DE2448909C3 - - Google Patents

Description

The invention relates to an electrical circuit arrangement of the type specified in the preamble of claim 1.

Speech recognition methods have already become known in which the ones obtained from speech analog electrical wave trains are evaluated. In DE-OS 14 72 038 such a method described in which both the zero crossings of the wave trains and the intervals of the zero crossings be determined. The intervals of the zero crossings determined during a certain period of time are sorted according to their length Then the number of intervals of the same length is determined, and based on the count result, the words are recognized via an allocation circuit.

With its disadvantages, this solution is representative of most of the known speech recognition methods, in which only a small number of words, mostly numerals and some commands, are processed because the The possibility of selecting the individual phonemes is too imprecise. To increase the vocabulary, would be a considerable amount of circuitry and storage space is required. But even then it is - without them Allowing errors to increase too much - only possible, voices of related characteristics and similar Recognize the speed of speech. This means that the procedures must be adapted to the respective speaker will.

In DE-AS 12 02517 a device is for automatic recognition of spoken syllables or words have been described in each one a measured variable exceeding a threshold value the occurrence or non-occurrence of a fundamental oscillation and a harmonic. The assessment is made with yes or no and is used as a code for classification of the examined sound is used in a sound group by means of a logic circuit. The phonetic group marking signals are arranged in the order of their appearance in a memory and after the end of the spoken syllable for their identification used.

This circuit arrangement also represents a solution. In other words, only a few words, e.g. B.

Numerals, can be recognized. The circuit arrangement does not provide a sufficiently precise way of distinguishing between words other than those selected to select.

To eliminate these disadvantages, a method has been developed in which the frequency spectrum in several frequency bands are broken down and the occurrence of formants in the frequency bands is determined DE-OS 15 47 027). In this process the respective end is "separated by formants" into many Frequency bands, determined by differentiating circuits In downstream interlocking circuits the formant ratios (rising, falling, unchangeable) for a large number of vowel vector sizes in the Language spectrum displayed. Furthermore, consonant criteria are in a frequency analyzer and based on of energy surges determined The formant energy and the formant criteria are determined in a consonant matrix stored The formants determined are also transferred to the consonant matrix via a form transmitter, so that a total of a large number of vector sizes - in the known embodiment 56 - be available.

With this speech recognition process it is possible to use a much larger vocabulary process, but this requires a circuitry effort that such a device for the makes practical use seem unrealistic

The invention is based on the object with little circuitry complexity from the Frequency spectrum of the speech signals to create criteria from which the spoken text - without Limitation of the range of vocabulary - can be recognized in a machine-evaluable manner.

This object is achieved by the invention specified in claim 1. Further developments are specified in the subclaims.

The idea of the invention is based on the knowledge that the essential for speech recognition Derive speech features from the energy density spectrum of the electrical speech signal. For the individual phonemes - that is, the smallest phonetic unit that unites in the same phonetic environment Difference in meaning - when speaking there are always three pronounced frequency focal points. These frequency emphases, called formants, differ little for different speakers from each other. Corresponding studies have shown that only the two formants with the lowest frequency are important for the intelligibility of a phoneme. The circuit according to the invention therefore allows each phoneme to be recognized with high reliability from these two formants alone.

The advantage of the invention is in particular that only very few components are required to to derive four criteria from the speech signals, with which independent of the speech speed and a reliable analysis of the amplitude of the speech signals is possible. These four criteria are sufficient for an arbitrarily large set of words

In the drawings, an embodiment of the invention is shown and will be explained in more detail below explained. Show it:

F i g. 1 shows a circuit arrangement designed according to the invention as a block diagram,

F i g. 2 to 7 the division of the voice band for formant recognition,

F i g. 8 a circuit arrangement for evaluating the criteria obtained according to the invention and

F i g. 9 and 10 curves for the selection of recognized syllables.

As shown in the block diagram of FIG. 1 can be seen, the spoken text of the circuit is offered via a microphone 1, which the sound vibrations in converts electrical signals The electrical signals amplified in an amplifier circuit 2 arrive to a first circuit part 10 of the circuit arrangement in which the lowest-frequency formant is determined is, also to a second circuit part 20, in which the formant of the next higher frequency is determined is, and at the same time to a further circuit part 40, in which speech pauses are automatically recognized.

To filter out the higher frequency of the two lowest frequency formants, this is done with a Bandwidth of approx. 200-2800Hz speech signal coming from the microphone in circuit 20 via a High-pass filter 11 passed in order to achieve a better emphasis on the higher frequencies. In one subsequent phase difference amplifier 12, the signal is phase shifted, and the highest formant is filtered out. Due to the size of the phase shift, the next higher formant can have a higher Frequency but much smaller amplitude can be safely suppressed.

The term "phase difference amplifier" denotes a circuit (also in the following) which consists of a differential amplifier, one input of which is the direct output of the preceding circuit and the second input of which is the output of the preceding circuit in a phase-shifted manner and which forms the frequency difference.

The output signal of the phase difference amplifier 12, which for voiced phonemes is a sawtooth Has course, is fed via a line 16 to a circuit 30 for vowel recognition Dem Phase difference amplifier 12 is followed by a pulse shaping circuit 13, for example from an overdriven phase differential amplifier and the sinusoidal signals in square pulses The signals representing these formants can be converted at the output 14 of the circuit part 20 be removed.

The analysis of the lowest frequency formant is achieved by frequency conversion in the circuit 10. The input signal coming from the microphone 1 is passed through a low-pass filter 3 to better emphasize the lower frequencies and in a downstream modulator 4 with carrier suppression with a A frequency of 1.2 kHz coming from a frequency generator 8 is mixed. Two are produced Mixing products, namely (denotes the frequency coming from the low-pass filter 3 with Fl) 1.2 kHz plus F1 and 1.2 kHz minus Fl. The product 1.2 kHz plus FI is suppressed by an active low-pass filter 5. From the remaining product 1.2 kHz minus Fl can be the - now reversed - lowest frequency as the highest frequency by a phase difference amplifier 6 are screened out.

The output of the phase difference amplifier 6 is in turn fed to the circuit 30 for vowel recognition via a line 15. Here, too, a pulse generator 7 ensures that the ji ⁿ u ^r , shaped signals are converted into square pulses.

Based on the F i g. 2 to 7 can be easily done recognize how the two formants are sifted out. In Fig. 2 shows the total Dragon band that the Circuit offered by the microphone. To the formant

recognition, however, only a part of this voice band is used, which is shown in FIG. 3 is shown. F i g. 4 shows the Sana used for the formant with the second lowest frequency, such as e «through the high-pass filter 11 of the circuit 20 in FIG. 1 is screened out. The low p: ^{> r} iUkt r 3 of the P "hold! 10 in I- i g. 1 da: the voice band for the formant of the lowest frequency (Fig. 5) is filtered out and then in the modulator 4 with carrier frequency suppression with a frequency of 1, The resulting mixed products are shown in Fig. 6. The two products 1.2 kHz plus the frequency band shown in Fig. 5 and 1.2 kHz minus the frequency band of Fig. 5 (reversed) can be seen 7 finally shows the formant band of the formant with the lowest frequency after the higher-frequency mixed product has been suppressed by the active low-pass filter 5.

Furthermore, in FIG. 1 shows the circuit 30 for vowel recognition, which is connected to the lines 15 and 16 of the two formant analyzers 10 and 20. For this purpose, use is made of the fact that vowels are voiced sounds and at the same time the formant of the lowest frequency exceeds a certain amplitude level. Each of these points alone is not sufficient for vowel recognition, since consonants can also be voiced (n, m) and, on the other hand, fricatives or explosive sounds have a high amplitude.

In the lower branch of the vowel recognition circuit 30, the formant of the lowest frequency is therefore derived from the Formant analyzer 10 of an amplitude test, in the upper branch the formant of the next higher frequency from the formant analyzer 20, subjected to a voicing test. The lowest frequency Formant arrives at an adjustable amplifier 17 via line 15 and is used in a subsequent Circuit 18 rectified and smoothed. The signal treated in this way is passed through a low-pass filter 19 and a pulse shaper circuit 21 is supplied to a conjunction 23 as a first input 22.

When checking for voiced sounds, it is assumed that the voiced sound has the period of "Pitch" frequency are excited and then have a sawtooth-like amplitude curve when it fades away, which is particularly pronounced in the higher formants The higher frequency formant is from Formant analyzer 20 via line 16 to a circuit 24 for rectifying and smoothing and then a Low-pass filter 25 supplied. The sawtooth pulses produced are transmitted by an amplifier 26 quadratic gain greatly exaggerated and to control a timing element 27 to bypass the Pulse pauses used After conversion in a pulse converter circuit 28, the signals form the second input 29 of conjunction 23. Here the conditions "amplitude" and "voiced" become The statement "vowel" combined, which can be taken from the exit 31 of the conjunction 23

Finally, the circuit arrangement also has the voice switch 40. This speech switch is required to recognize pauses in speech. An optimal line division of the amplifier 32 will result in enciohi, du ' ¹ also hissing and H3uch! .Iu; o a perfect switching through of the threshold value switch. On the off ^r . passage 36 of the voice switch 40, the corresponding signals can be picked up.

F i g. 8 shows a circuit configuration as an example in FIG according to which the criteria for speech recognition obtained from the speech signals according to FIG.

ιι · niint can be. From the formant recognition signals on lines 9 and 14, the vowel recognition signals on line 31 and the pause signals Speech curves are formed on line 36 which represent the features for the phonemes. The phonemes

is are in their transition behavior very strongly dependent on the neighboring phonemes, i. That is, the formant frequencies are different for the same sounds. Sc have e.g. B. Explosives (t, b, p, k) in connection with other vowels also have other formant frequencies. Around

2 (i can still reliably recognize the speech signals is deviated from the usual analysis of individual phonemes, and in the circuit according to FIG. δ become the phonemes into longer, artificial syllables independent of speaking time and volume (in the following

2 ^r > called the pseudosyllables). Syllables are formed that consist of vowels, consonants and pauses and the centers of which are consonants. A pseudosyllable can therefore consist of

Pause-consonant-pause
Pause consonant vowel
Vowel-consonant-vowel
Vowel-consonant pause

exist.

i "The output of the voice switch 40 is via a Inverter 37 adds two conjunctions 38 and 39 and forms their first input. The second The conjunction 38 is input with the lowest frequency formant from the formant analyzer IC

in and the second input of conjunction 39 with derr Formants of the next higher frequency from the formant analyzer 20 applied to the outputs of the Conjunctions 38 and 39 are fed to counters 41 and 42 in which the impulses of the two formants

^l 'i are counted during defined time intervals of 20 ms.

The counter readings are then stored in memories 45 to 48, separately Consonants and vowels. This distinction is made

■> "from the output of the circuit 30 for vowel recognition Their output signals operate switches 43 and 44 in such a way that when a vowel occurs, alsc Presence of an output signal on line 31 the switches 43, 44 in a the counter readings in the

• 3 stores 47, 48 are brought into a leading position while in the absence of a signal on line 31 it is in the position shown in FIG are located.

The counters for consonants of the frequenz

"<> lowest formants are therefore found in saliva 45 stored in memory 46 for consonants of the formant of the next higher frequency, while the counter readings for vowels of the lowest frequency formant in memory 47 and those for vowels de:

'■' ■ higher- frequency formants are stored in the memory 48. It has also been found to be useful that when the consonants are saved, the curve should be smoothed by averaging

between two successive counters'ηr: i'en can be appropriate.

The vowel counter readings stored one after the other in the memories 4 /, 1 * are each subdivided into a first and a second half so that double vowels (ei, au, eu) can be recognized. The outputs of these two memories are fed to a circuit 50 for forming vowel focal points, in which vowel points are generated in a manner described with reference to FIGS. 9 to 12 by averaging.

Likewise, the outputs of the memories 45, 46 for consonant counts of a circuit 49 for Consonant thinning supplied, in which it is connected to the output of the circuit 50 for the formation of vowel emphases combined and a treatment also described in connection with FIGS be subjected.

As a result, signals corresponding to the spoken syllables can be picked up at the output of the circuit 49, which are fed to a feature comparator 51.

in Fig. 9 shows a smoothed speech curve derived from the consonants and vowels of the spoken word Word "no" is formed by the fact that the output of the counter 41 for the lowest frequency formants and on the ordinate the output of counter 42 for the higher frequency Formant is plotted The division of the abscissa and ordinate corresponds to the formant frequencies. The curve thus formed is given by the circuit 49 for consonant thinning and the circuit 50 for Formation of vowel emphases thinned out to a few significant points. The curve in individual sections with precisely definable start and end points are broken down Curve points corresponding to vowels by χ and curve points corresponding to consonants by shown

Fig. 10 shows the output of the circuit 50 for forming vowel centroids. It can be seen that from the multitude of vowel points in Fig. 9 now a vowel centroid has been generated by forming a mean value

F i g. 11 shows the speech curve including the both vowel emphases. There are now two curves, their delimitation points are each formed from the zero point of the coordinate matrix and the vowel centroids.

Finally, FIG. 12 shows the output of the circuit 49 on the thinning of consons. The thinning of the two consonant curves takes place in this one Embodiment by deleting the curve points with the largest angles to the Neighboring points. The criteria for thinning are, on the one hand, preservation of the significant points and secondly, the same number of points for all spoken syllables, since these criteria are used to

Hi bene comparison with sample syllables is easiest is feasible. In this example, there are the two remaining curves 54 and 55 in FIG. 12 consequently from the four most characteristic points.

is The spoken syllable prepared in this way for a comparison reaches a comparator 51 (Fig. 8), to which 52 sample syllables can be fed from a feature memory. The feature memory 52 is For example, a read-only memory that contains all

2ü mend pattern syllables expressed by the same Contains the same number of points as the processed syllables (i.e. four in the example). By comparing the processed Syllables in the feature comparator 51 are selected from among all the sample syllables that is the least Shows a deviation from the processed syllable.

That is for those in the fig. 9 to 12 shown syllables in fig. 13 and 14 to be seen in more detail. In Fig. 13 the curve 55 appears again with its four dash-dotted lines significant points, while in FIG. 14 the

jo curve 54 is shown in dash-dotted lines in solid lines the closest pattern syllable is shown as curve 56, 57 in both figures. Of the The difference between the processed syllable and the sample syllable results from the hatched areas

J5 areas between the two curves. The smallest Area corresponds to the greatest similarity. Each sample syllable is assigned an identifier that is dated from Feature comparator 51 is passed on to output 53.

The pattern syllables determined in the manner described represent the recognized ones as a first approximation spoken syllables. In order to achieve a flawless, orthographically correct syllable output, is still another processing required, the z. B.

by combining the characteristics of the sample syllables into words and comparing them again with a Pattern word reserve can be made from a further read-only memory.

In addition 5 sheets of drawings

Claims (5)

Patent claims: 1. Electrical circuit arrangement for a device for machine-evaluable speech recognition, with which the frequency spectrum of the speech signals presented in the form of electrical signals in "Speech bands" are broken down, from which "formants" and from these "vowels" are determined and the one Has voice switch, characterized in that a first circuit part <10) for Determination of the number of vibrations of the "formants" lowest frequency in which the The lower voice band is filtered out and its frequency is converted, the lower sideband is screened out as well the phase difference between the lower sideband and the out-of-phase lower sideband be formed a second circuit part (20) for determining the number of oscillations of the "formant" of the next higher frequency, in which the upper voice band is filtered out and the phase difference between the upper voice band and the phase-shifted upper voice band is formed,
a third circuit part (30) for recognizing "vowels", in which the exceeding of a defined amplitude level of the "formant" of the lowest frequency in first signals and the presence of voicing are converted into second signals by evaluating the sawtooth pulses of the "formant" of the next higher frequency and common output signals are emitted when first and second signals occur simultaneously, and a fourth circuit part (40) for the automatic detection of Pauses in the speech signals through a threshold switch that depends on pauses in speech Supplies output signals, where for evaluation the output signals of the first and second circuit parts (10, 20) - controlled by the output signals of the fourth circuit part (40) - which can be counted separately Counting results through, the output signals of the third circuit part (30) after "vowels" and »Consonants« can be separated, from which sequences of features are formed and by comparison with Identify stored pattern feature sequences »syllables«. 2. Circuit arrangement according to claim 1, characterized characterized that the determination of the number of oscillations of the lowest frequency "formant" by connecting a low-pass filter (3), a modulator (4) with carrier suppression, to which the output of the low-pass filter and a frequency generator (8) are fed, one further active low-pass filter (5), a differential amplifier (6) to which the output of the low-pass filter (5) is fed both directly and out of phase, and a pulse shaper (7) takes place. 3. Circuit arrangement according to claim 1, characterized in that the determination of the number of oscillations of the "formants" of the next higher frequency by the series connection of a high-pass filter (11), a differential amplifier (12) to which the output of the high-pass filter (11) is both direct and is fed out of phase, nnH pinr-t: ImniiUfnrmpr *; iit \ prfrtttrf 4. Circuit arrangement according to claim 2 and 3, characterized in that the "vocal recognition" by checking whether a defined amplitude level of the "formant" of the lowest frequency is exceeded in a series connection of an amplifier (17), a rectifier (18) with Smoothing, a low-pass filter (19) and a pulse shaper (21) and by checking the voicing of the "formant" of the next higher frequency in a series connection of a rectifier (24) with smoothing, a low-pass filter (25), an amplifier (26), a timing element (27) and a pulse shaper (28) takes place, the Output signals of a conjunction (23) are fed. 5. Circuit arrangement according to claim 1, characterized in that the automatic detection the "speaking pauses" in a series connection of an amplifier (32), a rectifier (33) with smoothing, a low-pass filter (34) and a pulse shaper (35)