DE10302101A1 - Training of a Hidden Markov Model using training data vectors and a nearest neighbor clustering method based on condition parameters used to describe the Hidden Markov Model - Google Patents

Abstract

Method for computer-based training of a Hidden Markov Model (HMM) in which condition parameters are formed with which the HMM can be described, training data vectors are used to carry out a k-nearest neighbor clustering method based on the condition parameters, whereby the condition parameters are matched to the training data vectors, so that a trained HMM is formed. Independent claims are also included for the following:- (a) a device for training an HMM; (b) a computer program element for use on a computer for training an HMM and; (c) a computer readable storage medium for storing a computer program for training an HMM.

Description

The invention relates to a method and a device for training a hidden Markov model as well a computer program element and a computer readable storage medium.

Hidden Markov models can be found in one Numerous different technical areas of application, for example in the context of digital signal processing, especially speech recognition or pattern recognition, use.

So that a hidden Markov model in can be used in practice, it must be for example, for signal processing.

• – die Emissionswahrscheinlichkeitsverteilung,
• – die Zustandsübergangs-Wahrscheinlichkeiten und
• – die Initialwahrscheinlichkeiten sowie
• – die Anzahl der Zustände des Hidden Markov Modells.

The system parameters of the HMM are calculated as part of the training of a Hidden Markov Model (HMM). The system parameters of the HMM are:

• - the emission probability distribution,
• - the state transition probabilities and
• - the initial probabilities as well
• - the number of states of the Hidden Markov model.

For example, a hidden Markov model for a speaker-dependent Speech recognition system online, especially in an embedded system, be trained so it is desirable a training procedure with relatively few arithmetic operations required to be used to meet the system requirements as part of the training of the HMM if possible to keep low.

As a method of training one Hidden Markov model is known as the Baum-Welch method (compare [1]).

A disadvantage of the Baum-Welch process is in particular that this requires a very high computing effort. On Another significant disadvantage of the Baum-Welch process is in the numerical problems occurring, as part of this Procedure multiplied many probability values together become, which leads to very small numbers. The very small numbers are part of a common 8-bit, 16-bit or 32-bit processor, even with a 64-bit processor often no longer processable with the required accuracy.

The problem thus lies with the invention based on training a hidden Markov model, the training with reduced numerical problems and with reduced computing effort compared to the Baum-Welch method.

The problem is caused by the procedure and the device for training a hidden Markov model as well through a computer program element and a computer readable storage medium solved with the features according to the independent claims.

In a process for computer-aided training of a hidden Markov model, state parameters are formed with those states of the Hidden Markov model.

Using training data vectors becomes k-nearest neighbor clustering on the state parameters of the Hidden Markov model, where the state parameters according to the training data vectors be adjusted. Call this way the trained Hidden Markov Model formed.

A device for training a Hidden Markov model has a processor unit that is is set up that the process steps described above feasible are or are being carried out.

A computer program element points when executed by a processor unit, the ones described above Steps on to train a Hidden Markov model.

In a computer readable storage medium is a program stored when it is run by a processor unit accomplished which has the method steps described above for training of a hidden Markov model.

According to the invention, the problem becomes clear the determination of the system parameters of the HMM (in other words, the State parameters of the HMM) on the determination of representatives of the training data vectors. All Training data vectors with common meaning in a speech recognition system for example, represent a spoken language unit, are preferred to grouped together in a state. Within the training data vectors become representatives determined. The logarithmic emission probabilities are preferably obtained as the Euclidean distance between the representatives of the HMM and the Observation vector of the input signal. The state transition probabilities and the initial probabilities are according to an embodiment of the invention accepted as constants.

The training data vectors are common Vectors that describe the behavior of a technical system and preferably measured at the beginning of the process. As part of Speech recognition thus includes, for example, human expressions and the training data vectors become the recorded speech signal certainly. In the context of digital image processing, for example Images captured by a camera and the digitized Images are used to generate training data vectors from them.

In general, the training data vectors Data that is common describe a technical system in its actual behavior and serve the behavior of the technical system as possible to reproduce truthfully using the Hidden Markov model.

The invention makes it possible to in the Baum-Welch process Avoid occurring numerical problems as no more products must be determined from many probability values.

Thus, the invention itself is based on a processor with a relatively small word width and with a relative low computing power can be used. The invention is therefore suitable especially for use in an embedded system.

Preferred developments of the invention result from the dependent Claims.

The configurations described below The invention relates to the method and the device for training of a hidden Markov model as well as the computer program element as well the computer readable storage medium.

According to an embodiment of the invention the training data vectors are provided at the beginning of the method to group into different state groups, one state group is assigned to a state of the Hidden Markov model. The state parameters of a respective state of the Hidden Markov Models with training data vectors of the associated state group trained.

By assigning training data vectors into groups and thus by combining training data vectors in groups more similar to each other in terms of a given similarity criterion Training data vectors will converge the training process significantly accelerated and thus the need for computing power further reduced.

As part of the application of a hidden For example, Markov's language processing model means that the training data vectors form groups with a common one Meaning summarized. As part of language processing the training data vectors are grouped together, for example, where each group represents a spoken unit.

• – eine Symboldarstellung von einzelnen Zeichen, die eine lautsprachliche Äußerung repräsentieren,
• – Phonemsegmente,
• – Phoneme,
• – Allophone,
• – Diphone,
• – Halbsilben, oder
• – ganze Wörter.

Within the scope of the invention, a phonetic unit is understood to be a symbolic representation of a verbal utterance, for example:

• A symbol representation of individual characters that represent a spoken utterance,
• - phoneme segments,
• - phonemes,
• - Allophones,
• - Diphone,
• - half syllables, or
• - whole words.

For further simplification and thus for faster feasibility In one embodiment of the invention, the method is provided the mean value vectors of the probability density functions as state parameters to use which the respective state of the Hidden Markov model describe.

Is preferably used as the k-nearest-neighbor clustering method a 1-nearest-neighbor clustering process used (i.e. k = 1), in other words a winner takes all method (WTA) method.

In this embodiment of the invention thus only that becomes actual closest to the respective training data vector in the comparison space State parameters, in particular mean vector selected and adapted to the respective training data vector. That way further improves the convergence behavior of the training process.

According to another embodiment The invention provides for checking whether a state parameter sufficiently often as part of the k-nearest-neighbor clustering process has been adjusted and for the case that a state parameter is not sufficiently frequent in the Framework of the k-nearest-neighbor clustering process has been adjusted so that it is replaced by a new one State parameters. The new state parameter is preferably due to this characterizes that it is sufficiently frequent, preferably most frequently, adapted as part of the k-nearest-neighbor clustering process has been.

Through this embodiment of the invention the convergence behavior of the training process is further improved.

Alternatively, the new state parameter can also be generated randomly.

According to another embodiment The invention provides a state parameter in the frame of the k-nearest-neighbor clustering method to an even lesser extent adapt to the training data vectors, the more training data vectors be taken into account as part of the training process.

The method is particularly suitable for use in an embedded system.

• – in der Signalverarbeitung,
• – in der Mustererkennung,
• – in der Sprachverarbeitung, insbesondere in der sprecherabhängigen Spracherkennung oder der Sprechererkennung, insbesondere in einer Gerätesteuerung, beispielsweise in einem Telekommunikationsgerät, in der Home Automation, d.h. beispielsweise in einer Umgebungssteuerung,
• – in der Bildverarbeitung,
• – in der Schriftzeichenerkennung,
• – im Rahmen der Prognose des Verhaltens eines technischen Systems, welches mittels Hidden Markov Modells beschrieben wird.

Furthermore, the invention can be used in one of the following areas:

• - in signal processing,
• - in pattern recognition,
• In speech processing, in particular in speaker-dependent speech recognition or speaker recognition, in particular in device control, for example in a telecommunications device, in home automation, ie for example in environmental control,
• - in image processing,
• - in character recognition,
• - as part of the prediction of the behavior of a technical system, which is described using the Hidden Markov model.

The invention can be clearly illustrated therein can be seen that the representatives of the training data vectors to be determined according to the following Procedure to be determined.

• a) der in dem Vergleichsraum räumlich nächstliegende Repräsentant (Winner) zum aktuell betrachteten Trainingsdatenvektor im Sinne einer minimalen euklidischen Distanz ermittelt wird,
• b) der ermittelte Repräsentant mit steigender Zahl von Durchläufen, vorzugsweise pro Durchlauf, abnehmender Adaptionskonstante an den jeweilig verwendeten Trainingsdatenvektor angepasst wird.

First, all the training vectors that are assigned to a state are determined. This is followed by an initialization, preferably a random initialization of the desired number of representatives of the training data vectors. The representatives, ie the state parameters, are subsequently adapted to the probability density function of the training data vectors; in other words, adjusted by

• a) the representative (winner) spatially closest to the currently considered training data vector is determined in terms of a minimal Euclidean distance,
• b) the determined representative is adapted to the training data vector used in each case with increasing number of runs, preferably per run, decreasing adaptation constant.

Steps a) and b) are described in an arbitrary number of runs with all training data vectors carried out.

Are state parameters within of a run are not determined often enough to be the winner, so treats them separately, preferably in such a way that in the event that that a condition parameter is met too rarely; he through that Representatives, i.e. that state parameter that is replaced in the same pass most frequently when Winner was determined.

Finally, preferably each used adaptation constant (adaptation constant) to the Total number of considered Training data vectors adjusted. The adaptation constant is in of the type chosen that the adaptation behavior even in conditions with very different Number of training data vectors which correspond to the respective state are assigned the same and regardless of the scaling of the Training data vectors is.

An embodiment of the invention is shown in the figures and is explained in more detail below.

Show it

1 a block diagram of a speech recognition device according to an embodiment of the invention;

2 a diagram of a hidden Markov model according to an embodiment of the invention;

3A and 3B a flow chart in which the individual method steps for training the hidden Markov model are shown according to an embodiment of the invention.

4 a telecommunications device with a speech recognition device according to an embodiment of the invention;

5 a car radio with a speech recognition device according to an embodiment of the invention.

l shows a speech recognition device 100 according to an embodiment of the invention.

The speech recognition device 100 Depending on the operating mode, it works in a first operating mode as a speech recognition device, the speech recognition device making a spoken utterance in the speech recognition mode 101 , spoken by a user (not shown) of the speech recognition device 100 , recognizes using a speaker-independent speech recognition technique. In a second operating mode, hereinafter also referred to as training mode, the fol The described hidden Markov models that are used in the context of speech recognition are trained.

The speech signal spoken by the user is used in both operating modes 101 a microphone 102 supplied, and as a recorded electrical analog signal 103 a pre-processing, in particular a pre-amplification by means of a pre-processing unit 104 , in particular subjected to a preamplifier and as a preprocessed and amplified analog signal 105 an analog / digital converter 106 fed, there in a digital signal 107 converted and as a digital signal 107 a calculator 108 fed.

It should be noted in this context that the microphone 102 , the preprocessing unit 104 , in particular the amplification unit, and the analog / digital converter 106 as separate units or as in the computer 108 integrated units can be realized.

According to this exemplary embodiment, it is provided that the digitized signal 107 the calculator 108 via its input interface 109 is fed.

The computer 108 also has a microprocessor 110 , a memory 111 as well as an output interface 112 on all of them together using a computer bus 113 are coupled.

Using the microprocessor 110 the method steps described below, in particular the methods for training the hidden Markov models and for speech recognition are carried out using the trained hidden Markov models. An electronic dictionary that contains the entries that contain reference words as part of the speech recognition is in the memory 111 saved.

It is also connected to the computer bus 113 coupled digital signal processor (DSP) 114 provided that has a microcontroller that is specially specialized for the speaker-independent speech recognition algorithms used.

In the digital signal processor 114 a computer program is also stored, which is set up for speaker-independent speech recognition. Alternatively, the algorithms used can be implemented in hard-wired logic, that is to say directly in hardware.

Furthermore, the calculator 108 using the input interface 109 with a keyboard 115 as well as a computer mouse 116 via electrical cables 117 . 118 or a radio connection, for example an infrared connection or a Bluetooth connection.

Via additional cables or radio connections, for example an infrared connection or a Bluetooth connection 119 . 120 is the calculator 108 using the output interface 112 with a speaker 121 as well as an actuator 122 coupled.

The actuator 122 represented in 1 generally any possible actuator in the context of the control of a technical system, for example realized in the form of a hardware switch or in the form of a computer program in the event that, for example, a telecommunications device or another technical system, for example a car radio, a stereo system, a video recorder, a television, the computer itself or any other technical system is to be controlled.

According to the exemplary embodiment of the invention, the preprocessing unit has 104 a filter bank with a plurality of band-pass filters, which show the energy of the input speech signal 103 measure in individual frequency bands. So-called short-term spectra are formed by means of the filter bank, in that the output signals of the bandpasses are rectified, smoothed and sampled at short intervals, in accordance with the exemplary embodiment every 10 msec. The so-called cepstrurn coefficients of two successive time windows as well as their first time derivative and their second time derivative are determined and combined into a super feature vector and the computer 108 fed.

In an alternative embodiment, the preprocessing unit 104 in their function of mapping the speech signal onto the cepstrum coefficients in the computer 108 and in it in the microprocessor 110 and / or in the DSP 114 be realized.

In the calculator 108 is, as described above, a speech recognition unit in the form of a computer program, the speech recognition being based on the principle of the Hidden Markov models, for speaker-independent speech recognition, according to the exemplary embodiment in the DSP 123 ,

In a basic vocabulary at the beginning of the procedure in an electronic dictionary 114 A hidden Markov model is stored for each base entry, which is determined in each case in the following manner using a training data record, that is to say a set of training speech signals, spoken in by one or more training users.

• – einer erste Phase, in der die in der Trainings-Datenbank enthaltenen Sprachsignale segmentiert werden,
• – einer zweiten Phase, in der die LDA-Matrix (lineare Diskriminanzanalyse-Matrix) berechnet wird sowie
• – einer dritten Phase, in der das Codebuch, das heißt die HMM-Prototypen-Merkmalsvektoren für jeweils eine in einem Auswahlschritt ausgewählte Anzahl von Merkmalsvektor-Komponenten berechnet werden.

According to this exemplary embodiment, the training of the Hidden Markov models takes place in three phases:

• A first phase in which the speech signals contained in the training database are segmented,
• - a second phase in which the LDA matrix (linear discriminant analysis matrix) is calculated and
• A third phase in which the code book, that is to say the HMM prototype feature vectors, is calculated for a number of feature vector components selected in a selection step.

The entirety of these three phases is referred to below as the training of the Hidden Markov models (HMM training).

The HMM training is done using the DSP 123 as well as using predetermined training scripts, clearly illustrated by suitably set up computer programs.

According to this embodiment each spoken unit, that is, each Phoneme, divided into three consecutive phoneme segments, corresponding to an initial phase (first phoneme segment), a central one Phase (second phoneme segment) and a final phase (third phoneme segment) of a sound, that is of a phoneme.

In other words, each sound becomes one Sound model with three states, this means modeled with a three-state HMM.

2 shows an example of a hidden Markov model 200 according to the embodiment of the invention.

• – eine Emissionswahrscheinlichkeits-Verteilung b_j(o), mit der angegeben wird die Wahrscheinlichkeit, dass der Beobachtungsvektor o in dem Zustand j emittiert wird,
• – Zustandsübergangs-Wahrscheinlichkeiten α_ij für den Zustandsübergang von dem Zustand i in den Zustand j,
• – Eine Initialwahrscheinlichkeit π_i, mit der die Wahrscheinlichkeit angegeben wird, mit welcher ein Pfad durch das Hidden Markov Modell 200 in dem Zustand i beginnt.

The Hidden Markov Model 200  has a variety of conditions 1 . 2 . 3 . 4 . 5  on. Furthermore, the hidden Markov model 200  characterized by

• - one Emission probability distribution b_j(O) with which is specified the probability that the observation vector O in the state j is emitted,
• State transition probabilities α _ij for the state transition from state i to state j,
• - An initial probability π _i , which indicates the probability with which a path through the Hidden Markov model 200 starts in state i.

In the following, the universality the following boundary conditions are assumed.

In a left-right model as a hidden Markov model 200 , in which only the state transitions of a self-loop (selfloop), a one-step state transition (step) and a two-step state transition (skip) are taken into account, the state transition probabilities can be α _ij , which are not equal to the value “zero” , are used as constants.

In other words, this means that in this case for the exemplary Hidden Markov model 200 out 2 applies: α 12 = α 23 = α 34 = α 45 = C 0 (1) (for a one-step state transition) or α 22 = α 33 = α 44 = α 24 = C 1 (2) (for a self-loop and a two-step state transition).

Everyone else (in 2 (not shown) state transition probabilities α _ij are set to the value "zero".

In this connection it should be noted that it is not necessary to assign a constant value to the state transition probabilities α _ij . In an alternative embodiment of the invention, it is provided, for example, to encode the phoneme length in the state transition probabilities α _ij .

Furthermore, according to this exemplary embodiment, the following is defined: π 1 = 1 (3) and π i ≠ 1 = 0 (4) ie a path through the Hidden Markov model 200 can only ever in the first state 1 kick off. This assumption can always be made when using the Hidden Markov model 200 described system an initial condition exists.

wobei mit

• – c_jk ein Wahrscheinlichkeits-Verteilungskoeffizient der k-ten Wahrscheinlichkeits-Verteilung in dem Zustand j,
• – Mj die Anzahl von in dem Zustand j vorhandener Wahrscheinlichkeits-Verteilungen,

bezeichnet wird
und wobei

wobei mit

• – u_jk ein Mittelwertvektor der k-ten Wahrscheinlichkeits-Verteilung in dem Zustand j,
• – Σ_jk eine Kovarianzmatrix der k-ten Wahrscheinlichkeits-Verteilung in dem Zustand j,
• – n die Dimension des Beobachtungsvektors o,

bezeichnet wird.The emission probability distribution b _j ( o ) is modeled according to this exemplary embodiment by means of a continuous, multidimensional probability distribution, so-called continuous density mixtures, according to the following rule:

being with

• C _{jk is} a probability distribution coefficient of the kth probability distribution in the state j,
• Mj the number of probability distributions present in state j,

referred to as
and where

being with

• U _{jk is} an average vector of the kth probability distribution in the state j,
• - Σ _{jk is} a covariance matrix of the kth probability distribution in the state j,
• - n the dimension of the observation vector o,

referred to as.

Each probability distribution corresponds to a multi-dimensional Gaussian distribution. With such a continuous density mixture model, any continuous probability density function can be approximated with any accuracy. If the variances of the Gaussian distributions are set to a constant value and at the same time the covariances are zero - which is the case if the observation vector o was determined using a linear discriminant analysis (LDA), which is explained in more detail below - then the emission probability distribution is to be determined b _j ( o ) only the mean vectors μ _{jk are} required.

While In speech recognition, the three phoneme segments in a Bakis topology or generally a left-right topology strung together and on the concatenation of these three segments in a row the calculation was carried out as part of speaker-independent speech recognition.

As will be explained in more detail below, in the speech recognition mode a Viterbi algorithm is used to decode the feature vectors which result from the input speech signal 101 be formed.

After segmentation, the LDA matrix A is determined using an LDA matrix calculation unit.

• – x ein Merkmalsvektor,
• – A eine LDA-Matrix,
• – y ein Super-Merkmalsvektor,
• – y ein globaler Verschiebungsvektor

bezeichnet wird.The LDA matrix A is used to transform a respective super feature vector y to a feature vector x according to the following rule: x = A T · ( y - y ), (7) being with

• - x a feature vector,
• A an LDA matrix,
• - y a super feature vector,
• - y a global displacement vector

referred to as.

• – die Komponenten des Merkmalsvektors x im statistischen Durchschnitt voneinander im Wesentlichen unkorreliert sind,
• – die statistischen Varianzen innerhalb einer Segmentklasse im statistischen Durchschnitt normalisiert sind,
• – die Zentren der Segmentklassen im statistischen Durchschnitt einen maximalen Abstand voneinander aufweisen und
• – die Dimension der Merkmalsvektoren x möglichst, vorzugsweise Spracherkennungsanwendungs-abhängig, reduziert wird.

The LDA matrix A is determined in such a way that

• The components of the feature vector x are essentially uncorrelated from one another on a statistical average,
• The statistical variances within a segment class are normalized on a statistical average,
• - The centers of the segment classes have a maximum distance from each other on a statistical average and
• - The dimension of the feature vectors x is reduced as possible, preferably depending on the speech recognition application.

The method for determining the LDA matrix A according to these exemplary embodiments is explained below.

However, it should be noted that, alternatively, all known methods for determining an LDA matrix A can be used without restriction.

wobei mit N_j die Anzahl der in der Klasse j sich befindenden Super-Merkmalsvektoren y _j bezeichnet wird.It is assumed that J segment classes exist, each segment class j containing a set D _y -dimensional super feature vectors y , that is to say:

where N _j is the number of the class that are available super feature vectors y _j j.

wird die Gesamtzahl der Super-Merkmalsvektoren y bezeichnet.With

is the total number of super feature vectors y .

It should be noted that the super feature vectors y _j ^{k have} been determined using the segmentation of the speech signal database described above. According to this exemplary embodiment, each super feature vector y _j ^{k has} a dimension D _y of Dy = 78 (= 2 · 3 · 13) on, wherein 13 MFCC coefficients (cepstrums coefficients) are contained in the super feature vector y _j ^k , as well as their respective temporal first derivative and their respective temporal second derivative (this justifies factor 3 above).

Furthermore, each super feature vector y _j ^k contains the components of two temporally immediately consecutive time windows in the context of the short-term analysis (this justifies factor 2 above).

It should be noted in this connection that in principle any number of vector components adapted to the respective application can be contained in the super feature vector y _j ^k , for example up to 20 cepstrums coefficients and their associated temporal first derivatives and second derivatives.

The statistical mean or, in other words, the center of class j results from the following rule:

The covariance matrix Σ _j of class j results according to the following rule:

wobei p(j) als Gewichtungsfaktor der Klasse j bezeichnet wird.The average intea scattering matrix S _w is defined as:

where p (j) is called the weighting factor of class j.

als dem Durchschnitts-Super-Merkmalsvektor über alle Klassen.In an analogous manner, the average inter-scattering matrix S _{b is} defined as:

as the average super feature vector across all classes.

• – U eine erste Transformationsmatrix,
• – W eine zweite Transformationsmatrix und
• – V eine dritte Transformationsmatrix

bezeichnet wird.The LDA matrix A is broken down according to the following rule: A = U · W · V , (16) being with

• - U a first transformation matrix,
• - W a second transformation matrix and
• - V a third transformation matrix

referred to as.

The first transformation matrix U is used to diagonalize the average intra-scatter matrix S _w and is determined by transforming the positively definite and symmetrical average intra-scatter matrix S _w into its eigenvector space. In its eigenvector space, the average intea scattering matrix S _{w is} a diagonal matrix, the components of which are positive and greater than or equal to zero. The components whose values are greater than zero correspond to the average variance in the respective dimension defined by the corresponding vector component.

The second transformation matrix W is used to normalize the average variances and is determined according to the following rule:

The transformation U · W is also called whitening.

With B = U W (18) the matrix B ^T · S _w · B results in the unit matrix, which remains unchanged in any orthonormal linear transformation.

In order to diagonalize the average inter-scattering matrix S _b , the third transformation matrix V , which is formed in accordance with the following rule: V = B T · S b B, (9) where B ^T · S _b · B also represents a positively definite and symmetrical matrix, is transformed into its eigenvector space.

und eine diagonalisierte Durchschnitts-Inter-Streumatrix S _b:

wobei mit

eine D_y × D_y Diagonalmatrix mit den Komponenten c_d in der Zeile/Spalte d und sonst mit Komponenten mit dem Wert Null, bezeichnet wird.In the transformation space x = A T · ( y - y ) (20) The following matrices result: A diagonalized average intra-scatter matrix S _w :

and a diagonalized average inter-scattering matrix S _b :

being with

a D _y × D _y diagonal matrix with the components c _d in the row / column d and otherwise with components with the value zero.

die Gesamt-Durchschnitts-Varianz des Durchschnitts-Vektors x _j der J Klassen darstellt.The values σ ² _d are the eigenvalues of the average inter-scattering matrix S _b and represent a measure for the so-called pseudo-entropy of the feature vector components, which is also referred to below as the information content of the feature vector components. It should be noted that the trace of each matrix is invariant with respect to any orthogonal transformation, which results in the sum

represents the total average variance of the average vector x _{j of} the J classes.

The result is a determined pendency the pseudo-entropy of the feature vectors of those in the feature vector contained or considered Feature vector components.

According to this exemplary embodiment, a dimension reduction is then carried out by sorting the σ ² _d values in order of decreasing size and omitting the σ ² _d values, that is to say disregarding those that are smaller than a predetermined threshold value. The predetermined threshold value can also be defined cumulatively.

Have then the LDA matrix A ^T can be adjusted by the lines corresponding to the eigenvalues σ ² are sorted _d and the rows are omitted belong to the sufficiently "small" variances, and thus only a small information content (low Pseudoentropie).

According to this exemplary embodiment, the components with the 24 largest eigenvalues σ ² _d are used, in other words D _x = 24.

The four steps described above for determining the LDA matrix A are summarized in the following table:

The last procedure to partial procedure Clustering is part of the training of the Hidden Markov models the feature vectors, which is carried out by means of a cluster unit and which result has a respective codebook, each specific for one Training data record with a predetermined number of feature vector components. The entirety of the representatives The segment classes are referred to as the code book and the representatives themselves are also called prototypes of the phoneme segment class.

The prototypes, furthermore as prototype feature vectors are designated according to the following Procedure determined.

The 3a and 3b show in a flow chart 300 the individual process steps of the partial process for training the hidden Markov model 200 according to the embodiment of the invention.

According to one in 3a First step shown (step 301), the method is started.

Then, as explained above, the training data vectors acquired (step 302).

In a subsequent step the training data vectors in different state groups, which each have a common meaning, according to this embodiment a spoken unit or part of a spoken unit Represent unity, grouped (step 303) and a respective state that also represents the respective speaking unit.

This is followed by a random initialization of M _j mean vectors in a respective HMM state for all HMM states.

Usually are for a full word model with approx 2000 representatives, for a Phoneme model with 4000 to 10000 representatives at approximately 100 to 200 states approximately at initialization 20 mean vectors are provided per state.

• – w_start ein Gewichtswert der initialisierten Mittelwertvektoren und
• – rand() eine Zufallszahl-Erzeugungsfunktion,

bezeichnet wird.The randomization of the M _j mean vectors μ _jk is carried out according to the following rule: μ jk = w begin Rande (), (24) being with

• - w _start a weight value of the initialized mean vectors and
• Rand () a random number generation function,

referred to as.

After initialization of the M _j mean vectors μ _jk for each of the HMM states of the Hidden Markov model 200 an HMM status index j is initialized to the value 1 (step 305).

Then a training data vector index n with which a respective training data vector is uniquely identified is set to 1 (step 306).

To the first selected training data vector, to which the training data vector index n = 1 is assigned, the Average vector in the state j = 1 determined that the training data vector with n = 1 is closest in the comparison space (step 307).

In other words, for the training _{data vector} v _jn currently _{under consideration,} the closest mean _vector μ _{jw is determined} according to the following rule:

The closest mean value _vector (winner μ _jw ) is adapted to the position of the current training _{data vector} depending on the currently used training _{data vector} v _jn (step 308):

The adaptation coefficient a takes according to this embodiment with the number P of runs through all training data vectors from according to the following regulations:

1st pass: a 1 = a begin , a begin ≈ 0.1

All other runs: a p + 1 = a p · a diff , a diff ≈ 0.95, p = 1, ..., P.

The determination of the nearest mean vector μ _jw and the corresponding adaptation of the nearest mean vector to the current training _{data vector} (steps 307, 308) is carried out for all training data vectors in the training data group assigned to state j.

In other words, it means that in a test step (Step 309) is checked, whether the training data vector index n is equal to the value N, where with N the number of those contained in the respective status group Training data vectors is called.

If this is not the case, then in a subsequent step (step 310) the training data vector index n increased by 1 and steps 307 and 308 are for the new training data vector not yet considered n + 1 of the status group assigned to the HMM status j is carried out again.

Are all training data vectors of a status group has been checked in a subsequent step whether a predefined termination criterion, for example a predefined number of to be performed Iterations, fulfilled (step 311).

If the termination criterion is not met, a new iteration, starting in step 306, is applied to the Hidden Markov model 200 with a new adaptation coefficient a determined according to the following regulation (step 312): a p + 1 = a p · a diff

If the termination criterion is fulfilled, then in a subsequent step it is checked whether the procedure already does for all HMM states J performed has been.

In other words, it is checked in a further test step (step 313) whether the state index j is equal to the value J, with J being the number of states of the Hidden Markov model taken into account 200 referred to as.

If this is not the case, then the state index j is increased by the value 1 (step 314) and step 306, 307, 308, 309, 310 are for the training data vectors of the new state (j + 1) performed again.

In other words, it means that everyone conditions using those assigned to the respective status groups Training data vectors according to the above described procedures are trained.

Is the procedure for all J HMM states and doing it for the procedure is carried out every N training data vectors ended in a step 315.

The following will go to that described above Training methods of an HMM described two preferred modifications.

According to an alternative embodiment of the invention, it is provided that those mean value vectors that have not been determined often enough as the closest mean value vector (as a winner) within an iteration are treated separately. A threshold value _coefficient c _{giveup is} introduced which, together with the total number N of training _{data vectors of} the respective status group, defines the threshold below which an average value vector is considered to have been hit too rarely.

The threshold value N _giveup thus results according to the following regulation: N give up = c give up · N, (27) With c give up ≈ 0 ... 0.05. (28)

If an average _{vector was} not the winner more than N _giveup times after an iteration, it is replaced by the average _vector which was most often determined as the closest average _vector in the same iteration. According to the rule c _giveup = 0, an average _vector is only rejected if it has never been hit within an iteration.

This approach has in particular the advantage of better modeling of a fine structure in the Training data vectors.

wobei mit P die Anzahl der Durchläufe der Iterationen über alle N_j Trainingsdatenvektoren bezeichnet wird.According to another embodiment of the invention, the adaptation constant a is adapted differently to the total number N _j of training data vectors per state j. According to the following regulation, the adaptation constant a is chosen the smaller the more training data vectors are present:

where P denotes the number of iterations through all N _j training data vectors.

In this way it is achieved that in different states with a very different number of existing training data vectors the adaptation behavior of the cluster method in the respective status group is equal to.

wird in der Weise gewählt, dass die Adaptionskonstante a bei der ersten Iteration das 100-fache bis 10000-fache ihres Wertes bei der letzten Iteration beträgt, d.h. es gilt folgende Vorschrift:

The variable

is selected in such a way that the adaptation constant a in the first iteration is 100 times to 10,000 times its value in the last iteration, ie the following rule applies:

zu:

This is calculated

to:

Through the procedure according to the invention is the computing time required to train a Hidden Markov Model the Baum-Welch process is significantly reduced.

With a personal computer with a Clock frequency of 1 GHz, this means an effort of a few hours for training a hidden Markov model according to the invention against one Computing time for training the hidden Markov model according to the Baum-Welch method of a few weeks.

Furthermore, none occur according to the invention Numerical value problems more because the mean vectors are the same orders of magnitude assume like the training data vectors.

In the manner described above the basic entries of the electronic dictionary, this means the basic entries for speaker independent Speech recognition created and saved and the corresponding Hidden Markov models trained.

Thus there is one for each basic entry a hidden Markov model.

Using the Hidden Markov Models are in the speech recognition system in the speech recognition mode performed a view of known speech recognition.

4 shows a cellular telephone device 400 in which the in 1 shown speech recognition device 100 is integrated. Furthermore, in the mobile phone device 400 a PDA (Personal Digital Assistant) be integrated as well as other telecommunications functions, such as the sending and / or receiving of fax messages or SMS messages (Short Message Service messages) or MMS messages (Multimedia Message Service messages). Furthermore, the mobile phone device 400 can be expanded by additional multimedia functionalities, for example a camera in the mobile telephone device 400 be integrated.

5 shows a car radio 500 , in which (symbolically in 5 shown) a variety of different components are integrated, such as a navigation system 501 , a CD player 502 , a tape recorder 503 , a radio 504 , a hands-free phone 505 as well as the speech recognition device 100 as in 1 is shown. The information can both by means of the speech recognition device 100 as well as a screen 506 between the user and the car radio 500 be replaced.

Especially for controlling a system providing a multitude of different functionalities, such as a car radio provided with a multitude of different functions 500 The invention is very well suited, since an arbitrarily complicated voice dialogue structure can be set up and implemented very flexibly and independently of the speaker.

• [1] John Nerbonne and Wilbert Heeringa, Measuring Dialect Distance Phonetically, in: John Coleman (ed.) Workshop on Computational Phonology, Special Interest Group of the Association for Computational Linguistics, Madrid, 1997, S. 11 – 18
• [2] J. Nerbonne et al, Phonetic Distance between Dutch Dialects, Proceedings of CLIN '95, S. 185 – 202, Antwerpen, 1995
• [3] D. Hirschfeld, Comparing static and dynamic features for segmental cost function calculation in concatenative speech synthesis, ICSLP, Peking, 2000

The following publications are cited in this document:

• [1] John Nerbonne and Wilbert Heeringa, Measuring Dialect Distance Phonetically, in: John Coleman (ed.) Workshop on Computational Phonology, Special Interest Group of the Association for Computational Linguistics, Madrid, 1997, pp. 11-18
• [2] J. Nerbonne et al, Phonetic Distance between Dutch Dialects, Proceedings of CLIN '95, pp. 185-202, Antwerp, 1995
• [3] D. Hirschfeld, Comparing static and dynamic features for segmental cost function calculation in concatenative speech synthesis, ICSLP, Beijing, 2000

100

Speech recognition device

101

speech signal

102

microphone

103

Taped analog voice signal

104

preprocessing

105

preprocessed speech signal

106

Analog / digital converter

107

digital speech signal

108

computer

109

Input interface

110

microprocessor

111

Storage

112

Output interface

113

computer bus

114

DSP

115

keyboard

116

computer mouse

117

electric wire

118

electric wire

119

radio link

120

radio link

121

speaker

122

actuator

123

DSP

1

HMM state

2

HMM state

3

HMM state

4

HMM state

5

HMM state

α _ij

State transition probabilities

300

flow chart

301

begin

302

To capture Training data vectors

303

Group Training data vectors in state groups

304

Initialize Mean vectors of the HMM states

305

state index j = 1

306

Training data vector index n = 1

307

Determine the one closest to the training data vector n

Mean vector

308

To adjust the closest Mean vector

Training data vector

309

n = N?

310

n = n + 1

311

termination criterion Fulfills?

312

a _{p + 1} = a _p · a _diff

313

j = J?

314

j = j + 1

315

The End

400

Mobile phone device

500

car radio

501

navigation system

502

CD player

503

tape recorder

504

radio

505

Telephone device with hands-free system

506

screen

Claims (12)

Process for computer-aided training of a hidden Markov model, - at the state parameters are formed with which states of the Hidden Markov model are described, and - in which using k-nearest neighbor clustering method using training data vectors on the state parameters of the Hidden Markov model, the state parameters according to the training data vectors can be adjusted, which creates a trained Hidden Markov model becomes. Method according to claim 1, - at which the training data vectors at the beginning of the process in different Condition groups are grouped, one condition group each is assigned to a state of the Hidden Markov model, and - in which the state parameters of a respective state with training data vectors the associated Condition group to be trained. Method according to claim 1 or 2, in which the mean value vectors of describing the respective state of the hidden Markov model Probability density functions are used. Procedure according to a of claims 1 to 3, in which a k-nearest-neighbor clustering method with k = 1 executed becomes. Procedure according to a of claims 1 to 4, - at the checked is whether a state parameter is sufficiently frequent as part of the k-nearest-neighbor clustering method has been adjusted, and - at that for the case that a state parameter is not sufficiently frequent in the Framework of the k-nearest-neighbor clustering procedure has been adjusted, this will be replaced by a new state parameter, which is sufficiently frequent as part of the k-nearest-neighbor clustering process has been adjusted. Method according to claim 5, in which a state parameter is defined by the state parameter as new state parameter is replaced, which is part of the k-nearest-neighbor clustering process most frequently has been adjusted. Procedure according to a of claims 1 to 6, in which a state parameter in the context of the k-nearest-neighbor clustering method adjusted to a lesser extent the more training data vectors are available. Procedure according to a of claims 1 to 7, used in one of the following areas: - signal processing, - pattern recognition, - language processing, especially the speaker-dependent Voice recognition or speaker recognition, - image processing, - character recognition, - forecast the behavior of a technical system, which by means of hidden Markov model is described - Description of an ecological process - Description an economic one process - in environmental control, for example in the context of home automation, - in a Device control, in particular for controlling a telecommunication device. Device for training a hidden Markov model, with a processor unit which is set up in such a way that the following Process steps carried out become or feasible are: - it state parameters are formed with which states of the Hidden Markov model are described, and - under Using training data vectors becomes a k-nearest-neighbor clustering method on the state parameters of the Hidden Markov model, where the state parameters according to the training data vectors can be adjusted, which creates a trained Hidden Markov model becomes. Device according to claim 9, set up as an embedded system. Computer program element that it is a computer allows after it has been loaded into a memory of the computer, the following Steps for training a hidden Markov model: - it will State parameters are formed with which states of the Hidden Markov model are described, and - under Using training data vectors becomes a k-nearest-neighbor clustering method on the state parameters of the Hidden Markov model, where the state parameters according to the training data vectors can be adjusted, which creates a trained Hidden Markov model becomes. Computer-readable storage medium on which a program is stored, which allows a computer after it is in a Computer memory has been loaded to perform the following steps Training a hidden Markov model: - there are state parameters formed with which states of the Hidden Markov model, and - under Using training data vectors becomes a k-nearest-neighbor clustering method on the state parameters of the Hidden Markov model, where the state parameters according to the training data vectors can be adjusted, which creates a trained Hidden Markov model becomes.