-linearly scaled spectrum x(f,t) is the linear fraction function 51 (f t) A 5 x(ft) = S A (Equation 7 A) where A is the average value of the spectrum Sn(f,t) defined as follows:

20 1 31 A= E S N b 1 t) (Equation 78) 31 f= 1 25 where fb indexes over the frequency bands of the power spectrum.

This scaling function produces a soft threshold and gradual saturation effect for spectral intensities which deviate greatly from the short-term average A Mathematically, for 30 intensities near the average, the function is approximately linear; for intensities further from the average it is approximately logarithmic; and at the extreme values of intensity, it is substantially constant On a logarithmic scale, the function x(f,t) is symmetric about zero and the function exhibits threshold and saturation behavior that is suggestive of an auditory nerve firing-rate function In practice, the overall recognition system performs significantly 35 better with this particular non-linear scaling function than it does with either a linear or a logarithmic scaling of the spectrum amplitudes.

There is thus generated a sequence of amplitude transformed, frequencyresponse equalized, normalized, short-term power spectra x(f,t) where t equals 01, 02, 03, 04, seconds, and f = 1,, 31 (corresponding to the frequency bands of the generated power 40 spectra) Thirty-two words are provided for each spectrum; and the value of A (Equation 7 B), the average value of the spectrum values, is stored in the thirtysecond word The amplitude tranformed, short-term power spectra are stored, as indicated at 95, in a first-in, first-out circulating memory having storage capacity, in the illustrated embodiment, for 256 thirty-two-word spectra There is thus made available for analysis, 2 56 seconds of the audio 45 input signal This storage capacity provides the recognition system with the flexibility required to select spectra at different real times, for analysis and evaluation and thus with the ability to go forward and backward in time as the analysis requires.

Thus, the amplitude transformed power spectra for the last 2 56 secondsgre stored in the circulating memory and are available as needed In operation, ins' the illustrated 50 embodiment, each amplitude transformed power spectrum is stored for 2 56 seconds Thus, a spectrum, which enters the circulating memory at time ti, is lost or shifted from the memory 2 56 seconds later as a new amplitude transformed spectrum, corresponding to a time t, + 2 56, is stored.

The transformed and equalized short-term power spectra passing through the circulating 55 memory are compared, preferably in real time, against a known vocabulary of keywords to detect or pick out those keywords in the continuous audio data Each vocabulary keyword is represented by a template pattern statistically representing a plurality of processed power spectra formed into plural non-overlapping multi-frame (preferably three spectra) design set patterns These patterns are preferably selected to best represent significant acoustical 60 events of the keywords.

The spectra forming the design set patterns are generated for keywords spoken in various contexts using the same system described hereinabove for processing the continuous unknown speech input on line 10 as shown in Figure 3.

Thus, each keyword in the vocabulary has associated with it a generally plural sequence 65 1 603 927 of design set patterns, P(i)I, P(i)2,, which represent, in a domain of short-term power spectra, one designation of that i keyword The collection of design set patterns for each keyword form the statistical basis from which the target patterns are generated.

In the illustrated embodiment of the invention, the design set patterns P(i)j can each be considered a 96 element array comprising three selected short-term power spectra arranged 5 in a series sequence The power spectra forming the pattern should preferably be spaced at least 30 milliseconds apart to avoid spurious correlation due to time domain smoothing In other embodiments of the invention, other sampling strategies can be implemented for choosing the spectra; however the preferred strategy is to select spectra spaced by a constant time duration, preferably 30 milliseconds, and to space the nonoverlapping design 10 set patterns throughout the time interval defining the keyword Thus, a first design set pattern Pl corresponds to a portion of a keyword near its beginning, a second pattern P 2 corresponds to a portion later in time, etc, and the patterns P 1, P 2, form the statistical basis for the series or sequence of target patterns, the keyword template, against which the incoming audio data will be matched The target patterns tl, t 2,, each comprise the 15 statistical data, assuming the P(i)j are comprised of independent Gaussian variables, which enable a likelihood statistic to be generated between selected multiframe patterns, defined below, and the target patterns Thus, the target patterns consist of an array where the entries comprise the mean, standard deviation and area normalization factor for the corresponding collection of design set pattern array entires A more refined likelihood 20 statistic is described below.

It will be obvious to those skilled in the art that substantially all keywords will have more than one contextual and/or regional pronounciation and hence more than one "spelling" of design set patterns Thus, a keyword having the patterned spelling PI, P 2, referred to above, can in actuality be generally expressed as p(i)l, P(i)2, i = 1, 2,, M where each 25 of the p(i)j are possible alternative descriptions of the j class of design set patterns, there being a total of M different spellings for the keyword.

The target patterns t 1, t 2,, ti,, in the most general sense, therefore, each represent plural alternative statistical spellings for the ith group or class of design set patterns In the illustrated embodiment described herein, the term "target patterns" is thus used in the most 30 general sense and each target pattern may therefore have more than one permissible alternative "statistical spelling " Processing the Stored Spectra The stored spectra, at 95, representing the incoming continuous audio data, are 35 compared with the stored template of target patterns indicated at 96, representing keywords of the vocabulary according to the following method Each successive transformed, frequency-response equalized spectrum is treated as a first spectrum member of a multi-frame pattern, here a three spectrum pattern which corresponds to a 96-element vector The second and third spectrum members of the, pattern, in the illustrated 40 embodiment, correspond to spectra occurring 30 and 60 milliseconds later (in real time) In the resulting pattern, indicated at 97, then, the first selected spectrum forms the first 32 elements of the vector, the second selected spectrum forms the second 32 elements of the vector, and the third selected spectrum forms the third 32 elements of the vector.

Preferably, each thus formed multi-frame pattern is transformed according to the 45 following methods to reduce cross-correlation and decrease dimensionality, and to enhance the separation between target pattern classes This is indicated at 99 The transformed patterns in the illustrated embodiment are then applied as inputs to a statistical likelihood calculation, indicated at 100, which computes a measure of the probability that the transformed pattern matches a target pattern 50 Pattern Transformation Considering first the pattern transformation, and using matrix notation, each multi-frame pattern can be represented by a 96-by-1 column vector x = (x I, x 2,, x 96), where x 1, x 2, x 32 are the elements x(f,tl) of the first spectrum frame of the pattern, X 33, X 34, x 64 are the 55 elements x(f,t 2) of the second spectrum frame of the pattern, and X 65, X 66,, X 96 are the elements x(ft 3) of the third spectrum frame Experimentally most of the elements xi of the vector x are observed to have probability distributions that are clustered symmetrically about their mean values so that a Gaussian probability density function closely fits the distribution of each xi ranging over samples from a particular collection of design set 60 patterns corresponding to a particular target pattern However, many pairs xi, xj of elements are found to be significantly correlated, so that an assumption to the effect that the elements of x are mutually independent and uncorrelated would be unwarranted.

Moreover, the correlations between elements arising from different frames in the multi-frame pattern convey information about the direction of motion of formant 65 1 603 927 resonances in the input speech signal, and this information remains relatively constant even though the average frequencies of the formant resonances may vary, as from talker to talker As is well known, the directions of motion of formant resonance frequencies are important cues for human speech perception.

As is well known, the effect of cross correlations among the elements of x can be taken into account by employing the multivariate Gaussian log likelihood statistic -L = 1/2 (x-x)K-1 (x x)t + 1/2 1 n I Kr| (Equation 8 A) where x is the sample mean of x, K is the matrix of sample covariances between all pairs of elements of x defined by Kij = (xi xi) (xj xj), (Equation 8 B) and I| K I denotes the determinant of the matrix K The covariance matrix K can be decomposed by well-known methods into an eigenvector representation K = EV Et (Equation 8 C) where E is the matrix of eigenvectors ei of K, and V is the diagonal matrix of eigenvalues vi of K These quantities are defined by the relation Keit = vieit (Equation 8 D) Multiplication by the matrix E corresponds to a rigid rotation in the 96dimensional space in which the vectors x are represented Now if a transformed vector W is defined as (Equation 8 E) then the likelihood statistic can be rewritten as -L = 1/2 w V-Iwt + 1/2 in IIK 1 l 96 = V/2 I=l w 2 + n V + In v I (Equation 8 F) Each eigenvalue vi is the statistical variance of the random vector x measured in the direction of eigenvector ei.

The parameters Kij and xi are determined, in the illustrated embodiment, by averaging formed multi-frame patterns, for each of the indicated statistical functions, over a number of observed design set samples This procedure forms statistical estimates of the expected values of Kij and xi However, the number of independent parameters to be estimated is 96 mean values plus 96 x 97/2 = 4656 covariances Since it is impractical to collect more than a few hundred design set pattern samples for a target pattern, the achievable number of sample observations per statistical parameter is evidently quite small The effect of insufficient sample size is that chance fluctuations in the parameter estimates are comparable to the parameters being estimated These relatively large fluctuations induce a strong statistical bias on the classification accuracy of the decision processor based on w = E(x)t 1 603 927 10 equation 8 F, so that although the processor may be able to classify the samples from its own design set patterns with high accuracy, the performance measured with unknown data samples will be quite poor.

It is well known that by reducing the number of statistical parameters to be estimated, the effect of small sample bias is reduced To that end, the following method has been 5 commonly employed to reduce the dimensionality of a statistical random vector The eigenvectors ei defined above are ranked by decreasing order of their associated eigenvalues v;, to form a ranked matrix Er of ranked eigenvectors er so that e 1 is the direction of maximum variance Vr 1 and vr'-, se v', Then the vector x-x is transformed into a vector W as in equation 8 E, (using the ranked matrix Er), but only the first p elements of W 10 are utilized to represent the pattern vector x In this representation, sometimes termed "principal component analysis," the effective number of statistical parameters to be estimated would be in the order of 96 p instead of 4656 To classify patterns the likelihood statistic L is computed as in equation 8 F except that the summation now ranges from 1 to p instead of from 1 to 96 On applying the principal component analysis method to practical 15 data it is observed that the classification accuracy of the processor increases as p increases, until at a critical value of p the accuracy is a maximum; thereafter the accuracy diminishes as p is increased until the poor performance described above is observed at p= 96 (See Figure 4, graph (a) (training set data) and graph (b) (unknown input data) ).

The maximum classification accuracy achieved by the principal component method is still 20 limited by a small sample statistical bias effect, and the number of components, or dimensions, required is much larger than one would expect is really necessary to represent the data Furthermore it can be seen from the illustration (Figure 4) that the performance for design set pattern samples is actually worse than the performance for unknown samples, over a wide range of p 25 The source of the latter two effects is found in the fact that by representing the sample space with p components of the transformed vector w, the contribution of the remaining 96-p components has been left out of the likelihood statistic L A region where most of the pattern samples are found has thus been described, but the regions where few samples occur has not been described The latter regions correspond to the tails of the probability 30 distribution and thus to the regions of overlap between the different target pattern classes.

The prior art method thus eliminates the very information needed to make the most difficult classification decisions Unfortunately these regions of overlap are of high dimensionality, so it is impractical to reverse the argument above and employ, for example, a small number of the components of W for which the variance v; is smallest instead of 35 largest.

The effect of the unutilized components wp+ 1,,w 96 is estimated by a reconstruction statistic R in the following manner The terms dropped out of the expression for L (Equation 8 F) contain the squares of the components wi, each weighted in accordance with its variance vi All these variances can be approximated by a constant parameter c, which 40 can then be factored out thus 96W 2 96 W 2 a C L (Equation 8 G) 45 1 =p+l i=p+l The summation on the right is just the square of the Euclidean norm (length) of the vector 50 w' = (wp+,,ws 6) (Equation 8 H) 55 Define the vector wp to be wp = (w,,wp) (Equation 8 I) 60 Then 96 2 1 W 12 WP 2 (Equation 8 Bp) i 2 = w'(q i p 1 1 603 927 since the vectors w, w' and wp can be translated so as to form a right triangle The eigenvector matrix E produces an orthogonal transformation, so the length of W is the same as the length of x-x Therefore it is not necessary to compute all the components of w The statistic sought, which estimates the effect of the unutilized components upon the log likelihood function L, is thus 5 R (Ix-x 2,,Pl V 2)/2 (Equation 8 K) 10 This is the length of the difference between the observed vector x-x and the vector that would be obtained by attempting to reconstruct x-ix as a linear combination of the first p eigenvectors ei of K R therefore has the character of a reconstruction error statistic To 15 utilize R in the likelihood function it may simply be adjoined to the set of transformed vector components to produce a new random vector (w 1,w 2, w P,R) which is assumed to have independent Gaussian components Under this assumption the new likelihood statistic is 20 -L 1/2 L var ' -)+ 1/2 I in varmw) + M (Equation 8 L) 25 where -230 M = 112 (R R) + 1/2 in var (R) (Equation 8 M) var(R) and the barred variables are sample means and var( denotes the unbiased sample variance.

In Equation 8 L the value of wi should be zero, and var(wi) should be equal to vi; however 35 the elgenvectors cannot be computed or applied with infinite arithmetic precision, so it is best to remeasure the sample means and variances after transformation to reduce the systematic statistical bias produced by arithmetic roundoff errors This remark applies also to Equation 8 F.

The measured performance of the likelihood statistic L' in the same maximum likelihood 40 decision processor is plotted as graphs (c) and (d) of Figure 4 It can be seen that as p increases, the classification accuracy again reaches a maximum, but this time at a much smaller number p of dimensions Moreover the maximum accuracy achieved is noticeably higher than for the statistic L, which differs only by omission of the reconstruction error R.

As a further test of the efficacy of the reconstruction error statistic R, the same practical 45 experiment was again repeated, but this time the likelihood function employed was simply so L = M-M (Equation 8 N) 50 That is, this time the region in which most of the sample data lie was ignored, while the regions where relatively few samples are found was described The maximum accuracy obtained (graphs (e) and (f) of Figure 4) is very nearly as high as for the statistic L', and the maximum occurs at a still smaller number of dimensions p= 3 The result can be interpreted 55 to mean that any data sample lying in the space of the first p eigenvectors of K can be accepted as belonging to the target pattern class, and that there is little or no benefit to be gained by making detailed probability estimates within that space.

Statistical Likelihood Calculation 60 The transformed data w;, corresponding to a formed multi-frame pattern x, are applied as inputs to the statistical likelihood calculation This processor, as noted above, computes a measure of the probability that the unknown input speech, represented by the successively p 5 resented, transformed, multi-frame patterns, matches each of the target patterns of the keyword templates in the machine's vocabulary Typically, each datum of a target pattern 65 1 1 1 603 927 has a slightly skewed probability density, but nevertheless is well approximated statistically by a normal distribution having a mean value wi and a variance var(w 1) where i is the sequential designation of the elements of the kh target pattern The simplest implementation of the process assumes that the data associated with different values of i and k are uncorrelated so that the joint probability density for the datum x belonging to target pattern 5 k is (logarithmically) -21 ( 1 W 1) 1 L(tlk) = p(x,k)= j/2 In 2 m (var(w)) 1/2 10 (Equation 9) 15 Since the logarithm is a monotonic function, this statistic is sufficient to determine whether the probability of a match with any one target pattern of a keyword template is greater than or less than the probability of a match with some other vocabulary target pattern, or alternatively whether the probability of a match with a particular pattern exceeds a predetermined minimum level Each input multi-frame pattern has its statistical 20 likelihood L(tlk) calculated for all of the target patterns of the keyword templates of the vocabulary The resulting likelihood statistics L(tlk) are interpreted as the relative likelihood of occurrence of the target pattern named k at time t.

As will be well understood by those skilled in the art, the ranking of these likelihood statistics constitutes the speech recognition insofar as it can be performed from a single 25 target pattern These likelihood statistics can be utilized in various ways in an overall system, depending upon the ultimate function to be performed.

Selection of Candidate Keywords According to the preferred embodiment of the invention, if the likelihood statistic of a 30 multi-frame pattern with respect to any first target pattern exceeds a predetermined threshold; the comparison being indicated at 101, 103, the incoming data are studied further to determine first a local maximum for the likelihood statistic corresponding to the designated first target pattern, and second, whether other multi-frame patterns exist which correspond to other patterns of the selected potential candidate keywords This is indicated 35 at 105 Thus, the process of repetitively testing newly formed multispectrum frames against all first target patterns is interrupted; and a search begins for a pattern, occurring after the "first" multi-frame pattern, which best corresponds, in a statistical likelihood sense, to the next (second) target pattern of the potential candidate keyword(s).

If a "second" multi-frame pattern corresponding to the second target pattern(s) is not 40 detected within a preset time window, the search sequence terminates, and the recognition process restarts at a time just after the end of the "first" multi-frame pattern which identified a potential candidate keyword Thus, after the "first" multiframe pattern produces a likelihood score greater than the required threshold, a timing window is provided within which time a pattern matching the next target pattern in sequence 45 corresponding to the selected potential candidate keyword(s) must appear.

The timing window may be variable, depending for example upon the duration of phonetic segments of the particular potential candidate keyword.

This process continues until either ( 1) multi-frame patterns are identified in the incoming data for all of the target patterns of a keyword template or ( 2) a target pattern cannot be 50 associated with any pattern occurring within the allowed time window If the search is terminated by condition ( 2), the search for a new "first" spectrum frame begins anew, as noted above, at the spectrum next following the end of the "first" previously identified multi-frame pattern.

At this processing level, the objective is to concatenate possible multiframe patterns 55 corresponding to target patterns, and to form candidate words (This is indicated at 107).

The detection thresholds are therefore set loosely so that it is very unlikely that a correct multi-frame pattern will be rejected, and here, at this acoustic processing level, discrimination between correct detection and false alarms is obtained primarily by the requirement that a number of the pattern events must be detected jointly 60 Post-Decision Processing Processing at the acoustic level continues in this manner until the incoming audio signals terminate However, even after a keyword is identified using the likelihood probability test described above, additional post-decision processing tests (indicated at 109) are used to 65 1 603 927 decrease the likelihood of selecting an incorrect keyword (i e to reduce the false alarm rate) while maintaining the probability of a correct detection as high as possible For this reason, the output of the acoustic level processor, that is, a candidate word selected by a concatenation process, is filtered further by a mask of prosodic relative timing windows and/or a likelihood ratio test which uses information from the acoustic level processor 5 concerning all target pattern classes.

The Prosodic Mask As noted above, during the determination of the likelihood statistics, the time of occurrence of the multi-frame pattern having the local peak value of likelihood statistic 10 relative to the active target pattern is found and in the preferred embodiment is recorded for each of the selected patterns corresponding to the several successive target patterns of a candidate keyword Those times, ptl, pt 2,, ptn for each candidate keyword are analyzed and evaluated according to a predetermined prosodic mask for that keyword to determine whether the time intervals between successive patterns likelihood peaks meet predeter 15 mined criteria According to the method, the elapsed times between the times of the peak value of likelihood statistic, that is, pti-pti-1, for i = 2, 3,, n, are first normalized by dividing each elapsed time interval by: ptn-ptl The resulting normalized intervals are compared with a prosodic mask, that is, a sequence of allowable ranges of normalized interval length, for the candidate keyword, and if the interval lengths fall within the selected 20 ranges, the candidate word is accepted.

In the illustrated embodiment the prosodic mask timing windows are determined by measuring the elapsed intervals for sample keywords spoken by as large a number of different speakers as possible The prosodic pattern is then compared with the statistical sample keyword times using a statistical calculation wherein the mean and standard 25 deviation for each prosodic mask (corresponding to each keyword) are derived from the keyword design set pattern samples Thereafter, the likelihood statistic is calculated for deciding whether to accept and thus render a final decision with respect to the candidate keyword This likelihood statistic relates to the timing of events and is not to be confused with the likelihood statistic applied to the multi-frame patterns relative to the target 30 patterns.

In another embodiment of the invention, the ranges of normalize interval duration are loosely set, but are inflexibly fixed In this embodiment, a candidate keyword is accepted only if the normalized interval times fall within the fixed window boundaries Thus a candidate word is acceptable only if each of the normalized times fall within the set limits 35 Word-Level Likelihood Ratio Test In the preferred embodiment of the invention, each candidate work is also tested according to a likelihood ratio test before a final decision to accept the keyword is made.

The likelihood ratio test consists of summing a figure of merit over the sequence of selected 40 multi-frame patterns which have been identified with the candidate keyword The accumulated figure of merit, which is the sum of the figures of merit for each multi-frame pattern, is then compared with a decision threshold value.

The figure of merit for a detected multi-frame pattern is the difference between the best log likelihood statistic relative to any target pattern in the keyword vocabulary and the best 45 score relative to those which are permitted choices for the target pattern Thus, if the best scoring target pattern is a legal alternative for the pattern sought, the figure of merit has the value zero However, if the best score corresponds to a target pattern not in the list of alternatives for the selected candidate word target pattern (a given target pattern may have several statistical spellings depending upon accents, etc), then the figure of merit is the 50 difference between the best score and the best among those that did appear in the list of alternates The decision threshold is optimally placed to obtain the best balance between missed detection and false alarm rates.

Considering the word level likelihood ratio test from a mathematical point of view, the probability that a random multi-frame pattern x occurs, given that the input speech 55 corresponds to target pattern class k, equals p(xtk), read "the probability of x given k " The log likelihood statistic, then, of the input x relative to the kth reference pattern is L(xlk) and equals in p(x,k) as defined by Equation 9 Assuming that the detected multi-frame pattern must be caused by one of a group of N predefined target pattern classes, and assuming that either the classes occur with equal frequency or the N possible choices are considered to be 60 equally valid, then the probability, in the sense of a relative frequency of occurrence, of observing the event x in any case is the sum of the probability densities defined by the summation:

14 1 603 927 14 n 1 p p(x)= E p (xl k) -n (Equation 10) k=l 5 Of these occurrences, the proportion attributable to 'a given class, p(kjx) equals:

(k)1 10 p(k lx) P(Xjk)' N (Equation 11 A) 1 -P(x pli) N 15 i= 1 or logarithmically, 20 nIn p(klx) = L(xlk) ln p(xli) (Equation 11 B) i 125 If the decision processor is then given x, and for some reason chooses class k, then equation 11 A or 11 B above gives the probability that that choice is correct The above equations are consequences of Bayes' rule:

30 p(x,k) = p(xik) p(k) = p(kix) p(x), wherein p(k) is taken to be the constant 1/n 35 If one assumes that only one class, say class m, is very likely, then equation 10 is approximated by p(x) max { p(xli) À 1/n} = p(xjm) À 1/n (Equation 12) 40 i and we have 45 P (k,m,x) = L(xlk) L(xlm) ln p(klx) (Equation 13) Note that if the kth class is the most likely one, then the function P assumes its maximum value zero Summing over the set of presumed independent multi-frame patterns, the 50 accumulated value of f estimates the probability that the detected word is not a false alarm.

Hence, a decision threshold on the accumulated value of P relates directly to the trade-off between detection and false alarm probabilities and is the basis of the likelihood ratio test.

The accumulated value of f 3 then corresponds to the figure of merit of the candidate keyword 55 Reference is made to our copending application Nos 25735/78 (Serial No 1603925), which discloses and claims an analysis method as disclosed herein, 25736/78 (Serial No.

1603927), which discloses and claims an analysis method as disclosed herein and from which this application has been divided out, 8035763 (Serial No 1603929), and 8019337 (Serial No 1603928), which have also been divided out from Application No 25736/78 and which 60 disclose and claim analysis methods as disclosed herein.

Claims (1)

1. WHAT WE CLAIM IS:-

   1 In speech analysis apparatus for recognizing at least one predetermined keyword in an audio signal, each said predetermined keyword being represented by a template having at least one target pattern and each target pattern representing at least one short-term 65 1 603 927 15 power spectrum, an analysis method comprising the steps of forming a sequence of selected patterns from the audio signal, identifying a candidate keyword when said seq ience of selected patterns corresponds respectively to the sequence of target patterns of a said keyword template, normalizing the time duration spacings between selected patterns corresponding to said candidate keyword, and applying a prosodic test to said normalized 5 time duration spacings, wherein said normalized time duration spacings for a candidate keyword must meet the timing criteria imposed by said prosodic test before a said candidate keyword is accepted as a recognized keyword.

   For the Applicant 10 GRAHAM WAIT & CO, Chartered Patent Agents, 3, Gray's Inn Square, London, WC 1 R 5 AH.

   Printed for Her Majesty's Stationery Office by Croydon Printing Company Limited Croydon Surrey, 1981.

   Published by The Patent Office 25 Southampton Buildings, London, WC 2 A l AY from which copies may be obtained