DE1803621A1 - Facility for synthesizing human speech - Google Patents

Abstract

1,263,523. Speech synthesis. RCA CORPORATION. 12 Feb., 1969 [26 Feb., 1968 (2)], No. 7608/69. Heading H4R. A speech synthesizer has a library of phonemes stored at selected locations of a data store to be read out in any desired sequence to synthesize desired audible information. Each phoneme is truncated at the beginning and/or the end to provide continuity of formants across the junctions between the phonemes in the desired sequence. The phonemes are read out in sequence as determined manually, by punched card, paper, or magnetic tape, or computer output. Each phoneme to be reproduced is represented by three numbers, one defining the position of the phoneme in the library and the other two the beginning and end respectively of the desired portion of the phoneme. Preferably the phonemes are stored in digital, i.e. pcm, form on a magnetic drum with core memories used as an intermediate store during reading in and reading out of the stores. Fig. 9 shows a system used for recording phonemes the analogue waveforms of which are fed as 7 bit pcm signals into a core memory A. The core memory can store 4096 seven bit signals which represent one phoneme. The signals are fed into the core memory in defined locations commencing when a record command pulse 707 is received. When the core memory is full, flip-flop 709 is tripped and further storage ceases. After an index pulse is received from the drum 21 the core memory is read out in predetermined sequence to store on the drum 21, in locations defined by the sector pulses permanently recorded on the drum, the digital signals defining the phoneme. Further phonemes are recorded in further tracks of the drum. Fig. 10 shows apparatus for reading phonemes from the drum 821 under manual, 831, or paper tape, 833, control. The phenomes requested by the manual or paper tape control are read alternately into the core memories 17 and 827 and while one phoneme is being read into one core memory the desired portion of the previous phoneme, as defined by the second and third numbers read out of the manual or paper tape control, is read out from the other core memory, functions of the core memories being interchanged as soon as the phoneme has been read out of the core store for reproduction. Reading in from the drum to the core store takes place faster than reading out providing allowance for the truncation of the reproduced phoneme. It is suggested that the phoneme might be recorded in analogue fashion on the drum.

Description

6682-68 / Dr.ν.Β / Ε

RCA 59,488 / 593694

U.S. Serial Nos. 709,323 / 708,389

Piled February 26, 1968

Radio Corporation of America New York N.Y. (V.St.A.)

Facility for the synthesis of human language

The present invention relates to an apparatus for synthesizing human speech or others Aud i ο information.

Human speech is audio information that contains a series of complex sounds that pass through the larynx, the tongue, mouth and nasal spaces as well as the breathing strength can be controlled and determined. As a result of the skills learned speaking and understanding language, the difficult problems involved are easily underestimated. In the artificial generation of language, all factors must be so minor they may appear should be taken into account in so far as they contribute to the production of intelligible spoken words.

The recording of speech, like that of music, is usually analogue. The continuous changes the amplitude and frequency of the audio information are thus retained on the storage medium. The reproduction of the language can then simply by converting back the recorded

909836/0785

* · Ο -

Signals take place in audible sound information.

When synthesizing audio information, in particular Speech, more than simple reproduction is sought. Rather, the purpose of synthetic speech generation is in translating abstract facts or stored information into understandable language to convey the facts or information to be able to tell someone they want to know.

Various solutions to this problem are known. The most obvious solution is to have them all in one certain scope possible sentences that the user might want or need to be saved. For this, however, even with a small number of facts, those for storing all permutations and combinations of facts required storage capacity prohibitively large.

In order to reduce the storage capacity, it is known to save only parts of sentences instead of whole sentences. Here too for a relatively large storage capacity is still required. A further reduction in storage capacity is possible as a result, that you only save single words and put them together in sentences. However, this known measure has the disadvantage that the vocabulary remains relatively limited. Similar difficulties arise when using syllables.

The most successful solution, with which a large vocabulary can be achieved without large storage capacities required consists in the use of the smallest linguistic units, the so-called phonemes.

A phoneme is a group of the same or related sounds that fall under different phonetic behaviors. change conditions. There are 40 phonemes in the English language, which can be divided into seven groups.

The first three groups comprise the vowels. The first group contains the ten simple vowels; the second Group the six diphtongs and the third group the four semi-vowels and liquids.

909836/0785

The fourth group contains the six explosive sounds.

The fifth group contains the three nasal consonants,

The sixth group contains nine pricatives or spirants, created by a rubbing sound of breathing on one Part of the oral cavity are marked.

The lower group contains the two affricates. this are plosive or explosive sounds, followed by a slow separation of the sound-forming organs, so that the last part is a Pricative or Spirant with a corresponding organ position.

The phonemes listed above are given in groups in Table I below. Each phoneme is means a simple typical word and underlined for identification.

Table I.

Elementary sounds (phonemes) of the English language

simple vowels

(D fit (2) feet (3) let (Ί) asked (5) but (6) n £ t (7) law (8th) book (9) boat (10) bird

II. Diphtongs

(1) pain

(2) go

(3) hous e (H) toe

(5) boy,

(6) few

III. Half vowels and liquids

(1) y_ou

(2) we

(3) late rate

909836/0785

IV. (D V. (D Explosive sounds (2) (2) bath (3) (3) dive (4) give (5) jpot (6) toy £ at Nasal consonants may now sing

VI,

Pricatives

VII

(D jz he ο (2) vision (3) very (4) that (5) Has (6) fat (7) thing (8th) shed (9) sat Affricates

church

(D
(2)

However, for the artificial generation of speech it is not sufficient to simply use a sequence of recorded phonemes to reproduce. In order to generate a natural-sounding artificial language from phonemes, three more conditions have to be met be fulfilled:

1. The continuity of the speech oscillation must be guaranteed at the transitions between the phonemes;

2. The continuity of the pitch must be guaranteed at the phoneme boundaries and

3 · The continuity of the frequency components of the phonemes must be ensured at the transitions between the phonemes be.

The predominant frequencies, the formants, can be viewed as frequency components of a phoneme should be designated. As is well known, every complicated periodic oscillation can be made up of a combination of sinusoidal oscillations suitable frequencies, amplitudes and phase positions are put together. The characteristic sound of a phoneme can through the combination of no more than three formants, each of which may or may not change over time, recognizable

909836/0785

be practiced.

Synthesizing speech (or other classes of audio information) from phonemes thus requires choice a corresponding phoneme sequence and a merging of the pormants of the respective phonemes at their junctures, so that the resulting speech has no breaks.

Various possibilities are already known achieve a smooth transition between phonemes (U.S. Patents 2,771,509; 2,860,187; 3,158,685). With the known Solutions are made through the transitions between the phonemes special circuits produced, which bring about the necessary continuity at the transition, or there are each of one Phoneme different forms are stored so that one can choose a phoneme which gives a smooth transition.

The present invention is based on the object of a device for synthesizing the human Specify language that is characterized by a previously unattainable language quality without requiring large storage capacities or restrictions of the vocabulary have to be accepted. The artificial speech can through control signals which determine the information to be transmitted. Devices of the present type are suitable e.g. for converting the output data supplied by an electronic data processing system or another device into understandable language.

According to the invention is a device for synthesizing the human language characterized by a memory for storing coded phonemes, a Control device for extracting the coded phonemes in a desired order from the memory, and a conversion device to convert the extracted phonemes into continuous electrical signals.

According to a VJe it formation of the invention, the control device can when removing a phoneme from a memory

909836/0785

ΘΑΟ OKJüINAL

ί t * 1 it it

—6—

begin at a different location from the beginning of the stored phoneme and end at a specific location, which is different from the end of the stored phoneme. The places where the beginning and end are given in such a way that that the desired continuity of language results with the preceding and following phoneme.

According to another development of the invention, precautions have been taken to convert spoken phonemes into electrical To convert signals for storage in memory. For a digital storage of the phonemes a sampling or Be provided converter circuit.

The invention is explained in more detail below with reference to the drawing, which show:

Fig. 1 is a graphical representation of the approximate time course of the formants of the spoken word "WED";

2 shows a graphical representation of the approximate course over time of the formants of the spoken word "CALF";

Fig. 3 is a graph taken from how a transition is determined to ensure continuity of formants between phonemes can be determined;

FIG. 4 shows a graphical representation of the course over time of two mutually merging phonemes according to FIG. 3; FIG.

5 is a graph showing two periods of a typical complex oscillation are as they appear in language;

FIG. 6 is a graphical representation of the pulses generated during a periodic sampling of the one shown in FIG Vibration to be obtained;

Figure 7 is a block diagram of an embodiment of the invention employing digital storage;

909836/0785

-7-

Pig. 8 is a block diagram of an embodiment of the invention using analog storage; and

9 and 10 are block diagrams of further exemplary embodiments the invention.

In Fig. 1, formants 103, 107 and 109 are for the spoken word "WED" is represented as it would appear on, for example, a spectrogram; the solid lines give it the middle of the existing frequency bands. The lower frequency formants 107 and 109 between the zero point and the abscissa value 101 form the phoneme / W /, which in the word "WED" consists of two frequencies, both of which rise in the time interval 115. The vowel sound of the phoneme / EH / consists of three formants 103, 107 and 109 between the abscissa values 101 and 105. The Final consonant / D / occurs after the vowel after a short pause. For even understandable language is it is necessary that the formants of two adjacent phonemes at the junctions between the phonemes be continuous merge. The dashed line 101 in FIG. 1 denotes such a transition, namely the transition between the phonemes / W / and / EH /, the formants 107 and 109 continuously merge into one another and are placed over the junction continued without interruption.

Figure 2 is a similar representation of a spectrogram of the spoken word "WADE". The phoneme / W / consists of two lower frequency formants 207 and 209 in the time interval 215 between the origin and a dashed line 201 parallel to the ordinate. The phoneme / AY / consists of the three formants 203, 207 and 209 between dashed lines 201 and 205. The formants of the phonemes / W / and / AY / go at the junction indicated by the line 201 continuously merge into one another. A comparison of Figures 1 and 2 shows that the phoneme / W / in the word "WED" occupies a time span 115 which is longer than the time span 215 which the phoneme / W / in the word "WADE" lasts. In Fig. 1, the two lower shapes are also

909 836/0785

manten 107 and 109 of the phoneme / EH / at higher frequencies than the two lower formants 207 and 209 of the phoneme / AY / in Fig. 2. The formants of the phoneme / W / in the word "WADE" according to Fig. 2 are similar to the Formants of the phoneme / W / in the word "WED" according to FIG. I if one considers the same period of time. The main difference between the two phonemes / W / _λ is that this phoneme in Fig. 2 ends earlier than in FIG. 1.

Fig. 3 shows two phonemes which are not connected to one another but separated by a time interval. The in The sound shown in Fig. 3 thus consists of two complete phonemes, which are pronounced for themselves and separately.

If the two phonemes shown in Fig. 3 are to be connected in the artificial speech generation, there would obviously be a discontinuity at the junction, if the temporal end 311 of the first phoneme were allowed to coincide with the beginning 312 of the second phoneme. The formants 303, 307 and 309 of the first phoneme would end suddenly and the formants 313> 317 and 319 would follow immediately at frequencies different from those of the first formants distinguish, begin. Such frequency hops would create distortions which would severely affect the intelligibility of the artificially generated speech.

If the formants at the beginning of the second phoneme in FIG. 3 are lengthened by lines 327 and 329, these intersect Lines formants 307 and 309 of the first phoneme. The den Point in time 321 corresponding to intersection points is a point at which the first phoneme can be broken off by a continuous one transition of the formants of the first phoneme into those of the second phoneme to reach. If the points of intersection of the respective formants do not coincide in time, the beginning of the second Phonems changed to do it. Smaller discontinuities are permissible, so that the points of intersection are not exactly timed have to coincide.

If the first phoneme in Fig. 3 at time 321,

909836/0785

which corresponds to the described point of intersection, is aborted and this point in time with the beginning 312 of the second If phonems can coincide, there is a steady transition of the formants. This is in Pig. 4, in which the formants 407 and 409 run steadily at transition 421, the combined Phonemes begin at a point in time 401, which corresponds to the point in time 301 of the start of the first phoneme in FIG. 3, and the combined phonemes end at point in time 405, which is earlier than the end 305 of the second phoneme in FIG. 3.

A continuous sequence of phonemes can be put together in paired groups in the manner described above, to generate artificial speech. One way of creating such a sequence of phonemes is to use individual To store phonemes on a magnetic tape, cut off the beginning and the end of each phoneme and the resulting ones Assemble pieces of tape in a desired order. When you play such a tape you get an understandable one Language.

Another method of implementing this technique is to store the individual phonemes in such a way that access to a desired phoneme, clipping of the phoneme, and playback in a predetermined order is possible, for example under the control of a program-controlled computer system, in order to generate the desired spoken information.

In one embodiment of the present invention are the phonemes stored digitally by periodically? Samples the amplitude of the phonemes representing each Vibrations are picked up and the amplitude values of these samples are converted into binary numbers. The binary numbers obtained are then stored in sequence for each phoneme.

Fig. 5 shows two periods of a typical Schwin- mnv ,. The line 501 shows the time course of the amplitude of the oscillation between the origin and the end 503 of the

909836/0786

ßAD ORiGlNAL

-10th period. From the end of 503 to the end of the first period 505 of the second period, line 501 then has the same again Course. If the amplitude of such an oscillation is sampled periodically with a correspondingly high frequency, a sequence of numbers can be obtained which gives a very good approximation of the amplitude curve of the original oscillation. The more samples are taken during an oscillation period, the more accurate the reproduction.

FIG. 6 shows an example of the result of scanning the vibration shown in FIG. 5. Each sample 6OI corresponds to an instantaneous value of the continuously changing amplitude of line 501 in FIG. 5 to a corresponding one Time.

7 shows an exemplary embodiment which operates according to the method explained above. A phoneme goes into a microphone 701 spoken and the received electrical signal is amplified by means of an amplifier 703. The output of amplifier 703 is the input of a gate circuit 705, which is periodically gated by an oscillator 711 will. The output signal of the gate circuit 705 consists of a series of pulses whose time interval is reversed is proportional to the oscillation frequency of the oscillator and its amplitude to the respective instantaneous value of the output signal of amplifier 703 corresponds.

The pulse sequence shown in FIG. 6 would therefore correspond to the output signal of the gate circuit 705 if the The output of the amplifier 703 is that shown in FIG Course would have.

In the present exemplary embodiment, the oscillator frequency is 14 kHz. The period of the output pulses gate circuit 705 is therefore about 71.5 ps. Any phoneme 4096 memory locations are assigned. The maximum allowable length of a phoneme is therefore 292.8 ms. Phonemes that have a duration longer than this maximum value are paused sounds,

909836/0785

thus periodically, and can be queried several times if necessary will. All aperiodic phonemes are shorter than the maximum value given above.

The output of amplifier 703 also becomes rectified by a diode 712, smoothed to a direct current by a low-pass filter 709 and a follower control device 715 supplied. The control device 715 has a switch 725. When the switch 725 is in the illustrated INPUT position, the lack of a DC output signal from the low pass filter 709 causes the controller 715 to a memory address register 714 of a digital Store 710 supplies the address for the beginning of the next phoneme. The start address and the end address are from a Reading device 719 is provided which is known from any known Device can exist, e.g. a punched tape reader, card reader, magnetic tape reader, etc. One can also use an electronic calculator or other more complicated control device. The end address is in a register stored in the controller 715.

The memory address register 714 consists of registers which can be set to the successive addresses by the output signal of the oscillator 711. The output signal of the oscillator 711 is fed to the registers via a gate circuit 727, which is gated on by the control device 715 while a phoneme is being stored.

The output of the gate circuit 705 is through an analog-to-digital converter 707 is converted into a binary number. The amplitudes of the pulses can, for example, in 128 Sections are divided. Each amplitude value can then be represented by a binary number of seven bits; these Binary numbers start with the value 0 (0000000) and end with the Maximum value (1111111). The AC voltage zero value of the output signal of the amplifier 703 is accordingly 64 (10000OGQ

909836/0 785

In practice, the AC voltage zero value is assumed to be offset by approximately 5%. The direction depends on the number of phase reversals in the amplifier. The reason for this offset is that the amplitude of the sound vibrations caused by the expulsion of breath is greater than that caused by the effects of the muscles in the larynx.

The seven-digit binary number from the converter 707 \ d.rd stored in a memory output register 713 to be stored by restoring into memory what takes place by a clock signal which is fed from the oscillator 711 via the gate circuit 727. If the phoneme in all it has been stored in the memory locations assigned to it in the form of such a sequence of digitally coded signals, the storage of the current phoneme is considered finished. The termination of the storage is indicated by a comparison circuit 717 determined, which supplies a corresponding signal to the control device 715 if it determines that the in the Control device 715 stored end address matches the address in the memory address register 714.

When all the phonemes necessary to reproduce the speech have been stored in memory, will the switch 725 in the lower playback position in FIG. 7 knocked down.

A correspondingly prepared recording medium is inserted into the reading device for the artificial generation of speech 719 used. The record carrier contains two binary numbers for each desired phoneme, the first of which is the starting address and the second indicates the ending address of the phoneme in memory. The starting and ending addresses are not required the first and the last memory location address of the relevant stored phoneme. For each phoneme needed two such addresses appear and the phonemes follow one another in such a way that the desired language information results.

909836/0785

The start and end addresses are chosen so

1. the binary number stored in memory the value 1000000 and that

2. The formants in the phoneme are at frequencies that have a continuous transition with the frequencies at the start address at the end of the preceding phoneme and at the end address there is a continuous transition with the frequencies of the following phoneme guarantee.

The reading device 719 can be a program-controlled digital data processing system and the addresses can form the result of calculations in the data processing system, so that the resulting language information about delivers the result of the invoices in the data processing system.

The slave controller will be under control of a command 715 the first two addresses supplied by the reading device 719. The start address is in the address register 714 saved.

The addressed phoneme information is read out of the memory and stored in the output register 713. A digital / analog converter 721 is connected to the output register and supplies an output voltage and its amplitude determined by the value of the binary number fed to its input will.

The memory address register 714 is then through the pulses supplied by the oscillator 711 via the gate circuit 727 are incremented and switched accordingly the following binary numbers are read out from the memory 710 and fed to the input of the digital / analog converter 721 The amplitude of the output voltage of the converter 721 changes continuously according to the values of the binary numbers fed to the input of the converter and it becomes an electroacoustic VJandler, e.g. fed to a loudspeaker 723, the

909836/0785

SAD OFUGiMAL

produces audible sounds.

The memory address register JIk is switched on until it has reached the second address, that is to say the end address that was stored in the control device 715. The match between these addresses is determined by the comparison circuit 717, which supplies a corresponding control signal to the control device 715.

When the control signal that reaching the end address indicates arrives, the control device 715 immediately takes the next two addresses from the reading device 719 on. The phoneme identified by these new addresses is then read out of the memory and written to in the Way converted into an audible sound.

The successive phonemes are processed in the manner described until the control device 715 receives a halt signal from reader 719 or other external source.

Fig. 8 shows an embodiment of the invention in which no digital storage device is required.

The device of Fig. 8 includes a drum 801 on which the various phonemes are each on separate Tracks are stored in analog form. A combined reading / speaking head 803 is assigned to each track. The drum also contains a clock track that is assigned to its own head 805 anyway.

In the synthetic generation of speech, a reading device 809 supplies two addresses to a control device 807. The addresses are decoded in the control device and the output of a corresponding header 803 is via an amplifier in device 307 with a headset 313 connected, which is used to record analog information on a magnetic tape 821.

By using an Inde :: mark and that of the control device 807 from the clock track head 805 supplied clock pulses

C B Ϊ?, S / t;

BATH ORIGINAL

becomes the correct starting position of the addressed phoneme of the drum 801 is determined.

When the starting point is set, a drive motor 811 for the tape 821 is operated by the control device 807 switched on. The control device 807 then effects the transmission of the information to the headset 813 so that the selected Part of the addressed phoneme is recorded on the tape 821 now moving. If the ending address is replaced by the Control device 807 is perceived with the help of the clock pulses from head 805, the motor 811 and thus the tape stopped. The transmission of the signal from the selected head to the speaking head 813 is also blocked.

Then the next two addresses are fetched from the reading device 809 and the next phoneme is processed in a corresponding manner.

When the last phoneme has been processed, the tape 821 is read by a signal from the reading device 809 or from an outside source. Then the belt is driven in the forward direction again, the recorded speech signals are scanned by a reading head 815, amplified in an amplifier 817 and transmitted through a speaker 819 or the like.

The last embodiment described can be refined in various ways. For example, the The direction of rotation of the drum 801 is reversed and the phonemes are retrieved in the reverse order. When playing needs then the tape does not get rewound. Besides that, it can Query the phonemes from the drum and record them on the tape at a later playback speed exceeding Speed.

FIGS. 9 and 10 show a further installation in which the invention can be applied. The first thing to do is Loading of the system are described.

9 0 9 8 3 6/0785

BATHROOM OBlGTMAL

The one in Pig. The device shown in FIG. 9 contains an audio signal source 701 'which is connected to an input of an analog / digital converter 703'. The other input of this converter is connected to a clock signal source 705 'and the output signal of the converter accordingly reproduces the instantaneous values of the amplitude of the audio signal at the time of the clock pulses. The pulse amplitudes can be divided into 128 sections, for example. Each size value can be represented by a seven-digit binary number that goes from the minimum value 0000000 to the maximum value 1111111. The AC voltage factor corresponds to the decimal value 64 or the binary number 1000000. In practice, the reference level, as already explained above, is offset by 5%.

The signal from the clock signal source 705 'causes, under the control of a bistable circuit 709' * which in turn is gated by memory command pulses from a source 707 ', firstly the switching of a tactile address counter 711' via a gate circuit 713 'and secondly that a core memory 17 via a gate circuit 215 'are supplied with memory command signals. Each pulse from the clock signal source 705 'advances the address counter 711 by the address of a memory location, causes the output signal of the converter 703' to be fed to the core memory 17 and delivers a memory command _{pulse via the gate circuit 715}} which causes the output signal of the converter 703 'is stored in the memory location of the core memory 17 designated by the address counter 711'.

In this way, consecutive digital Signals are stored that represent the phoneme to be stored. When the address counter 711 'is at the highest address switches in memory, indicating that the memory capacity is exhausted, the bistable circuit 709 'is again switched. The gate circuit 713 'is blocked

90 9836/0785

-17-

and prevents the address counter 711 'from advancing. the Gate circuit 715 'is also disabled, so that the core memory 17 no further spear commands can be supplied. The state of the bistable circuit 709 'after the described second switchover enables a second bistable to be switched over Circuit 719 '. The second bistable circuit 719 'is toggled after being through the first index pulse of a storage drum 21 has been made responsive. By switching over the second bistable circuit 719 ', a Gate circuit 723 'made ready to pass at the other input a sector clock signal from drum 21 is present. A sector clock signal occurs for each digital signal to be stored on the drum 21 on. Each of the digital signals stored in the core memory 17 exists in the present embodiment from seven binary digits. The seven binary digits each Signals are stored in parallel, i.e. at the same time, in the core memory 17 or read from it. On the drum 21 the seven binary digits of each signal are serialized, i.e. saved one after the other. The sector clock signal from the drum 21 advances the address counter 711 'by one memory location and supplies a control signal (read command) which causes the core memory 17 to send a digital signal to a parallel / Serial converter 725 'to be delivered. These two punctures of the sector clock signal are only carried out if the gate circuit 723 'is permeable through the second bistable circuit 719' has been made. Another puncture of the pulse-shaped Oektortaktsignales consists of the one from the core memory 17 to pass the signal read out to the parallel / serial converter 725 '.

The parallel / serial converter 725 'only exists from a seven-stage shift register into which the output signal of the core memory 17 is stored in parallel, while the output signal from the last stage in which the stored Signals are shifted in sequence, is picked up. For every bit that is transferred from the converter 725 'to the drum 21

909836/0785

BAD CRIGfNAL

-18-

is transmitted, the drum 21 delivers a clock pulse. For each digital signal that is read out from the memory 17 if j seven clock pulses are required from drum 21, in order to store the seven binary digits one after the other on the drum 21. For every seven clock pulses a pulse occurs Sector clock signal.

The address counter 711 'used in the present embodiment works modulo 4096. "This means that the address counter 711 switches from the decimal number 4095 (corresponding to the binary number 111111111111) to 0 (in binary digits 000000000000). The loading of the core memory 17 is finished, x ^ When the address counter 711 'has reached the count value 4095. The first sector clock signal from the gate circuit 723' then switches the address counter 711 'to 0, so that the extraction of the successive digital signals begins at the first address of the core memory 17. The number of digital signals that from the core memory 17 via the converter 725 'is transferred to the drum 21, may be smaller than the 4096th It is therefore necessary to address counter 711' ^f by the storage command pulse from the source 707 reset prior to the loading of the core memory 17th

The transfer from the core memory 17 to the drum 21 continues until a further index pulse from drum 21, indicating that the drum is a complete Has carried out rotation, the second bistable circuit 719 'switches over. The gate circuit 723 'is blocked and prevents the address counter from advancing and further command signals from being fed to core memory 17.

The digital signals, each consisting of seven binary digits and representing a phoneme, are thus recorded in series on a track of the drum 21 while it is making one revolution. Further phonemes are stored on other tracks of the drum 21 with other heads ₃ which are distributed in the axial direction along the drum surface. In the present example there are 128

BAD CRlGiNAL

such data tracks exist. The index sector and clock pulses are each recorded on a separate track. There are seven clock signals between the sector signals on and between the index pulses, which occur once during each revolution, there are approximately 4000 sector pulses. Every Track contains a single phoneme. During loading, the tracks can be selected manually, for example with Help from switches. By selecting one of the data heads, the associated phoneme can be recorded and later retrieved.

After all the phonemes to be used have been recorded, a desired sequence of phonemes can be added to the synthetic Speech generation can be read from the drum and alternately fed to one of two core stores. the Phonemes are then extracted from the core memories in the same order, clipped to make a steady transition from the Ensure sounds, and translate them into audible sounds. The transfer of a phoneme from the drum to a core memory and the removal of a phoneme from the other core memory for conversion into an audible sound takes place at the same time. How this is done in detail will now be explained with reference to FIG.

Each phoneme to be reproduced is assigned three numbers. The first number indicates which track of the drum 21 read, i.e. which phoneme should be selected. The second number gives the starting point and the third number the end point at. The second and third digits supplied to indicate the beginning and the end of the selected phoneme, are delayed until the selected phoneme has been read from the drum in the manner to be described below is. In the embodiment shown, these numbers can be controlled by a manually controllable device 831 or by a punched tape reader 333 can be entered. Of course these numbers can also be supplied by a more complicated arrangement, e.g. by a digital data

909836/0785

BATH ORIGINAL

processing plant. The start and end points are chosen so that the resulting clipping of the respective phoneme ensures a steady transition with the preceding or following phoneme ♦ The start and end addresses are like this chosen that

(1) the value of the binary number deviates by no more than 5% from 64 (1000000), and that

(2) the pormants in the phoneme have frequencies that at the beginning address continuously into those at the end of the preceding one Phonems, and at the end address they continuously change to the one at the beginning of the following phoneme.

The first number is obtained from the hand control device 831 or the paper tape or punch tape reader 833 a phoneme selection register 835 supplied. The second number becomes a read counter via a starting position intermediate register 839 837 supplied. The third number is fed to a stop or end point register 841.

The phoneme assigned to the number of start and end positions in registers 839 and 841 is read by drum 21 read and stored in one of the two core memories 17 or 827.

Two processes take place at the same time: First, is one phoneme is transferred from the drum 21 to the core memory 17 or 827 and, secondly, the phoneme is made from the other Core memory read out and converted into an audible sound. Which register for the first and which for the second operation is involved is determined by a bistable circuit 843. For explanation it is assumed that the A output 847 of the bistable circuit 843 is positive ("correct") and the B output 845 is negative ("incorrect"). Purely the operations to be described it doesn't matter which outcome is positive (correct) first.

The first operation in which a phoneme is transferred from the drum 821 to the one core memory is in progress as follows. The track to be read of the drum 821 is through

909836/0785

the phoneme selection register 835 is determined. An index pulse from the Drum 21 sets a write counter 811 to O. The binary digits of the phoneme signal from the selected head of the drum 21 are fed into a serial / parallel converter with the aid of the clock pulse track 849 headed. The parallel output of converter 849 appears on seven output lines that connect to the core memories 17 and 827 are connected. The positive A output 847 of the bistable circuit 843 causes the binary digits from the converter 849 are stored in the B-memory 827, in the gate circuits 851 and 853, which are used for storage in the memory 827, made ready for passage. The address for storing each character consisting of seven binary digits is the core memory 827 from the counter SU via an open gate circuit 851 supplied. Each seven clock pulses from the drum 21 are temporally accompanied by a sector clock signal that the Core memory 827 is fed through an open gate circuit 853 to the storage of the seven binary digits from the converter 859, the counter 811 is also activated by the sector signal moved one step further. The successive seven-digit binary phoneme signals are thus from of the drum 21 is stored in the core memory 827 via the converter 849. The counter 8II switches back from 4095 to 0, so that a maximum of 4096 characters can be transmitted. When the drum 21 has made one complete revolution, they are all Characters representing a complete phoneme are transferred to the B-memory 327. The counter 811 is activated by an index pulse is reset to 0 and, if the state of the bistable circuit 343 does not change, the same sequence of characters retransmitted without changing the contents of the B-memory 827.

The second operation running with the above-described first operation Clcichzeitic, in which the signals auο read out the other core spreader and into an audible one To be implemented aloud, proceeds as follows:

The positive Λ output 047 of the bistable circuit 343 also makes gates 855, 357 and 859 through

909836/0785

BATH

ready to leave, those for reading out of the A-Kernspebher 17 are provided. The address from which the characters are to be read out of the core memory 17 is determined by the read counter 837 via the opened gate circuit 855. The original setting of the 837 counter was made from the outside, either by means of the manual control device 831 or the paper tape reader 833. The clock pulses for reading out from the A-core memory 17 are provided by a 14 kHz oscillator 861 is supplied via the opened gate 859. The clock pulses also switch the read counter 837 by one place Further. The output signal of the core memory 17 is a seven-digit binary character that is sent to the input of a digital / analog converter 863 is supplied via the opened gate circuit 857. The output of the digital-to-analog converter 863 is a continuously changing electrical signal, the amplitude of which is determined by the digital input signal. That continuously changing output signal of converter 863 is amplified by an amplifier 865 and converted into an audible sound by a transducer, e.g. a loudspeaker 867. Phoneme signals are read out one after the other from the core memory 17 until the number in the read counter 837 equals the Number in the terminal register is 841. The agreement of these numbers is determined by a comparison circuit 869, the output signal of which the bistable circuit 843 switches over and causes reader 833 to read the second and third numbers for the phoneme that has just been transferred from drum 21 to B-core memory 827 is transmitted and to supply the first number for the next phoneme to be transmitted in this way.

By switching the bistable circuit 843, the B output 845 , which was previously negative (“false”), becomes positive (“correct”) and the A output 347, which was previously positive, becomes negative.

The track of the drum 21 selected by the phoneme selection register 035 is now read, the signal read becomes the series / parallel. ^1st - converter 849 eL ^ preul - .. - ■: I del above

O ft O w '"-" ~ -' ■ ft '' i Q c-

ÖAD ORiGfNAL

timing control described above, and the output signal of the converter 849 is fed to both memories. Of the positive B output 845 of the bistable circuit 843 opens the Gates 871 and 873, which are provided for storage in the A-core memory 17. More specifically, the sector clock pulses supplied via the open gate 871 and the address via the open gate 873 from the memory counter 8II. The corresponding Gates 835 and 851 of the D-core memory 827 are now blocked because the A output 847 of the bistable circuit 843 is negative. The first operation is now carried out with the other core memory.

The second operation is also carried out with a different core memory as before, since the positive B output 845 of the bistable circuit 843 opens the gate circuits 875, 377 and 879, which are used to query the B core memory 327. The memory 827 receives the clock pulses from the oscillator 861 via the open gate circuit 875 and the address from the read counter 337 via the opened gate circuit 377. while the output signal of the core memory 827 is fed to the input of the digital / analog converter 863 via the open circuit 879. The corresponding gate circuits 855 , 859 and 857 of the A core memory 17 are blocked because the A output 847 of the bistable circuit 843 is now negative.

The alternating querying and saving is continued with the two core memories until all the desired phonemes. that have been chosen have been converted into audible sounds. In the embodiment described, the drum 21 runs at a speed of 1300 rpm. One revolution of the drum is required to transmit an entire phoneme. The transfer of a phoneme from the drum to the core memory therefore takes approximately 34 ms. In addition, a latency period, that is to say a viarte period, is provided for the index pulse, which is approximately 34 ms . The frequency with which the information is read out from the other core memory is 14 kKz, so that in each 7I ₃ 5 one character

909836/0785

BATH ORIGINAL

is read out. The transfer of a phoneme into a core memory thus requires a period of time in which about 950 characters can be read from the other core memory. That The maximum capacity of the core memory is 4096 characters each, but that many characters are never queried because the Phoneme at the beginning and at the end is trimmed. However, the number of characters read out always exceeds the minimum, that is necessary to feed the other core storage from the drum. The period of time required for setting a new starting point is so short that there is no discontinuity can be perceived in the sound produced. The output signal of the digital / analog converter 863 is retained as long as that minor discontinuities that could otherwise occur under certain circumstances are avoided.

The numbers supplied by the punched tape reader 833 or a corresponding control device, such as a digital computer system ₉ , are selected so that the phonemes generating the artificially generated speech are selected in the correct order and trimmed in the manner described for optimal understanding.

In a further possible embodiment of the invention that contains only one core memory, the Drum read only that part of the various phonemes that is to be reproduced, and all the phoneme parts read off are stored in a large core memory as standard. The query from the drum and the transfer to the core memory then starts at the first or lowest core memory address. When the last or highest core memory address is reached, the transmission starts again with the first address. Reading out the digital signals to be converted into analog signals the core memory starts with the first address and continues up to the last address, whereupon the reading out starts again the first address begins. A suitable arrangement allowing reading of a phoneme from the drum for a certain number

prevented by drum revolutions, ensures the correct time relationships.

909836/0785

Claims (9)

Claims 1. A device for synthesizing audio information, in particular human speech, characterized by a memory (710., 801 _s 21) for storing coded phonemes, a control device (715, 807) for extracting the coded phonemes in a desired order the memory, and a device (721; 815, 863) converting the extracted phonemes into continuous electrical signals. 2. Device according to claim I _5, characterized in that the phonemes are stored in the memory as digitally coded signals. 3. Device according to claim 1 or 2, characterized in that the control device when reading out the individual phonemes or parts of these, the respective start and end positions of the relevant Part of the respective phoneme in the memory denotes which places a continuity of the language with the before and after Ensure phonemes removed from the memory. ¹ L device according to claim 1, 2 or 3, there · - characterized in that the Umsetzvorrichtuni; a second memory for the temporary storage of more than one extracted coded phoneme and that the control device causes the reading of the phonemes stored in the second memory for conversion into continuous electrical signals. 5. Device according to claim 4, characterized characterized in that the second memory has two Storage units (17, <°> 27) contain / ilts the respective signals that form a complete phoneme., able to store, and that the control device (3 ^ 3) the transmission of desired sequences, of signals representing complete phonemes and of 909836/0785 BATH ORIGINAL have been removed from the first memory (821), alternately in the first and second storage units (17 »327) and ' that the control device also causes the two memory units to alternately receive digitally encoded phoneme signals in the order in which the phonemes were extracted from the first memory. 6. Device according to claim 3, characterized in that a second converter is provided for the storage of phonemes in the first memory for obtaining electrical signals from phonemes and that the control device directly causes the signals obtained in the memory in sequential order and in groups, which can be addressed as individual phonemes, are stored, the groups being stored in any predetermined order. 7. Device according to claim 6, characterized in that the second unisetting device a gate circuit (705) for sampling the amplitude of a continuous electrical signal obtained from phonemes at suitable time intervals and a coding device (707) for converting the periodic values obtained during the sampling Amplitude samples in digitally encoded signals for storage in memory (710). 8. A method for the synthetic generation of audio information, in particular human speech, characterized in that signals are obtained represent the phonemes which occur in the artificially generated audio information that the individual Signals representing phonemes are stored in locations on a storage medium in such a way that these signals can be extracted is possible, starting at any given place and ending at any subsequent predetermined place that from the storage medium determined PgI .-. o "of signals can be taken, which show a desired sequence in front of you. 909836/0? BAD ORlGiN ¹ AL -27- represent, whereby the gain of the individual phonemes representing Signals begins and ends at special points, which a continuity of the pormanten over the junctions between ensure successive phonemes in sequence, and that the extracted phonemes are converted into audible sounds. 9. The method according to claim 8, characterized in that that the phonemes are represented by digital signals. J 909 8 3 6/0785