EP0157903B1 - Method and apparatus for speech synthesizing - Google Patents

Abstract

1. Process for the synthesis of speech in which combinations of a consonant followed by a vowel are stored and read out as required, characterised in that - every consonant occurring in the language is stored together with a weak uniform vowel "&", - all the vowels occurring in the language are stored individually, - the desired consonant/vowel combination is formed by staggered reading out of consonant and vowel in two channels, with the uniform vowel "&" being masked by the vowel that is read out.

Description

The invention relates to a method for speech synthesis and an arrangement for carrying it out.

It is known to individually store the phonemes (consonants and vowels) occurring in a language and then to read them out sequentially as required. Since the number of phonemes is relatively small, it is possible to work with relatively little storage capacity. It is disadvantageous, however, that a sound that deviates greatly from natural language is produced because a "click" sound can be heard between successive consonants and vowels, so that this principle is practically not applied.

In order to produce a synthetic language which is largely approximated to human language, the procedure is therefore that combinations formed from the occurring consonants and the occurring vowels, so-called di-phonemes, are stored and read out. It is understood that a very considerable storage capacity is required for this, which for an Indo-European language is of the order of about 500 or more di-phonemes. Accordingly, the use of the method is limited to those cases in which considerable effort is justified.

Based on the last-mentioned method, the invention has for its object to significantly reduce the amount of memory, so that the method even for mass consumer goods, for. B. toys, (dolls' voices and the like) is applicable.

Claim 1 defines the method according to the invention. It can be seen that each consonant is only stored in combination with a single unit vowel, which is to be referred to here and below as "&", and which has approximately the "e" in the German word "all", in the English word "the", in the French word "le". It has been shown that when reading out with a corresponding time delay, the & is masked completely or at least to such an extent by the actual vowel then reproduced that the resulting syllable largely comes close to natural pronunciation.

In the usual digital storage of the phonemes, it is expedient to store a command in the consonants in the correct place, which initiates the beginning of the reading of the vowel memory. Furthermore, it can be expedient to use different filters to amplify or attenuate the frequencies of the consonants on the one hand and the vowels on the other to improve the masking of the &.

Claim 5 defines an arrangement according to the invention with which speech synthesis can be carried out.

• Fig. 1 zeigt anhand eines Beispiels das Prinzip des Verfahrens,
• Fig. 2 zeigt Frequenzgänge von Filtern für Vokale und Konsonanten,
• Fig. 3 zeigt schematisch ein mögliches Speicherformat für Konsonanten und Vokale,
• Fig. 4 zeigt eine bevorzugte Hüllkurve für die Konsonantenerzeugung,
• Fig. 5 ist ein Blockdiagramm einer Anordnung zur Ausführung des Verfahrens, und
• Fig. 6 ist ein Diagramm zur Darstellung des Zeitablaufs bei der Synthese eines einfachen Wortes.

The invention will be explained in detail below with reference to the accompanying drawings.

• 1 shows an example of the principle of the method,
• 2 shows frequency responses of filters for vowels and consonants,
• 3 schematically shows a possible storage format for consonants and vowels,
• 4 shows a preferred envelope for the consonant generation,
• 5 is a block diagram of an arrangement for carrying out the method, and
• Fig. 6 is a diagram showing the timing of the synthesis of a simple word.

Since the reading process is delayed, this means that the reading of a vowel already begins while the reading of the consonant di-phoneme (namely its & part) is still in progress, one works with two reading channels. In Fig. 1 the upper diagram shows the envelope curve of the consonant channel, the lower one that of the vowel channel, the simple word "DATO" being chosen as an example. It can be seen that the & parts of the consonant channel and the vowels are reproduced at the same time, and the weak & sound is strongly masked by this alone. This masking can also be supported by further measures.

Fig. 2 represents a first means for this. It is known that the frequency spectrum of the consonants and vowels is different; e.g. For a male voice, for example, the maxima of the consonants are in the range of approximately 600 ... 3000 Hz, the vowels in the range of approximately 200 ... 1000 Hz. Accordingly, filters with the passbands shown in FIG. 2 are assigned to the two channels , where the filtering can be done either during recording or during playback.

Fig. 3 shows schematically the format for the storage. When recording, the sounds are digitized, that is, with a clock of z. B. 10 KHz or more amplitude sampled and the data thus obtained are stored in successive memory locations for serial readout. However, two memory locations for command data are kept free, namely a "continue" command and a "end" command. The "continue" command means the point in time at which the other channel is to continue reading; this command is in the consonant data when the actual consonant sound changes to the & part, while in the vowel data it is close to the end of the data string. The command "end" goes without saying, but is necessary because the individual phonemes have different durations. The "continue" command can be used to increase the masking effect by using its occurrence the envelope of the channel just read is damped, as indicated in Fig. 4, for which one can use a conventional analog attenuator consisting of diode, resistor and capacitor.

Fig. 5 shows in block form an embodiment of a speech synthesizer which - as can be seen - is extremely simple. The phonemes to be reproduced are selected by external means, for example a microprocessor, and do not form part of the subject matter of the present invention; an external control circuit is therefore only indicated here as block 1.

The arrangement comprises two mutually identical channels, only one of which is described below.

A memory address counter 2 is set by the control circuit 1 to a specific phoneme start address. A phoneme memory 3 contains all the phonemes required for a given language, with thirty-six phonemes being sufficient for many languages. After filtering during recording (as explained above), the phonemes are digitized and stored in the format shown in FIG. 3; For example, the codes "0" or "1" can be reserved for the commands "Continue" or "End". A clock generator 5 generates the read clock of z. B. 10 KHz, for both channels. The read data go to a decoder 4, which determines whether it is data or one of the commands "continue" or "end". Data pass through a digital-to-analog converter 6 and a multiplier 7 to a summing element 8 and from there to an amplifier-loudspeaker unit 9.

When the "End" command is decoded, the incrementation of the address counter 2 is blocked via an AND gate 10.

If the command "further" is decoded, a phoneme request flip-flop 11 is set for the other channel; it is reset by the external control circuit when the next start address is entered. Furthermore, when the command "continue" is switched a damping flip-flop 12, which connects an envelope generator 13 with its output F to the one, with its output F to the other channel, which acts on the multiplier 7, so that the output of the channel in question is gently damped without, however, the "click" noise. The outputs of both channels are combined in the summing element 8.

The respective set state of the flip-flop 12 is also transmitted to the external control circuit in order to signal the latter which of the two channels can be occupied, for example at the beginning of a readout cycle after the circuit has been started up.

Before a synthesis process is explained in detail with reference to FIG. 6, possible modifications of the block circuit shown in FIG. 5 are pointed out.

The memory expenditure can be halved if only one phoneme memory 3 is provided for both channels and the reading is done in time division multiplex. The multiplier 7 is already included in certain commercially available digital-to-analog converters, so that the output of the envelope generator 13 need only be connected to the corresponding input of the converter. The circuit can also be largely implemented in a microprocessor, in which case either the two envelope generators and the two converters remain outside or only a single, common converter, while all other processes are carried out digitally by the microprocessor.

• - in Zeile (a) den Takt des Taktgenerators 5; dieser Takt kann starr sein, kann aber auch von der Steuerschaltung 1 variiert werden, um eine der natürlichen Sprache noch ähnlichere Phrasierung zu erzielen,
• - in Zeile (b) Formate aus dem ersten Kanal, hier die Phoneme "D&" und "T&",
• - in Zeile (c) den Logikpegel am Ausgang des Flipflop 12,
• - in Zeile (d) den Logikpegel am Ausgang des Flipflops 11 des zweiten Kanals,
• - in Zeile (e) Formate aus demselben zweiten Kanal, hier Phoneme "a" und "o",
• - in Zeile (f) den Logikpegel am Ausgang des Flipflops 11 des ersten Kanals,
• - in Zeilen (g) bzw. (h) die Hüllkurven, erzeugt von den Hüllkurvengeneratoren 13 des ersten bzw. des zweiten Kanals, und
• - in Zeilen (i) bzw. (k) die analogen Ausgangssignale des ersten bzw. zweiten Kanals; dabei sind die Hüllkurven nicht als repräsentativ für die tatsächlich erzeugten Laute "D", "A", "T" oder "0" zu verstehen; das Diagramm dient nur der Erläuterung des zeitlichen Ablaufs.

Fig. 6 shows

• - In line (a) the clock of the clock generator 5; this clock can be rigid, but can also be varied by the control circuit 1 in order to achieve a phrase that is even more similar to natural language,
• - in line (b) formats from the first channel, here the phonemes "D &" and "T &",
• in line (c) the logic level at the output of the flip-flop 12,
• in line (d) the logic level at the output of the flip-flop 11 of the second channel,
• - in line (s) formats from the same second channel, here phonemes "a" and "o",
• in line (f) the logic level at the output of the flip-flop 11 of the first channel,
• - In lines (g) and (h) the envelopes generated by the envelope generators 13 of the first and second channels, and
• - in lines (i) and (k) the analog output signals of the first and second channels; the envelopes are not to be understood as representative of the sounds "D", "A", "T" or "0" actually generated; the diagram only serves to explain the chronological sequence.

Claims (8)

1. Process for the synthesis of speech in which combinations of a consonant followed by a vowel are stored and read out as required, characterised in that

- every consonant occurring in the language is stored together with a weak uniform vowel "&",

- all the vowels occurring in the language are stored individually,

- the desired consonant/vowel combination is formed by staggered reading out of consonant and vowel in two channels, with the uniform vowel "&" being masked by the vowel that is read out.

2. Process according to claim 1, characterised in that when the combinations of consonant and uniform vowel are stored vowel-typical frequencies are attenuated and that when the vowels are stored consonant-typical frequencies are attenuated.

3. Process according to claim 1 or claim 2, characterised in that the amplitudes of the uniform vowels "&" are attenuated during reading out.

4. Process according to claim 1, 2 or 3, characterised in that the reading out is effected at a variable pulse frequency.

5. Arrangement for the synthesis of speech, comprising a memory device for storing consonants and vowels and a reading out circuit for reading out combinations of a consonant and a following vowel, characterised in that

- every consonant occurring in the language is stored in the memory device together with a weak uniform vowel "&",

- all the vowels occurring in the language are stored individually in the memory device,

- the reading out circuit comprises two read-out channels that can be activated alternately and a switching over circuit that can be operated by a command read out in one channel for activating the other channel, and reads out the desired consonant/vowel combination from the two channels in a staggered manner, and that

- an envelope generator that can be activated by the switch-over command is provided for each channel, which generator effects attenuation of the amplitude in order for the uniform vowel "&" to be masked by the vowel that is read out.

6. Arrangement according to claim 5, characterised in that each channel includes a memory for all the sounds required (phonemes and diphonemes).

7. Arrangement according to claim 5 or claim 6, in which the sounds (phonemes and diphonemes) are stored digitally in each memory and can be read out sequentially, characterised in that at least in the case of the diphonemes the changeover command is stored in the read-out sequence between the consonant interval and the uniform vowel interval.

8. Arrangement according to claim 5 in which the sounds (phonemes and diphonemes) are stored digitally in each memory and can be read out sequentially, characterised in that at the end of each read-out sequence a command "end" can be read out, by means of which command a subsequent reading out operation can be initiated.

Images