DD218969A1 - ARRANGEMENT FOR FORMING SENSE GROUPS - Google Patents

Abstract

Arrangement for forming speaker groups, in particular for automatic speech recognition systems, by forming mean value functions. The object of the invention is to use the different similarity relationships existing between the individual speakers in order to create groups of speakers. According to the invention, the object is achieved in that the speech signals of each speaker from a signal source, a microphone or magnetic tape via a chain circuit consisting of a preamplifier, a preemphasis filter, a built-up from a system of band pass filter bank, an analog multiplexer and an analog-digital Translators in the form of an 8-bit data word to be transferred to the digital computer. Over a defined period of time, the amplitude values are added up channel by channel and finally normalized to the number of time windows considered. The resulting mean value spectrum is loaded into an external memory. From the stored average spectra of N speakers examined, a similarity matrix is determined, using as a measure of similarity the city block distance, the Euclidean distance or other known distance mass. The speaker groups are formed in a subsequent stage. For this purpose, each distance measure is compared line by line in a further step with a similarity threshold. Fig. 4

Description

Arrangement for forming speaker groups, in particular for automatic speech recognition systems, by the formation of mean value functions.

Elements of the invention may be used in speaker identification systems and speaker verification systems.

Characteristic of the known technical solutions

The first research in the field of automatic speech recognition date back to the 1950s. Since that time, a variety of automatic recognition systems have become known which achieve more or less reliable recognition results using nearly as many speech features.

(HOIMGREET, G.L. Speaker Recognition, Speech Characteristics, Speech Evaluation and Modification of Speech Signals - A Selected Bibliography

in: IEEE Transactions of Audio and Electroacoustics, 1966, p. 32-39).

At the same time, the individuality of human language appears as a disruptive factor from the beginning. In order to take this into account, a number of restrictions have been necessary so far.

The currently implemented automatic speech recognition systems mainly use words as the smallest recognizable units (single word connoisseurs). There are still systems whose smallest units to recognize syllables or

Represent phonemes.

The limitations mentioned, which are due to the speaker individuality of the human language, led to the development of the individual word connoisseurs in two basic directions:. '

1. Purely each speaker using the speech recognition system is formed a separate reference data set. It is necessary that for each speaker the complete learning cycle must be processed. Before a new speaker comes to the system, first the

; selected for him reference record. This reference data selection can be done automatically (via the automatic speaker recognition.) Or it can be communicated to the recognition system via a control panel. The disadvantage of this method is that a complete reference data set must be determined for each speaker using the system. This requires a relatively large effort, since the computing time required for the learning phase is usually higher by a multiple than for the detection phase.

(US-PS 3261916)

LIEK, W ·: The recognition of speech by electronic systems

in: News Electronics, 1978, No. 9).

          \

2. In this method, a common reference data set is formed for all speakers. In order to be able to take account of the different speaker types, it is therefore necessary to store a plurality of reference patterns for each word. As a result, the storage space requirement necessarily increases by a multiple compared to the first method. In order to be able to reduce this effort at least partially, representatives of reference words with the same meaning of words are formed with the aid of cluster methods. However, the storage space requirement remains relatively high. The fact that the number of reference words compared to the first method is many times higher, also increases the detection time.

(DE-OS 2240557

SCOTT, P.B .: VICI-A Speaker Independent Word Recognition System in: Int. Conf. on ASSP, 1976 ,. Pp. 210-213).

The number of words and speakers to be processed is determined by the available memory space for storing the reference data. In addition, there must still be the possibility to perform the actual recognition process in a relatively short time (possibly real time). This means that without a significant increase in memory space, an increase in the number of speakers simultaneously results in a reduction in the number of words to be recognized. Likewise, an increase in the number of words to be recognized inevitably entails a reduction in the number of speakers. Since the storage space available in a digital computer is currently very limited and, moreover, relatively expensive, the limitations mentioned adversely affect a broader application. -

Attempts to compensate these specialized parts by means of a suitable feature vector normalization or feature vector transformation are still in their infancy and have hitherto been unsuccessful.

(VIHCZJUE, T.K ·; KULIAS, Α · Ι .: Sadaca podstrojki,

pod diktora pri raspoznavanii reci

in: OBRABOTKA I RASPOZHAVAHIE SIGHAXOV, Kiev, 1975

RAMISHVILI, G.S. Recevoj signal i individualnost golosa Isdatjelstvo "Mezniereba", Tbilisi, 1976).

The previous solutions are liable to the deficiency that with a constant size of the main memory an increase in the number of speakers is possible only by reducing the number of words to be recognized. On the other hand, an increase in the number of words to be recognized results in a reduction in the number of speakers.

Object of the invention

The aim of the invention is that without a significant increase in the necessary storage space requirement of the main memory a gpossi ^ e number of speakers with the car "-

- 4 matic speech recognition system can work.

Explanation of the essence of , invention

The object of the invention is to use the different similarity relationships existing between the individual speakers in order to create groups of speakers which are distinguished by the fact that the members of a speaker group are similar in their speech characteristics used for the speech recognition. According to the invention, the object is achieved in that the speech signals of each speaker from a signal source, a microphone or magnetic tape via a chain circuit consisting of a preamplifier, a preemphasis filter, a built-up of a system of bandpass filter bank, an analog multiplexer and an analogue Digital converter in the form of an 8-bit data word to which digital computers are transferred. Over a defined period of time, the amplitude values are added channel by channel and then normalized to the number of time windows considered. A similarity matrix is determined for the resulting center lumpsum of IT speakers and the city block distance, the Euclidean distance, as a similarity measure or other known pitch dimensions. .

The speaker groups are formed in a subsequent stage. For this purpose, each distance measure is compared in a further stage with a poor threshold. All indices of the distance measures which fall below the similarity threshold are passed on to a stage for outputting the speaker numbers of the determined speaker groups, sorted in ascending order and output. In automatic speech recognition systems, the formation of a mean value spectrum for pre-selection of the reference data set used for the classification takes place in a further stage. In an external memory, a selection of reference data sets of different speaker types is stored for this purpose. In averaging, all previously spoken recognition units are included. After each spoken word, it is checked whether the reference data selection made is still valid. A new review

The reference data set is loaded into main memory if it better satisfies the similarity conditions.

To form speaker groups, characteristic vector sequences of amplitude histograms, zero crossing functions and LPC analysis functions are used as speech features.

When an automatic speech recognition system is used by a first-grade speaker, the set of reference data that is used is that record which provides the best recognition result with high reliability.

Embodiment '..'''

The invention will be explained in more detail below using an exemplary embodiment. , , .

In the accompanying drawings:, ^l,

Pig. 1: Block diagram of the arrangement for formation. of speaker groups

Pig. 2: Mean spectrum of a female and a male speaker.

Pig. · 3 '· - Principle of speaker group formation in general

Pig. 4: Principle of speaker group formation in an automatic speech recognition system

The starting point for * the formation of speaker groups form mean value functions of different speakers. These mean value functions are formed by component averaging of parameter terms. In this case, the parameter functions must be representable as Yektorfolgen constant number of components · These conditions meet ζ. B. Pilter spectra, Hulldurchgangsfunktionen, amplitude or level histograms and iPC coefficients.

The use of patter spectra for automatic speech recognition is very widespread. In this Pail mean value spectra are used to determine the speaker groups. In Pig. 1, an arrangement is shown with which average value spectra of different speakers can be formed. The speech signals of each speaker come from a signal source 1 (microphone ^ magnetic tape o.

amplifier 2, to a pre-emphasis filter 3 · This filter raises the speech signal by about 6'-12 dB per octave, to counteract the amplitude drop typical of speech to high frequencies. Thereafter, the speech signal passes to the parallel filter bank 4 · This parallel filter bank consists of a system of bandpasses. The investigations were carried out with a 20-channel filter bank having the cutoff frequencies and center frequencies given below. '

iz (Hz) bandwidth (Hz)

70 - 170 170 - 270. 270-370 370 - 470 470 - 570 570 - 680 680 - 810 810 - 970 970 - 1170. 1170-1400 1400 - 1680 1680 - 2000 2000 - 2400 2400 - 2880 2880 - 3460 3460 - 4150 4150 - 4980. '4980 - 5980 5980 - 7170 7170 - 8600

An adjoining analog multiplexer 5 samples the filter outputs step by step every 12.5 ms. The resulting analog function is converted by the analog-to-digital converter into an 8-bit long data word. These 8 bits allow a resolution of 128 amplitude levels. This corresponds to a dynamic range of 48 dB. The further processing is done by computer. The digitized signal enters the computer 7 (KRS 4201) and is averaged there.

Channel number I littenfreq · 1 109 2 214 • / 3. 316 4 417 .5 518   6 623 7 742 a 886 9 1065 10 1280 11 1535 12 1833 13 2191 14 2629 15 3157 16 3789 17 ' 4546 18 5457 19 6548 20 7853

This happens in such a way that the amplitude values are summed over the entire period of time to be considered channel-wise and finally normalized to the number of time slots loaded. Break sections are not taken into account. The so-called Hittelwertspektrum can in an external memory 8 (magnetic tape, magnetic disks o. The like.) Stored and / or z. B. from a graphical output unit 9 (plotter, XY-vision or the like.) Are drawn.

Fig. 2 shows the mean value spectra of a male and a female speaker. The punctuation is based on a 30 words comprehensive speech. As a rule, a stable mean value spectrum has developed after a '20 - 30 s long speech, which is independent of the spoken text and characteristic for the respective speaker. The speaker-typical maxima are determined in particular by the high-energy vocal parts of the speech material. The vowels contain essential speaker-describing information. The word-meaningful information within the speech spectrum is attenuated by averaging over a longer, multi-word speech text. Speakers with a similar shape of the center value spectra therefore have substantial similarities with respect to articulation. "Such speakers can use a common reference data set for automatic speech recognition. Speakers with such a common reference record form a speaker group ·

Pig. 3 illustrates the principle of speaker group formation. In a digital computer (eg ERS 4201), the mean value spectra of Έ examined speakers 10 are stored in a suitable manner. Subsequently, the determination of a similarity matrix 11 takes place. The measure of similarity is the city block distance or the Euclidean distance. These similarity measures are very computationally easy to handle (simple arithmetic operations and thus low calculation times) and on the other physically very well interpretable. In addition, a variety of other distance measures, the calculation of which is much more complex (Pischerkriterium, Mahalanobis distance, correlation measure, etc.), apply.

r-, r * by αr> r, n ,. "^ π it *

  - 8th -

the. ^;

The city block distance can be determined by the relationship

JL

'' ^CD ij ⁼ 2_i! ^S i ⁽¹⁾ " ^X j ⁽¹⁾ 1 = 1

and the Euclidean distance after the relationship

I 'j' --- ^; - ~ -,

• 1 = 1-

determine. In this case, i and j denote the pair of speakers for which the distance is determined. The component of the parameter function is indicated by 1. L is the number of components that make up the parameter vector χ.

In stage 13, the formation of speaker groups takes place. In this case, the procedure is such that each distance measure is compared line by line with a similarity threshold level 12. The similarity threshold is determined experimentally. It should be chosen so that about 2-5 speakers are assigned to a group. After the entire line of the similarity matrix has been processed, all indices of the distance measures that fall below the similarity threshold are passed on to the level 14. The indices of the distance measures agree with the hammers of the considered speakers. In stage 14, the speaker numbers transmitted after processing each line of the similarity matrix are sorted in ascending order. The resulting polings are compared to exclude repetitions. Finally, the speaker numbers of each speaker group are output.

Figure 4 shows the arrangement for speaker group formation in an automatic speech recognition system. "A significant advantage of the system is the dynamic speaker adaptation. There is a constant updating of the reference data. For the input of the speech data, what has been said with reference to FIG. 1 applies. This concerns stages 1-6.

While the input vector sequence is used for classification within the computer, on the other hand the formation of a mean value spectrum for the preselection takes place

^r - 9 -

of the reference data set used for the classification in a stage 15. An external memory 16 (digital magnetic tape, digital cassette, magnetic disk or the like) stores a selection of reference data records of different types of speakers in a suitable manner. With the aid of the mean value spectrum formed in step 15, the reference data record is selected whose mean value spectrum (formed over the entire reference data record) has the greatest similarity to this newly determined average value spectrum. All previously spoken identification units are included in the averaging. After each spoken word, a check is made as to whether the reference data from Y / ahl17 is still valid. Otherwise, another reference data set that better satisfies the similarity condition is loaded into main memory. As a result, an adaptation of the recognition system to the new speaker is achieved when the speaker changes. Since most speech recognition systems provide for the possibility of rejecting a misrecognized word level 18 ", it is even possible to constantly update the selected reference data set.Whenever a recognition unit has been recognized correctly, the determination of a new reference data word takes place in which the mean value between the Thus, it is achieved that the speech changes typical of a speaker are taken into account during the normal speech process. In step 20, the reaction of the recognition system intended by the speech input takes place.

Table 1 contains the recognition results of 10 male speakers. A total of 30 different words were used, each with 7 implementations per word. This gives a total of 210 words per speaker.

With the arrangements described in FIGS. 1 and 3, the similarity matrix was determined for these speakers. The distance measure used was the city block distance. The similarity matrix is included in Table 2. The similarity threshold for speaker group formation was chosen so that groups of a maximum of two speakers are formed. For larger groups of speakers, a higher number of speakers is needed. The

Table 3 contains the recognition results for two speaker groups (group 1 consisting of speakers 2 and 8, group 2 made up of speakers 5 and 7).

The recognition results of the speakers within a speaker group are over 90 %. During the recognition process, only the reference data set for a speaker group is in the main memory. Which reference data record this is, is decided during the current recognition process, live the reference speech data of the respective speaker group, the mean value spectra of all available reference data sets of the different speaker groups are also stored. -

Parallel to the recognition process, the average range of the user is determined. In this case, previous utterances are included in the averaging. In each recognition process it is checked to which average range of the reference data sets of the speaker groups the greatest similarity exists. This reference data set is loaded from an external memory (magnetic tape, magnetic disk, magnetic cassette or the like) in the main memory. Thus, the recognition system works to a certain extent 'speaker-adaptive. When changing the speaker, you can automatically access a suitable reference data record.

It should be clarified with reference to Tables 1 and 2 how a foreign speaker uses the automatic speech recognition system. As reference data sets are for example patterns of speakers 1 -. 5 · In general, Pail may also be reference records of speaker groups. The speakers 6-10 represent foreign speakers. Thus, the first 5 lines of Table 1 and 2 apply.

From Table 2 it can be seen that, for example, for speaker 10 (speaker) the speaker 3 provides the smallest distance measure (0.81). Table 1 shows that using the speaker 3 reference record of speaker 10, the highest recognition score (86.3 %) is achieved.

For speaker 8, the reference data set of speaker 2 (distance measure 0.73) yields the highest recognition result (87.6 %). For speaker 7, the reference data set of speaker 2 is therefore used

Speaker 5 (distance 0.78) reaches the best recognition result (90.4 SS) '.

ITutzer

1 1 2 3 4 5 6 7 8th 9 10 2 96.2 81, 8 61.9 , 83.8 91.9 85.2 78.0 83.3 54.7 50.4 3 81.9 99.0 78.1 77.4 84.4 79.5 79.4 87.6 72.4 70.0 Qj Λ O • 4 74.3 74.2 99.0 82.4 85.2 72.9 74.2 86.4 67.5 86.3 φ U 5 71. 9 72.2 65.2 97.1 80.0 78.1 76.1 72.4 69.0 71.4 Pt co 6 79.5 80.3 77.0 78.6 99.0 71.4 90.4 62.5 66.0 62.4 U φ 7 72.1 71, 2 76.3 82.2 78.3 97.8 69 ,, 8 72.9 71.8 78.3 8th 60.9 65.3 63.2 > 9.1 86.3 67.3 98.1 72.1 66.3 62.1 φ 9 63.2 75.3 69.1 68.2 62.1 64.1 62.7 97.3 73.2 77.3 © 10 59.8 62.3 60.3 69.8 72.1 67.3 69.1 73.2 97.8 80.3 57.9 53.8 49.7 65.2 68.3 79.2 62.7 85.1 84.2 98.5

Table 1 (figures in %)

speaker

1 1 2 3 4 • 5 6 7 8th 9 10 2 0.00 1, 23 1.69 1.19 1.22 1.41 1.39 1, 10 1.11 1, 26 3 1.23 0.00 1, 16 1, 02 0.93 1.32 1.02 0.73 1.22 1.33 4 1.69 1, 16 0.00 1, 06 0.99 1.00 1.34 0.93 1.14 0.81 U 5 1.19 1, 02 1, 06 0.00 0.81 0.84 0.98 0.93 1, 08 0.99 £} 6 1.22 0.93 0.99 0.81 0.00 1.03 0.78 1, 02 1, 09 1.12 W (S U 7 1.41 1.32 1.00 0.84 1.03 0.00 1.28 1.19 1.12 0.95 Oi Kl 8th 1.39 1.02 1.34 0.98 0.78 1, 28 0.00 1.31 1.36 1, 47 9 1, 10 0.73 0.93 0.93 1, 02 1.19 1.31 0.00 0.87 0.78 10 1.11 1, 22 1.14 1, 08 1.09 1.12 1.36 0.87 0.00 0.95 1.26 1.33 0.81 0.99 1.12 0.95 1.47 0.78 0.95 0.00

Table 2 (figures without unit of measurement (normalized),)

& ⁴ Φ

attached. spreader Learn pher sr vn ro he" 03 U) _i 03 03 95, ' ro • Φ ro co ro VD U) ω • φ 00 ω U) vn U? VD CP. U) VJL VD · £ - 03  s to oo VD O 03 VD -J ro V ». f * v. 03 vn VD VD 75.7 ι O

φ H

ro

Claims (4)

Arrangement for forming speaker groups, in particular for automatic speech recognition systems, characterized by the formation of mean value functions characterized in that the speech signals of each speaker are from a signal source (1), a microphone or magnetic tape, via a chain circuit consisting of a preamplifier (2 ), a pre-emphasis filter (3), a filter bank (4) constructed from a system of band passports, an analog multiplexer (5) and an analog-to-digital converter (6) in the form of an 8-bit data word to the digital computer (7) be transferred over a defined period of time the amplitude values channel by channel and finally to the , normalized time window, the resulting mean value spectrum is loaded into an external memory (8) and from the stored average value spectra (10) of IT. examined speakers, a similarity matrix (11) is determined and the city block distance is used as a measure of similarity, compared in a following stage (12) with a similarity threshold, all indices of the distance measures which fall below the similarity threshold are sent to a stage (14). in automatic speech recognition systems the averaging for pre-selection of the reference data set used for classification is carried out in one stage (15) »a selection of reference data sets of different speaker types is stored in an external memory (16 In the averaging, all previously spoken recognition units are included, after each spoken recognition unit. Word is checked whether the made reference data selection (17) is still valid '* and a new reference data set is loaded into the main memory, if it satisfies the similarity conditions better. 2. Arrangement for forming speaker groups according to item 1, characterized in that are used as language features feature vector sequences of amplitude histograms, ITulldurchgangsfunktionen and LPC analysis functions. -H- 3 · Arrangement for the formation of speaker groups according to. Point 1 is characterized by the fact that the measure of similarity is the Euclidean distance, the degree of cross - correlation and. the Mahalanobis distance can be used. 4. Arrangement for the formation of speaker groups according to item 1, characterized in that when using an automatic speech recognition system by a foreign speaker from existing sets of reference data that record for detection · is used, which provides the best recognition result with high certainty. For this 3 sheets of drawings.