Method and Apparatus for Generating Speech Pattern Templates

Our invention relates to speech processing and, more particularly, to speech pattern generation for automatic speech synthesis and recognition.

While people find speech a convenient and efficient means for communication, machines are generally organized by programmed arrangements to communicate by means of switches, keyboards and displays in coded formats. Advances in electronic technology have made it possible to provide a speech interface for machines such as computers or automatic control equipment. In accordance with known techniques, machine signals are supplied to a speech synthesizer having a stored vocabulary of coded speech elements. The synthesizer concatenates a speech pattern corresponding to the machine signals from the stored elements. The speech pattern is then converted into a spoken message via an electroacoustic transducer. A speech recognizer receives an utterance and compares the utterance pattern to previously stored speech patterns. the utterance may be identified as the best matching stored pattern and a machine signal corresponding to the matched pattern is generated. Voice communication with machines is thereby accomplished.

In one form of speech synthesis, speech patterns are recorded to provide a stored vocabulary. Responsive to a coded signal, selected patterns are retrieved from the stored vocabulary of speech elements and assembled into a message signal. The message signal is then converted to a sound sequence. The basic speech units in the stored vocabulary may be large message units. If an extensive vocabulary is required, the amount of storage for the different message speech patterns becomes uneconomical.

The use of submessage unit patterns allows greater flexibility in formation of messages and smaller vocabulary storage requirements. Subword speech units such as phonemes, allephones, demisyllables or dyads, for example, allow a large variety of messages to be generated from a very limited stored vocabulary. For example, a relatively small set of syllable subunit patterns or demisyllable subunit patterns may be used to synthesize any message. As is well known in the art, subword speech units are generally produced from word length patterns of ordinary speech by time consuming human analysis, that includes intervention and modification by experts. An arrangement illustrating a combination of manual and automatic determination of speech pattern subunits is described in the article "Automatic Labeling of Continuous Speech With a Given Phonetic Transcription Using Dynamic Programming Algorithms", by Michael Wagner, appearing in the Proceedings of ICASSP '81, pp. 1156-1159, April 1981.

Most speech recognition systems are adapted to receive input speech signals and to transform these speech signals into a sequence of prescribed acoustic features. The input speech acoustic feature sequence is compared to stored sets of previously obtained reference features for identified speech units. The speech signal is identified when the input speech feature sequence matches a stored set of reference features in accordance with predetermined criteria. Where the reference speech units are spoken words, reference speech feature templates may be derived directly from utterances of the spoken word. Word, phrase or sentence templates are useful for limited vocabularies. The template storage requirements for larger vocabularies, however, becomes uneconomical.

As in speech synthesis, it is desirable to limit the template storage requirements of speech recognizers through the use of submessage unit speech patterns. Stored submessage unit templates can be used directly in a speech recognizer to identify an input speech pattern as in U.S. Pat. No. 3,770,892 issued to G. L. Clapper on Nov. 6, 1973 or can be selectively concatenated to form longer message pattern templates. In recognizers adapted to speech patterns of particular speakers, it is necessary to obtain a vocabulary of submessage unit reference templates for each user of the system. Speaker independent arrangements adapted to recognize the speech patterns of any speaker require a set of submessage unit templates representative of a large population. Consequently, a number of submessage unit templates must be generated after the averaging or clustering of speech pattern features obtained from many speakers.

Ideally, a set of submessage unit templates should be generated for each individual whose speech is to be recognized. But the complex and time consuming template formation procedure makes it difficult to produce more than a few template sets in this manner. U.S. Pat. No. 4,060,694 issued to M. Suzuki et al on Nov. 29, 1977 discloses a speech recognition arrangement adapted to a plurality of different speakers. In this system, a set of phoneme characteristic parameters are obtained for a single speaker. Instead of forming additional phoneme characteristics for other speakers, the previously generated phoneme characteristics are artificially varied to produce a range of phoneme characteristic sets.

Prior to recognition of an utterance, the most similar phoneme set to the utterance is chosen from the artificially generated range. In this manner, the time consuming phoneme generation procedure is avoided. The artificial phoneme set, however, only approximates actual differences in voice and pronunciation characteristics to an uncertain extent. Consequently, the use of artificially formed phoneme sets increases the degree of uncertainty in recognition results. It is an object of the invention to provide improved submessage unit template formation without manual analysis or resort to artificial variations of a known subunit speech pattern.

The invention is directed to a speech pattern subunit generation arrangement in which a plurality of reference speech pattern templates are stored. Each template comprises a sequence of acoustic feature signals representative of a prescribed spoken reference speech pattern. For each reference speech pattern template, a set of the time of occurrence of signals representative of a predetermined subunit in said reference speech pattern template feature signal sequence is also stored. An utterance of a prescribed reference speech pattern is analyzed to generate a sequence of acoustic feature signals representative of the utterance. Responsive to the stored reference speech pattern feature signal sequence and the reference speech pattern utterance feature signal sequence, a set of signals representing the time alignment of the utterance feature signals with the stored feature signals is formed. The sequence of utterance feature signals corresponding to the predetermined subunit in the stored reference speech pattern template is determined responsive to the time alignment signals and the stored predetermined subunit representative signals.

In accordance with one aspect of the invention, the reference speech patterns are reference word patterns and the subunit speech patterns are subword unit patterns such as those obtained for phonemes, allephones, demisyllables or other well known components of spoken words. In an embodiment illustrative of the invention, template signals representative of time frame sequences of acoustic feature signals of vocabulary words spoken by a first speaker are stored together with signals representative of the frame boundaries of syllable, demisyllable or other subunits of each spoken word. Another speaker's utterance of a vocabulary word is analyzed to form a frame sequence of acoustic features corresponding thereto. The acoustic feature signals of the other speaker are then transformed so that the number of frames is the same as the first speaker's speech pattern and a signal corresponding to the ratio of the other speaker's utterance length to the first speaker's utterance length is stored. The transformed acoustic feature sequence of the other speaker is dynamically time warped to the acoustic feature sequence of the first speaker and the frame locations of the other speaker's word subunit are determined responsive to the stored frame boundaries and the utterance length ratio signal. The frames of the other speaker's utterance of the word subunit are addressed and the acoustic feature signal sequence of the other speaker's subunit portion are determined for use in speech synthesizers and speech analyzers. Advantageously, there is no need to generate a separate phonetic transcription for each different speaker's utterance and match the phonetic transcription to the speech pattern.

FIG. 1 depicts a block diagram of a speech pattern generator illustrative of the invention;

FIG. 2 shows a detailed block diagram of the demisyllable segment selector of the circuit of FIG. 1; and

FIG. 3 shows a flow chart of a speech pattern generation process illustrative of the invention.

FIG. 1 depicts a block diagram of a subword unit speech pattern generator adapted to provide a set of subword unit templates for different speakers on the basis of a set of previously stored reference templates. It is to be understood that the stored reference templates may be phrases or messages and the subunits may be words or other message subunits. Store 150 contains a set of predetermined reference word templates previously derived by techniques well known in the art. Store 150 may comprise the INTEL 2716 2K.times.8 EPROM or similar device. Each template in store 150, e.g., R.sup.I, contains a sequence of acoustic feature signals

representative of the reference word as spoken by one or more speakers and signals corresponding to the boundary frames NR1(I) and NR2(I) of a prescribed subword unit portion, e.g., demisyllable of the reference word. For example, the phonetic reference word may be "back" and the prescribed subword unit portion may be the final demisyllable "ack". The boundary frame denote the location of the feature signals for the demisyllable "ack."

The flow chart of FIG. 3 illustrates the demisyllable speech pattern generating process. Initially, a speaker provides a start signal ST to controller 140 and a speaker identification signal SP to demisyllable template store 125 via code generator 103 as shown in operation box 300. The reference word identification signal I is set to 1 in word counter 165 as in index initialization box 305 and an instruction message requesting the speaker to input his utterance of reference word I is generated in message generator 135 as per operation box 310. the speaker's utterance is analyzed and a time frame sequence of utterance feature signals

is generated in feature signal generator 115 as per operation box 315. The length of the utterance is compared to a prescribed range to determine any endpoint errors in box 315. In the event the utterance duration is outside the prescribed range, control is transferred to operation box 370 via decision box 320 and a repeat utterance message is produced. The control operations are performed in controller 140. The utterance request for word I is then generated again in operation box 310 and the repeated utterance of word I is analyzed in box 315.

Referring to FIG. 1, an identified speaker SP for whom demisyllable templates are to be generated provides a momentary start signal ST to controller 140 via code generator 103 as well as an identification signal SP. Generator 103 may comprise a keyboard code generator well known in the art. Controller 140 may comprise a microprocessor arrangement such as described in the article "Let a Bipolar Processor Do Your Control and Take Advantage of Its High Speed" by Stephen Y. Lau appearing on pages 128-139 of Electronic Design, 4, February, 1979. The conroller is adapted in accordance with permanently stored instructions to provide the sequence of control signals shown on the right side thereof responsive to the control signal sequence applied thereto. The stored instructions for the operation of the controller are listed in FORTRAN language in Appendix A.

Responsive to the ST signal, controller 140 generates control pulses WRS (word counter reset) and MST (message start). The WRS pulse resets word counter 165 to its I=1 state so that the first reference template in store 150 is addressed. Counter 165 may be the type 74193 integrated circuit. The MST pulse is applied to message generator 135 which is operative as is well known in the art to produce a synthetic speech message instructing speaker SP to utter reference word I=1. Message generator 135 may comprise the National Semiconductor type DT1050 DIGITALKER.TM. Standard Vocabulary Kit. Upon completion of the instruction message, generator 135 supplies signal MD (message done) to controller 140 to indicate that the speaker's utterance is expected at microphone 101.

In response to signal MD, controller 140 generates control pulse AS. The AS pulse is applied to feature signal generator 115 to initiate its operation. The speech signal from microphone 101 is supplied to low-pass filter and sampler 105 which limits the bandwidth of the speech signal to the range, e.g., from 100 Hz to 3200 Hz, and periodically samples the bandlimited speech signal at, e.g., a 6667 Hz rate. Each successive speech signal sample is converted into a digitally coded signal in analog-to-digital converter 110 as is well known in the art.

Feature signal generator 115 receives the coded digital samples from converter 110 and produces the time sequence of linear predictive feature signals T.sup.I representative of the speech signal. Generator 115 also provides a WR (write) pulse for each successive feature signal frame to FIFO stores 120 and 122. Responsive to the sequence of WR pulses, the feature signals of T.sup.I are successively entered into First-in, First-out stores 120 and 122. Each of these FIFO stores may be the Signetics type 8.times.60 FIFO RAM controller.

Feature signal generator 115 may comprise the Linear Prediction Coefficient Generator described in U.S. Pat. No. 4,092,493 issued to L. R. Rabiner et al May 30, 1978 and also includes an endpoint detector such as disclosed in U.S. Pat. No. 3,909,532 issued to L. R. Rabiner et al Sept. 30, 1975 and assigned to the same assignee. The endpoint detector is adapted to detect the termination of the utterance applied to microphone 101. In the event an endpoint does not occur within a prescribed period of time, an EPE (endpoint error) signal is produced by generator 115. The EPE signal is supplied to message generator 135 via OR-gate 137. The ERR output of OR-gate 137 causes controller 140 to produce an MST signal which, in turn, initiates the generation of a repeat word message. After the MD (message done) signal is obtained from generator 135, controller 140 supplies an AS pulse to generator 115. The AS pulse initiates feature signal generation for the expected repeat utterance.

As is well known in the art, articulation and the rate of speech varies significantly from person to person. The stored reference word feature pattern R.sup.I has M.sub.I frames while the utterance feature pattern I.sup.I may have N.noteq.M.sub.I frames. In order to remove differences between the acoustic feature patterns caused by speech rate variation, the utterance feature signals are linearly scaled so that the utterance duration is changed to M.sub.I frames as per operation box 325 of FIG. 3. The scale factor N/M.sub.I is retained to permit inverse scaling at a later time. The modified utterance pattern U.sup.I is an M.sub.I frame pattern. The articulation differences, however, result in a nonlinear relationship between the modified utterance feature signal sequence U.sup.I and the reference word feature sequence R.sup.I.

In order to determine the relationship between the modified utterance pattern frames and reference word pattern frames, the modified utterance feature signal sequence is dynamically time warped to the reference word feature sequence in operation box 330. A signal representative of the best correspondence between the reference word and the input utterance feature signal patterns and a set of signals .delta. representative of the sequence of utterance frames best corresponding to reference word frames are produced by the dynamic time warping. The correspondence signal is compared to a predetermined threshold TH1 in decision box 335. If the modified utterance and reference word patterns are not sufficiently similar, box 370 is entered and the utterance is repeated as previously described.

Referring to FIG. 1, generator 115 supplies an AD pulse to controller 140 upon termination of the feature signal generation for the utterance of word I. At this time, each of stores 120 and 122 contains the set of feature signals T.sup.I for the applied utterance. Linear time warp processor 145 which may comprise the circuit disclosed in U.S. Pat. No. 4,020,332 issued to R. E. Crochiere et al Apr. 26, 1977 and assigned to the same assignee is adapted to modify the feature signals from store 122 so that the utterance feature signals occupy the same time interval as the stored reference feature signals for the same word in reference template store 150. In effect, the number of frames of the applied utterance (N) is changed to the number of frames (M.sub.I) of the reference template for the word in store 150. A signal SF corresponding to the scale factor N/M.sub.I is produced and stored in processor 145.

Upon termination of the feature signal generation in generator 115, control pulse AD is applied from generator 115 to controller 140. Responsive to the AD pulse, controller 140 supplies control pulse LS to linear time warp processor 145 whereby the linear time warp operation is initiated. Processor 145 receives the sequence of feature signals from FIFO store 122 and is operative to modify these feature signals so that the sequence of M.sub.I frame modified feature signals

are produced. The modified utterance feature signals are stored in a frame addressable memory in processor 145.

The time alignment between the modified utterance speech pattern and the stored reference word speech pattern must be determined to account for the differences in articulation between the reference word R.sup.I and the modified utterance U.sup.I. As is well known in the art, the modified utterance feature signals may be time warped to the reference feature signals to form the best matching correspondence signal ##EQU1## where w(i) is the optimal time warping factor between feature vectors U.sup.I and R.sup.I and

d(i,j) is the local distance between the feature vectors of frame i of modified utterance features U.sup.I and the frame j=w(i) of the stored reference features R.sup.I. w(i) is usually restricted so that the slope of warping factor w(i) corresponds to the range of expected speech pattern frames. After the process of Equation 4 is completed, D.sup.I (R.sup.I,U.sup.I) represents the best similarity between the modified utterance features and the reference word features and the warping factor w(i) corresponds to the time alignment path between the modified utterance features and the reference features of Equation 4.

In the dynamic time warping process, the possible time alignment paths between the stored reference features and the modified utterance features are mapped through points (i,j) in accordance with ##EQU2## 1.ltoreq.i.ltoreq.M.sub.I are the reference word feature frames; 1.ltoreq.j.ltoreq.M.sub.I are the modified utterance feature frames and D.sub.A (i,j) is the accumulated distance signal for a path from point (1,1) to point (i,j). d(i,j) is the local distance (Equation 5) between the modified utterance feature vectors at modified utterance feature frame j and the reference word feature vectors at reference word feature frame i.

The rightmost term in Equation 6 represents the accumulated distance up to reference word frame i-1 and modified utterance frame j-I.sub.2 .ltoreq.k.ltoreq.j-I.sub.1. Frame I.sub.1 is chosen to conform to the allowable articulation differences. It is usually set at 1 if w(i-1)=w(i-2) and is set to zero otherwise. I.sub.2 is typically set to 2. Thus k extends over the range from j-2 to j-1 if w(i-1)=w(i-2) and extends over the range from j-2 to j otherwise. Dynamic time warping arrangements are described in greater detail in the article "Minimum Predication Residual Principle Applied to Speech Recognition" by F. Itakura, IEEE Transactions on Acoustics, Speech and Speech Processing, Vol. ASSP-23, pp. 67-72, February 1975 and the article "Considerations in Dynamic Time Warping Algorithms for Discrete Word Recognition" by L. R. Rabiner, A. E. Rosenberg and S. E. Levinson, IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-26, pp. 575-582, December 1978.

In the circuit of FIG. 1, dynamic time warp processor 155 is adapted to generate an overall distance signal D.sup.I in accordance with Equation 4. As aforementioned, signal D.sup.I is a measure of the best similarity between the modified utterance features U.sup.I and the reference word features for reference word R.sup.I. A set of time alignment signals indicative of the time warping path w(i) between the best corresponding feature signals U.sup.I (j) and R.sup.I (i) is also produced.

DTW processor 155 may comprise an arrangement such as described in the Microproducts Hardware System reference published by Data General Corporation, Westboro, Mass., and copyrighted by Data General Corporation, 1979, or other processor systems well known in the art. Processor 155 may include the Micro-NOVA MP 100 system processing unit, the MP 1100 4K/8K dynamic random access memory, the MP/100 8K programmable read only memory and one or more model 422 digital I/O interface units. The operation sequence of the processor is determined by the permanently stored instructions in the read-only memory thereof. These instructions are listed in FORTRAN language form in Appendix B. Processor 155 operating in accordance with the permanently stored instructions of Appendix B performs the dynamic time warping operation of Equation 6 for each reference word frame i. The selection of reference word frames for the time alignment is performed in accordance with Equation 6.

Upon termination of the utterance feature scaling in processor 145, Control pulse LD causes controller 140 to change state and to supply signal DST to processor 145. The DST signal initiates the dynamic time warp processing for the sequence of feature signal time frames. At each reference word frame i, the reference word feature signal R.sup.I (i) is transferred from template store 150 to processor 155. A sequence of selected modified utterance frames j within the range of frames from

is addressed and the feature signals U.sup.I (j) are supplied to processor 155. The cumulative distance signals D.sub.A for the range of Equation 7 are generated in processor 155 and a signal .delta.=i-k representative of the displacement between w(i-1) and w(i) is produced and stored for each (i,j) point. At the last reference word frame M.sub.I, the minimum cumulative distance signal D.sub.I is selected and applied to comparator 160.

Comparator 160 produces a TLE signal if the selected distance signal D.sub.I exceeds a predetermined threshold value. In this event, there is not sufficient similarity between the stored reference word pattern and the modified utterance pattern to determine the frame interval for the prescribed demisyllable in the modified utterance pattern. Consequently, a new input utterance of the word I must be requested. Signal TLE causes OR-gate 137 to produce conrol signal ERR which control signal activates message generator 135 as previously described. Message generator 135 is then rendered operative to synthesize a repeat utterance message and an utterance of I message. Controller 140 then supplies control pulse AS to feature signal generator 115 to initiate an analysis of the new utterance.

Where the distance signal D.sub.I from processor 155 does not exceed the threshold TH1, control pulse DD from processor 155 in FIG. 1 causes controller 140 to generate a BTS pulse and a BT control signal. The BTS pulse is applied to dynamic time warping processor 155 and the back-track operation indicated in operation box 340 of FIG. 3 is initiated in the processor. While processor 155 is in its back-track mode, the sequence of reference word frames are generated in reverse order. The reference word frame signals, m=M.sub.I, M.sub.I -1, . . . , i, . . . 2,1 are output from the processor as coded signals F.sub.R. When the reference word frame i=M.sub.I is addressed, the corresponding frame Fs=M.sub.I of the modified utterance is produced. At reference frame M.sub.I -1 the corresponding utterance frame is M.sub.I -.delta.(M.sub.I). The processing proceeds for each reference word frame j and the modified utterance frame

appears as signal Fs. The back-tracking operations of processor 155 are controlled by stored instructions listed in FORTRAN language in Appendix B.

As the backtracking proceeds, the demisyllable portion of the modified utterance is selected in demisyllable frame selector 180 shown in greater detail in FIG. 2. For each reference word I in store 150, the signals representative of the demisyllable portion of the reference word are available from the store. In the circuit of FIG. 1, a signal corresponding to the beginning frame of the prescribed demisyllable portion, NR1(I), is supplied to one input of comparator 102 in selector 180 from store 150. A signal corresponding to the endframe of the demisyllable portion, NR2(I), is applied to one input of comparator 203 in selector 180 from store 150. Each of comparators 201 and 203 may comprise one or more of the type 7485 integrated circuit.

The sequence of reference frames F.sub.R from processor 155 during the back-track operation is applied to both comparators 201 and 203. When reference word frame signal F.sub.R equals demisyllable endpoint frame signal NR2(I), the output of comparator 203 is enabled at the time that the corresponding modified utterance frame from processor 155, F.sub.S, is applied to latch 212. AND-gate 207 turns on jointly responsive to control signal BT and the enabled output of comparator 203 whereby the modified utterance frame coresponding to the reference word demisyllable endframe is inserted into latch 212. Latch 212 and each of the other latches in FIG. 2 may comprise one or more of the type 74273 integrated circuit.

The sequencing of reference word frames in processor 155 continues. When frame signal F.sub.R equals beginning frame signal NR1(I), the corresponding frame F.sub.S for the modified utterance is placed in latch 210 responsive to the enabled output of comparator 201. Processor 155 continues to perform the back-track operation until reference frame signal F.sub.R =1. At that time, processor 155 produces control pulse BTD which indicates the termination of the back-track operation.

Latch 210 now contains the modified utterance frame N1 corresponding to the beginning frame of the prescribed demisyllable and latch 212 contains modified utterance frame signal N2 corresponding to the endframe of the prescribed demisyllable in the modified utterance. The modified utterance, however, was the result of scaling in linear time warped processor 145 which equalized the durations of the utterance and the reference word. During the scaling operation, signal SF representative of the ratio of the modified utterance duration N to the input utterance duration M.sub.I was produced.

The N1 signal in latch 210 is now applied to one input of multiplier 215 and the scaling factor SF is appied to the other input thereof. Multiplier 215 is operative to form signal N1' which represents the beginning frame of the prescribed demisyllable in the original utterance. Similarly the N2 signal from latch 212 is multiplied by scale factor SF in multiplier 218 to produce a signal N2' representative of the endframe of the prescribed demisyllable in the original utterance. Thus, the prescribed demisyllable portion of the original utterance is determined as the interval between frames N1' and N2' (operation box 345 of FIG. 3). Each of multipliers 213 and 215 may be the TRW MPY-16HJ integrated circuit.

At the end of the back-track operation of operation boxes 340 and 345 in FIG. 3, the boundary frames of the demisyllable portion of the utterance are available at the outputs of multipliers 215 and 213. The utterance acoustic feature signal sequence is stored in FIFO store 120. As indicated in operation box 350 of FIG. 3, the prescribed demisyllable portion of the utterance is selected and stored in demisyllable store 125 responsive to signals N1' and N2'. Store 125 may be the Mostek type MK4802 2kx8 RAM circuit.

Control pulse BTD is applied to controller 140 when the back-track operation is terminated. Responsive to pulse BTD, controller 140 changes state and produces control pulse XR and control signal WDS. The XR pulse resets both demisyllable frame counter 226 and utterance frame counter 230 to their first states. These counters may comprise one or more of the 74193 integrated circuit. The WDS signal is applied to the write enable input of demisyllable feature template store 125 to permit the storage therein of the utterance feature signals from FIFO store 120.

At this time, the F.sub.s =1 frame of the I.sup.th word for identified speaker SP is addressed in template store 125. Controller 140 generates a sequence of XI pulses which successively increment counter 230. The C output of counter 230 representative of the successive utterance frames is supplied to one input of comparator 220 and to one input of comparator 222. The output of multiplier 215 representative of the prescribed demisyllable beginning frame N1' is applied to the other input of comparator 220. Until utterance frame signal C from counter 230 is equal to frame signal N1' from multiplier 215, the output of comparator 220 is disabled and the FC frame output of demisyllable counter 226 remains unaltered. Consequently, the succession of feature signals from FIFO store 120 are loaded into the first frame position of reference word I in demisyllable feature template store 125.

When the C output of utterance frame counter 230 is equal to the demisyllable begin frame signal from multiplier 215, comparator 220 is enabled and subsequent XI count pulses increment demisyllable counter 226 via AND-gate 224. The utterance feature signals for the prescribed demisyllable begin frame N1' are stored in the FC=1 location for reference word I in template store 125. Comparator 220 remains enabled so that demisyllable counter 226 is successively incremented. In this way, the sequence of utterance feature signals for the prescribed demisyllable is placed in the successive locations of template store 125.

Comparator 222 is enabled when the output of utterance frame counter 230 is equal to the N2' endframe of the prescribed demisyllable. The XD signal from comparator 222 occurring at the prescribed demisyllable endframe is supplied to controller 140. The controller is operative to end the generation of XI counter increment pulses. Thus, after all the utterance feature signals in the prescribed demisyllable are stored in template store 125 the operation of demisyllable frame selector 180 is terminated.

Controller 140 supplies control pulse WI to word counter 165 responsive to control pulse XD. The WI pulse increments word counter 165 as per index box 355 in FIG. 3 so that the next reference word template in store 150 is addressed. Comparator 170 is enabled if the word count output I is greater than the last word count IMAX (decision box 360). In this event, signal WD is obtained which places the circuit of FIG. 1 in a wait mode (operation box 365) until an ST (start) signal is received from input circuit 103. Where I does not correspond to the last reference word, signal WD is obtained from comparator 170. The WD signal causes controller 140 to generate signal MST so that an utterance request for the next reference word is produced by message generator 135. After the subword unit patterns are stored in template store 125, they are transferred to utilization device 130 such as a speech recognizer or a speech synthesizer for use therein. The utilization device may, of course, be any speech processing apparatus in which speech pattern templates are used. Alternatively, the templates may be stored in a read only memory that may be incorporated in speech processing equipment.

The invention has been described with reference to a particular embodiment thereof. It is to be understood, however, that various changes in form and detail may be made by those skilled in the art without departing from the spirit and scope of the invention. For example, the reference speech patterns may correspond to a continuous pattern of concatenated words and the subunit speech patterns may correspond to individual words.