DE10257475A1 - Computer-based method for determining the speech processing separation between similar sounding words using hidden Markov model distances - Google Patents

Abstract

Method for computer-based determination of the ease with which similar words can be differentiated in speech recognition technologies in which multiple realizations of different dictionary entries are generated and these are then used with dictionary entry generation models to determine a separation or difference between first and second dictionary entry generation models so that the separation between words can be characterized. Independent claims are also included for the following:- (a) a device for determination of the ease with which similar words can be differentiated in speech recognition technologies; (b) a speech recognition device; (c) a computer program storage medium and; (d) a computer program element for determining the separation or distance between similar sounding words.

Description

The invention relates to a method and a device for determining speech processing separability between a first dictionary entry and a second dictionary entry, a speech processing device, a computer-readable storage medium and a computer program element.

As part of a voice application, for example voice recognition or speaker recognition, which is implemented in an embedded system often occurs Problem on being similar to each other sounding words using a speech recognizer or a speaker recognizer cannot be separated well and thus often classified incorrectly.

One reason for this is in the optimizations the data generated for speech recognition or speaker recognition see which ones are required because of the target hardware used on the the speech application is implemented in terms of memory size and computing capacity, which to disposal standing, fixed limits. The optimizations concern in particular the design of the code book, also below as an electronic dictionary with the dictionary entries and the associated Hidden Markov Models (HMMs), each with a Hidden Markov Model a dictionary entry assigned. The code book should be as small as possible to be available in the existing memory and to match the available memory computing capacity to be processed.

For this reason, the language is usually not modeled with sufficient quality in the following sense:
Although it would make sense to use as many and as detailed as possible Hidden Markov models, this is not possible for the reasons mentioned above, so that different methods for data reduction are used. The data reduction in turn leads to the above-mentioned problem that words with a similar sound are often confused with one another, ie the words with a similar sound may have only a limited ability to be separated in the context of speech processing.

For a given electronic dictionary, i.e. For a given vocabulary, the language processing separability of the Words of the dictionary, Expressed differently dictionary entries, tested be made by voice recordings in a so-called database test be analyzed using a speech recognition system. As part of the recognized words with the words actually spoken or Dictionary entries compared. From the comparison, i.e. using the comparison results, a so-called confusion matrix is formed. In the confusion matrix is compared to a spoken word, depending on how much Percent correct the words contained in the electronic dictionary were recognized. Ideally, it always became a spoken word only the word itself recognized, i.e. in this case there was no confusion, Expressed differently no mistake on. Unfortunately, this is usually not the case. A statement can now be made on the basis of the confusion matrix determined what words yourself cheap with regard to the language processing separability within the framework of the Speech processing behave and which do not, so in this Case Ideally another word or another dictionary entry is to be used.

The problem is in the use case a speech dialogue clearly, in which in a speech dialogue the Design of the electronic dictionary special attention should be paid.

Usually a voice dialog is modeled using a dialog state machine, where a selection of words is assigned to each dialog state, which can be expressed in the respective state. It is possible, all words, the in all states of the state machine are provided in an electronic dictionary merge together. This approach increases however the likelihood of misrecognition as it is for one Viterbi decoder in an embedded system becomes all the more difficult determine the correct word, the more words he has available to find the most likely word.

So it is very efficient for everyone State of the dialog state machine its own electronic dictionary to create.

But even in this case it is appropriate selection of dictionary entries in the frame of making the electronic dictionary of considerable importance.

Suppose by means of a dialog state machine becomes a CD player controlled. The list of active words could between 1 and 20 words be long and include the following instructions: "Start, stop, Pause, rewind, next Song back End ". The list contains all necessary instructions. Unfortunately, they are similar the corresponding words in the English language "back" and "track" (play, stop, pause, track, next, back, exit) very phonetically, which can cause possible error detections can, i.e. that the word "track" was spoken, but the word "back" is decoded and vice versa.

In the context of an error detection, this error is referred to as a substitution error. It is therefore necessary to replace the word "back" in the dictionary with the word "last", with which this substitution error would no longer occur and the words in the modified electronic dictionary would now have improved speech processing separability.

In the example above, the probabilities are of a substitution error between the words "track" and "back".

In general, however, it is not so simple, such words, which have only a relatively low level of language processing separability, to investigate. For this reason, the usual procedure is a dictionary with regard to the language processing separability in this dictionary check included dictionary entries, a size Number of statements record and then a database test on the electronic dictionary perform.

Then, as described above, the confusion matrix is formed to show which word was uttered and which word was recognized. The ideal mix-up matrix, ie the mix-up matrix that describes error-free detection, is a pure diagonal matrix. 6 shows a confusion matrix 600 for the digits in German.

It should be noted that according to the confusion matrix 600 confusion has occurred between the three digits, "one", "two" and "three" and between the digits "zero" and "nine". The confusion between the digits "one", "two" and "three" are obviously due to the diphtongen "ei" common between them. However, the confusion between the numbers "zero" and "nine" is not so obvious. The confusion matrix 600 was determined after performing a database test with 615 statements totaling 6284 digits. Such a database test takes about 30 minutes on today's PC.

A disadvantage of this known method is that to every dictionary a representative collection of utterances available, i.e. must be saved, the so-called database. This means that from a large number of speakers who in their properties, the structure of the speakers of a language if possible reproduce well for every word of the electronic dictionary one or more recordings must be collected. This process requires high costs and is inflexible since every new word in the electronic dictionary in the manner described above before being added to the database needs to be edited.

The problem thus lies with the invention the language processing separability of two dictionary entries in one electronic dictionary to determine without the need for a database.

The problem is caused by the procedure and the device for determining speech processing separability between a first dictionary entry and a second dictionary entry, by a speech processing device and by a computer readable Storage medium and a computer program element solved with the features according to the independent claims.

In a method for computer-aided determination the speech processing separability between a first dictionary entry, a first dictionary entry generation model is assigned, and a second dictionary entry to which a second Dictionary entry-generation model is assigned to an electronic dictionary for electronic Speech processing using a dictionary entry generation model a realization of the respective dictionary entry can be generated from the first dictionary entry generation model a plurality of implementations, preferably a plurality of Realizations.

Using the realizations created of the first dictionary entry creation model an intra-generation model distance of the first dictionary entry generation model determined. For the Case that a hidden Markov model is used as the generation model a mean intra-HMM distance is determined as the intra-generation model distance.

Using the realizations created the first dictionary entry generation model and a plurality of implementations, preferably a plurality of Realizations of the second dictionary entry generation model becomes an inter-generation model distance between the first dictionary entry generation model and the second dictionary entry generation model determined. For the case that the generation models are set up as hidden Markov models are a medium Inter-HMM distance in this step between the two dictionary entry HMMs determined.

Using the intra-generation model distance of the first dictionary entry generation model and the inter-generation model distance a separability value is determined which is the speech processing separability the first word and the second word in an electronic Speech processing characterizes, clearly the distance of the describes both HMMs in the comparison space.

An apparatus for determining the speech processing separability between a first dictionary entry, to which a first dictionary entry generation model is assigned, and a second dictionary entry, to which a second dictionary entry generation model is assigned, in an electronic dictionary for electronic speech processing, using a dictionary entry generation model A realization of the respective dictionary entry can be produced has a processor unit which is set up in such a way that the method steps described above can be carried out are or are being carried out.

A language processing facility has a processor unit for performing the speech processing as well as a memory in which an electronic dictionary stored with dictionary entries, where each dictionary entry at least one dictionary entry creation model is assigned, using the implementation of the respective Dictionary entry can be generated, the electronic dictionary to that described above Procedure has been subjected.

There is also a control device to control a technological system with one described above Speech processing device provided, wherein in the electronic dictionary the control commands intended to control the technical system saved as dictionary entries are.

Furthermore, a telecommunications device with a Control device described above provided.

In a computer readable storage medium is a computer program stored to determine speech processing separability between a first dictionary entry, a first dictionary entry generation model is assigned and a second dictionary entry to which a second Dictionary entry-generation model is assigned in an electronic dictionary for electronic language processing, using a dictionary entry generation model a realization of the respective dictionary entry can be generated, which computer program, if it is executed by a processor unit, has the method step described above.

A computer program element for Determine the speech processing separability between a first one Dictionary entry a first dictionary entry generation model is assigned and a second dictionary entry to which a second Dictionary entry-generation model is assigned to an electronic dictionary for electronic Speech processing using a dictionary entry generation model a realization of the respective dictionary entry can be generated, has the process steps described above if it is from a processor unit is executed.

The invention makes it the first time possible, a reliable, because mathematically justified Statement about the separability of the dictionary entries in one electronic dictionary to meet and thus a minimum separability in an electronic dictionary to ensure.

According to the invention, there is no database with one Numerous predefined utterances different speakers are more necessary and so is no more database time-intensive and memory-intensive test required.

This is illustrated according to the status the database test required by the technology described above probabilistic calculation of distance measurements in the comparison space replaced, namely the calculation of the intra-generation model distance and the inter-generation model distance. According to one Embodiment of the invention creates a mean HMM distance matrix, which are the same based on the existing HMM code book Contains information like the confusion matrix according to the state of the technique.

According to the invention, a developer of a Speech dialogue, for example a developer of a speech dialogue state machine, now able to use the dictionary that he developed immediately to test after its creation. The same goes for changes, those on the electronic dictionary be made, i.e. these can also be immediately tested for their suitability be tested.

Another advantage is that this Procedures also on possible Development partner as part of a larger development project can be passed on without problems, without the sometimes complicated Ownership rights to language databases are affected.

The invention can be implemented in software, i.e. by means of a computer program or in hardware, i.e. by means of a special electronic circuit, or in any hybrid form, where part of the solution implemented in software and the other part implemented in hardware his.

Preferred developments of the invention result from the dependent Claims.

The configurations described below The invention relates to the method, the device, the control device, the telecommunication device, the speech processing device, the computer-readable storage medium as well as the computer program element.

According to an embodiment of the invention can the generation models using a respective implementation generate spoken units, especially using at least one of the following spoken units:

• - one phoneme,
• - one Phoneme,
• - one allophone
• - one diphone,
• - one Demisyllable,
• - one Syllable, or
• - one Word.

Preferably, the dictionary entry creation models are designed as Hidden Markov models.

To ensure a minimum quality for the whole electronic dictionary is the method described above according to an embodiment of the invention for all contained in the electronic dictionary Dictionary entry generation models carried out.

According to another embodiment The invention is created using a variety of realizations of the first dictionary entry generation model a mean intra generation model distance of the first dictionary entry generation model determined.

According to another embodiment the invention contemplates using a variety generated implementations of the first dictionary entry generation model and a plurality of generated implementations of the second dictionary entry generation model to determine an average inter-generation model distance.

According to these embodiments the invention is illustrative of a word of the intra-HMM distance of the assigned to the respective word HMMs determined. In addition, a pair-wise calculation of inter-HMM distances from Realizations of the first dictionary entry HMM to realizations of all others, i.e. second dictionary HMMs. Based on that the relative minimum discriminatory distance between the HMMs determined to be bigger is said to be a predetermined threshold to ensure adequate separability of the words in the electronic dictionary to care. In this case the procedure uses the middle one HMM distance, which is explained in more detail below.

The procedure becomes clear at the usual Speech recognition, in which the speech signal is first extracted using a feature is converted into a sequence of so-called feature vectors, inverted.

As part of speech recognition the sequence of feature vectors, for example using a Viterbi algorithm with the Hidden Markov models in the electronic dictionary saved words compared in the search space, and that Hidden Markov model with the greatest probability of a hit, this means with the greatest agreement, is considered the recognized word.

According to the invention, the hidden Markov models are vivid not used to recognize spoken utterances, but conversely, an HMM becomes a sequence of feature vectors generated. Such a sequence or chain of feature vectors will hereinafter also referred to as the implementation of the HMM (HMM implementation). For any two HMM implementations a distance is determined, preferably by means of a DTW method (Dynamic Time Warping algorithm) the pairwise alignment of the two feature vector sequences is determined in which the sum of the distances of the feature vectors is minimal is. The minimum distance is the distance between the two realizations. The average distance between two hidden Markov models is preferred determined by about a multitude of realizations of the two hidden Markov models, generally of the two generation models, formed an average becomes.

It is possible to change the distance between to calculate different realizations of the same HMM that is called, the intra-HMM distance, as well as the distance between different ones HMMs, the inter-HMM distance.

According to the invention, the electronic entries for all dictionary entries dictionary the distance between different realizations of the same HMM determined, i.e. the intra-HMM distance, as well as the distance between different HMMs, i.e. the inter-HMM distance. The basis of the separability between two words is the need that the inter-HMM distance should always be considerably larger than the intra-HMM distance. In other words, this means that two realizations of the same Word closer should be against each other than the realization of a word in comparison to another word.

A dictionary entry in the electronic dictionary is evaluated in terms of separability by the fact that the minimal difference between the intra-HMM distance to the inter-HMM distances in The relationship is set to the intra-HMM distance. This ratio has to be bigger as a predetermined threshold so that the word within the HMM search sufficient is easy to separate.

According to the invention it is thus possible to use words for which the Separability is not good enough from the electronic dictionary to remove and replace with another word, which is better from the others in the dictionary saved words is separable.

Thus, according to the invention, the problem the vocabulary active in speech recognition or speaker recognition to optimize so that it is as possible little like each other sounding dictionary entries included are automatically solved in a simple manner.

According to an embodiment of the invention it is intended to be checked becomes whether the separability value meets a predetermined separability criterion enough and for the case that the separability value meets the specified separability criterion not enough becomes the first dictionary entry and / or the second dictionary entry from the electronic dictionary deleted.

According to another embodiment The invention provides that it is checked whether the separability value a predefined separability criterion and in the event that the separability value meets the given separability criterion is not sufficient that the separability criterion changed is, so that also represents a poorer separability Separability value the changed Separability criterion is sufficient. This becomes clear by removing dictionary entries that too similar to each other are, from the electronic dictionary, a improved with regard to the separability in the context of language processing electronic dictionary educated. Furthermore, in order not to delete too many dictionary entries have to, also the separability criterion itself can be changed to suit the application, thus a worse one, but still sufficient to ensure good separability.

The method is preferably carried out in Framework of a computer-aided Speech processing used, preferably in speech recognition or in speaker recognition.

The method is particularly suitable for use in an embedded system, because due to the fact that no database test is required anymore and therefore none Database more stored in the memory of the speech processing system must be, even in a low-performance and limited storage capacity Environment can be used. The speech processing device is preferably set up as an embedded system. Alternatively or additionally it as a speech recognition device or as a speaker recognition device be trained.

Embodiments of the invention are shown in the figures and are explained in more detail below.

Show it

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

2 a sketch of an electronic dictionary according to an embodiment of the invention;

3 a sketch of a comparison space in which implementations of dictionary entry HMMs and implementations of dictionary entry candidate HMMs according to a first embodiment of the invention are shown;

4 a confusion matrix according to an embodiment of the invention;

5A to 5D a flow chart in which the individual method steps of an embodiment of the invention are shown; and

6 a confusion matrix according to the prior art.

1 shows a speech recognition system 100 according to an embodiment of the invention.

The speech recognition system 100 Depending on the operating mode, it works in a first operating mode as a speech recognition device, with a spoken utterance in the speech recognition mode 101 , spoken by a user (not shown) of the speech recognition system 100 is recognized by the speech recognition device. According to this exemplary embodiment of the invention, speech recognition takes place using a method for speaker-independent speech recognition, alternatively using a method for speaker-dependent speech recognition.

In a second operating mode, also referred to below as training mode, using a spoken utterance 101 , as will be explained in more detail below, the speech recognition system 100 trained, according to this embodiment, this means that individual hidden Markov models for an utterance by means of the spoken utterance 101 be 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 preamplification by means of a preamplification unit 104 subjected and as an amplified analog signal 105 an analog / digital converter 106 fed, there in a digital signal 107 converted and as a digital signal 107 a feature extraction unit 108 fed which is the digital signal 107 undergoes a spectral transformation and to the digital signal 107 a sequence of feature vectors for an utterance 109 which forms the digital signal 107 represent.

Every feature vector 109 has a predetermined number of feature vector components.

The feature vectors 109 become a calculator 110 fed.

It should be noted in this context that the microphone 102 , the preamp unit 104 , in particular the amplification unit, and the analog / digital converter 106 as well as the feature extraction unit 108 as separate units or as in the computer 110 integrated units can be realized.

According to this exemplary embodiment of the invention, it is provided that the feature vectors 109 the calculator 110 via its input interface 111 are fed.

The computer 110 also has a microprocessor 112 , a memory 113 as well as an output interface 114 on all of them together using a computer bus 115 are coupled.

Using the microprocessor 112 process steps described below guided.

In an electronic dictionary 200 (see. 2 ) which is in memory 113 are saved to every word entry 201 (Words i, i = 1, ..., n, where n denotes the number of words stored in the electronic dictionary), which can only be recognized by the speech recognition algorithm in the context of speech recognition, corresponding specifically to this word entry trained Hidden Markov models 202 (Hidden Markov models i (H (word i)) included or saved.

A digital signal processor can also be used be provided, the speech recognition algorithms used in each case has implemented and a specialized microcontroller can have.

Furthermore, a text-to-speech converter device (not shown) is provided for converting any phoneme sequence into a speech signal, which can be converted into an analog speech signal and then using the loudspeaker described below 122 can be issued to a user. The text-to-speech converter can be in the microprocessor 112 or be integrated in the digital signal processor or be provided as an independent hardware component.

Furthermore, the calculator 110 using the input interface 111 with a keyboard 116 as well as a computer mouse 117 via electrical cables 118 . 119 or a radio connection, for example an infrared connection or a Bluetooth connection.

Via additional cables or radio connections, for example using an infrared connection or a Bluetooth connection 120 . 121 is the calculator 110 using the output interface 114 with a speaker 122 as well as an actuator 123 coupled.

The actuator 123 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 110 itself or any other technical system to be controlled.

According to the exemplary embodiment of the invention, the feature extraction unit has 108 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, according to the exemplary embodiment every 10 msec, alternatively every 15 msec.

That by means of the feature extraction unit 108 formed cepstrum coefficients, the 13 coefficients of the feature vectors 109 are formed as feature vector components of two successive time windows of the size of 10 msec or 15 msec in the feature vector 109 saved. Furthermore, as feature vector components in the feature vector 109 in each case the temporal first derivative and the temporal second derivative of the cepstrum coefficients in the feature vector 109 summarized as a super feature vector and are the computer 110 fed.

In the calculator 110 a speech recognition unit is implemented in the form of a computer program and in the memory 113 stored, which speech recognition unit is based on the principle of the Hidden Markov models. A speaker-independent speech recognition thus takes place by means of the computer program.

Using the respective hidden Markov models 202 , each with a hidden Markov model 202 a dictionary entry 201 in the electronic dictionary 200 Realizations are uniquely assigned, that is to say a sequence of according to the probability information for the generation probabilities and the state transition probabilities of the respective Hidden Markov model 202 generated phonemes.

3 shows a comparison room 300 , in which the distance calculation for calculating the distance between two implementations of two HMMs is usually carried out in a clear sketch.

The individual hidden Markov models 202 are in the form of clusters 301 in 3 shown and symbolize a plurality of realizations of the respective HMMs 202 what realizations within the comparison space 300 according to the respective probability distribution in a given local environment around the prototype (s) of the respective HMM 202 lie.

In the 5A to 5D are in a flow chart 500 the individual process steps of the embodiment of the invention are shown.

The method according to the invention only uses the information that is already available in the code book anyway. A basic idea of the invention can be seen in that as below explained is calculated, the mean intra-HMM distance of each dictionary entry and with the word spacing of the dictionary entry to the other dictionary entry HMMs is compared.

After starting the method (step 501), a threshold value Θ _{S is set} with the relative minimum separability distance (step 502).

Then a sentence, ie a plurality of dictionary entries in the electronic dictionary 200 entered or adapted to predetermined requirements (step 503), which means a set of Dictionary entries result according to the following rule: W = {w 1 , w 2 , ..., w n }. (1)

As described above, each dictionary entry w _i is assigned a hidden Markov model, which represents the corresponding dictionary entry w _i , in accordance with the following rule: H = {H 1 , H 2 ,..., H n }. (2)

In a subsequent step (step 504) the electronic dictionary becomes 200 a dictionary entry 201 selected and in the event that there is still no HMM for the dictionary entry 201 is formed, the hidden Markov model for the dictionary entry is formed, ie trained (step 505).

Is the dictionary entry 201 already an HMM as a dictionary entry HMM 202 assigned, the corresponding dictionary entry HMM 202 selected in step 505.

To ensure that each dictionary entry has a dictionary entry HMM clearly assigned to it and trained on it 202 is checked in a subsequent step (step 506) whether a corresponding dictionary entry HMM has already been formed for all dictionary entries.

If this is not the case, then a next dictionary entry from the electronic dictionary 200 selected, for which no dictionary HMM 202 has been formed (step 507) and an associated HMM is trained and thus formed for the selected dictionary entry.

Steps 505, 506 and 507 are thus for all dictionary entries of the electronic dictionary 202 carried out.

Are for all dictionary entries 201 in the electronic dictionary 200 Dictionary entry HMM 202 is formed, then in a subsequent step (step 508) from the electronic dictionary 200 a dictionary entry HMM is selected as the first dictionary entry HMM.

Using the selected first Dictionary entry HMM will be a variety of realizations of the dictionary entry HMM is formed and using the realizations determined Intra-HMM distance, i.e. the distance between realizations of the first Dictionary Entry HMM calculated (step 509).

Determining the distance between two implementations of an HMM take place in a corresponding manner, as below explains how to determine the distance between two realizations different HMM.

The respective intra-HMM distance of the chosen first dictionary entry HMMs H ^ ∈ H becomes hereinafter referred to as d (H ^).

In a next step (step 510), another dictionary entry HMM, that is different from the selected first dictionary entry HMM, becomes the electronic dictionary 200 is selected as the second dictionary entry HMM and, using a plurality of implementations of the first dictionary entry HMM and a plurality of implementations of the selected second dictionary entry HMM, the mean inter-HMM distance between the first dictionary entry HMM is explained in more detail below and the second dictionary entry HMM (step 511).

To illustrate the procedure for Calculate the mean HMM distance (i.e. both the mean Intra-HMM distance as well as the mean inter-HMM distance) first briefly dealt with some basics in speech recognition.

The goal of speech recognition is to translate a spoken utterance into a sequence of symbols such as words or phonemes. Conventional speech recognition systems based on HMMs are set up such that an input speech signal, as explained above, is sampled and the sampled values s (t) are mapped into a sequence of feature vectors f (t) by means of the above-described method of Extraction, whereby the sequence of feature vectors f (t) results according to the following rule: s (t) → f (T). (3)

The sequence of the feature vectors is mapped into a sequence of symbols w _k using the Viterbi decoding algorithm in accordance with the following regulation: f (t) → w k , (4)

The Viterbi decoding method is based on the hidden Markov models. It should be noted that a time parameter is denoted by t, the index k being used to denote that an asynchronous relationship exists at time t. The Viterbi decoding method is described in more detail below explained.

Hidden Markov models are stochastic models that describe the relationship between a sequence of feature vectors and a symbol, that is to say that a symbol w _k corresponds to a hidden Markov model H _k . The expression "hidden" results from the "true" symbol w T / k, which is unknown per se and which was initially uttered, and from which the sequence of feature vectors was formed.

The statement is shown as the following figure: w T k → s (t). (5)

This statement should be recognized by means of the speech recognition system. In the ideal case, the determined symbol w _k , which was determined by means of Viterbi decoding, is equal to the expressed symbol w T / k.

An HMM can be used to Language using a variety of different spoken language Model units, such as words, syllables, half-syllables, phonemes, down to phoneme segments.

Conversely, a word model can be constructed using an HMM by concatenating phoneme segments, phonemes, half-syllables, syllables to words. Thus the dictionary entries w _k 201 representing the searchable and recognizable words found in the electronic dictionary 200 are stored, by means of the respective HMM H _k 202 modeled.

From a mathematical point of view HMMH the following items

• 1. The number n _{H of} the HMM states,
• 2. The state transition probability a ^H (j | i) for a transition from an initial state i to a final state j, and
• 3. The emission probabilities p ^H ( f | i) for emitting a feature vector f in the state i.

wobei mit c H / il die Mischgewichte, mit (σ H / il)² die Mischvarianzen und mit x H / il die Misch-Zentren bezeichnet werden. Die Vektoren x H / il werden auch als Prototypen bezeichnet. Die Gesamtheit aller Prototypen und Mischgewichte aller gespeicherten HMMs wird als Codebuch bezeichnet.The emission probabilities p ^H ( f | i) are modeled as continuous Gaussian mixed probability densities according to the following regulations:

where c H / il denotes the mixing weights, (σ H / il) ² the mixing variances and x H / il the mixing centers. The vectors x H / il are also referred to as prototypes. The entirety of all prototypes and mixed weights of all stored HMMs is referred to as a code book.

According to this embodiment the following simplifications are assumed.

First, the state transition probabilities are set in such a way that they essentially model independent values, that is, the following applies: a H (j | i)) = a trans (8th) where trans represents either a self-loop (j = i), a one-step state transition (j = i + 1) or a two-step state transition (skip, j = i + 2) or larger state transitions, i.e. state - Transitions with more than one skipped subsequent state (j> i + 2 and j <i).

und aother = 0. (10) To simplify the calculation, the following is assumed without restricting the general validity:

and a other = 0. (10)

This choice of parameters enables state transition probabilities to be late at any time Ignore their calculation provided that there is no state transition other than a self-loop, a one-step state transition, and a two-step state transition.

In addition, the variances of all mixed probability densities are set to a fixed value, that is, the following applies: σ H il = σ. (11)

wobei mit c ~ H / iw der skalierte Logarithmus des Mischgewichts und mit w der "beste" Prototyp bezeichnet wird, der sich ergibt gemäß folgender Vorschrift:

In this case, it is very easy to calculate a normalized negative logarithm of the probability, which is approximated according to the following rule:

where c ~ H / iw denotes the scaled logarithm of the mixed weight and w denotes the "best" prototype, which results from the following rule:

The variable q ^H ( f | i) is called the local result.

For a given HMM H becomes the global result in the following way calculated.

Alignment A is given in accordance with the following regulation: A ( f (t)) = i, (14) that is, a mapping of each feature vector to the sequence of an HMM state, the sum of all local results g ^H ( f (t) | i) in the alignment A ( f (t)) in the variable Q ^H ( f (t ), A) is saved. The minimum of the total sums is called the global result Q ^H ( f (t)) and results according to the following rule:

It should be noted that the optimal alignment A ^H corresponds to the global result Q ^H ( f (t)).

Using the Viterbi algorithm used for speech recognition, the HMM H ^{* is} removed from the list of all in the dictionary 200 contained HMMs H _k 202 determined which has the lowest global result corresponding to a given sequence of feature vectors f (t), that is, the following rule is sufficient:

The method described above is called Viterbi decoding and H ^* is called the recognition result.

The mean HMM distance is based on the reversal of the method described above for determining the best HMM H ^* in the context of speech recognition.

Starting with and using a given HMM H, sequences of feature vectors f (t), referred to as realizations of an HMM, are generated and compared with one another.

It is believed that an HMMH n states has and is assigned a symbol w and those previously described Simplifications is subject.

According to this exemplary embodiment, a realization of the hidden Markov model H, that is to say a sequence f (t) of feature vectors, is generated by carrying out the following method, which is also referred to as Monte Carlo scanning:
First, the state i = 1 of the HMM H is started and the variable t = 1 is set.

Then the 1st of the m H / i probability densities according to the mixed weights c H / il selected.

The emission of the feature vector f (t) is then generated by scanning the Gauss probability PH / il ( f ).

Then the variable t turns around increases the value 1.

A fifth step is to choose whether for the next Time step as part of the state transition a self-loop, a one-step state transition or a two-step state transition carried out and the value i becomes corresponding to the selected state transition customized.

If i <n then return to step 2 otherwise T = t - 1 set and the procedure is ended.

In this way, a sequence of feature vectors f (t) is generated, the number T of generated feature vectors f (t) depending on n, but is usually not equal to n.

The distance between two implementations, that is to say between two feature vectors f ₁ (t) and f ₂ (t) of any two HMMs, is calculated in the following way.

First, in the same way like Viterbi decoding an alignment A, as described above, between the two sequences formed by feature vectors, that is, a mapping takes place each vector of a first feature vector sequence onto a vector one second feature vector sequence to be compared.

Alignment A has according to this embodiment the following properties:

• 1. Alignment A is clearly defined, that is, A applies: f ₁ (t) → f ₂ (t) exists for all f ₁ (t), and
• 2. Alignment A is injective, ie for all f ₂ (t) there exists f ₁ (t) with A: f ₁ (t) → f ₂ (t).

It should be noted that usually T ₁ ≠ T ₂ and therefore the mapping is not necessarily bijective.

Then the given alignments A of the Euclidean vector distances | f ₁ (t) - A · f ₁ (t)) | ² to the opening gap _A d (A, f ₁ (t), f ₂ (t)) using the following rule accumulated:

The minimum of all accumulated distances is the distance d ( f ₁ (t), f ₂ (t)) between the two HMM implementations f ₁ (t) and f2 (t) according to the following rule:

This value can be calculated in a simple manner using the DTW method (Dynamic Time Warping method), which fulfills the properties described above. It should be noted that the definition of the distance between two HMM implementations is similar to the calculation of the global result of a sequence f (t) with respect to an HMM H when the scaled mixture weight is set to zero.

wobei gilt: f _a(t) ⊂ H_a. Diese Vorschrift ist eine Bezeichnung des Ausdrucks "f _a(t) ist eine Realisierung des HMMs H_a". Es ist anzumerken, dass die mittlere Intra-HMM-Distanz d(H) definiert ist gemäß folgender Vorschrift: d(H) = d(H, H). (20) The mean HMM distance (mean HMM distance) between two HMMs is defined by forming the mean over the distances between realizations of the two HMMs, namely according to the following rule:

where: f _a (t) ⊂ H _a . This regulation is a designation of the expression " f _a (t) is a realization of the HMM H _a ". It should be noted that the mean intra-HMM distance d (H) is defined according to the following rule: d (H) = d (H, H). (20)

The mean intra-HMM distance needs not to be zero.

Hence the mean inter-HMM distance between the first dictionary entry HMM and the second dictionary entry HMM formed (step 511). In the same way, the respective mean intra-HMM distance calculated in step 509.

It is then checked whether the respective inter-HMM distance from the selected first dictionary entry HMM to all dictionary entry HMMs otherwise stored in the electronic dictionary has been determined (step 512), and if this is not the case, the next one another dictionary entry HMM from the electronic dictionary 200 selected as the second dictionary entry HMM (step 513) and the inter-HMM distance between the selected first dictionary entry HMM and the now second dictionary entry HMM is determined for this selected dictionary entry HMM (step 511).

In other words, an inter-HMM distance is determined from the selected first dictionary entry HMM to all other dictionary entry HMMs 202 in the electronic dictionary 200 ,

The mean inter-HMM distance is hereinafter referred to as d (H ^, H ') designated.

Is the inter-HMM distance d (H ^, H ') from the selected first dictionary entry HMM to all other dictionary entry HMMs of the electronic dictionary 200 that second dictionary entry HMM is determined which is closest to the selected first dictionary entry HMM in the search space used, ie its inter-HMM distance d (H ^, H ') is minimal. In other words, the second dictionary entry HMM H is determined according to the following rule:

The closest second dictionary entry HMM is saved in a subsequent step (step 514) (compare 5C ).

The difference between the mean inter-HMM distance between the determined second dictionary entry HMM and the selected first dictionary entry HMM and the intra-HMM distance of the selected first dictionary entry HMM is then determined and stored (step 515):

Using the determined difference, a relative separability distance is calculated in a subsequent step (step 516) and stored in accordance with the following rule:

If the determined and stored relative separability distance is greater than the predetermined threshold value of the relative minimum separability distance Θ _S (test step 517), then the two checked dictionary entries, that is to say the two checked dictionary entry HMMs, are sufficiently dissimilar to one another, that is to say in the search space spaced apart sufficiently and the first dictionary entry HMM and associated dictionary entry are accepted (step 518).

so wird das erste Wörterbucheintrags-HMM und der zugehörige Wörterbucheintrag als kritisch markiert.However, if the relative separability distance is not greater than the specified threshold value, ie the following requirement is not met:

the first dictionary entry HMM and the associated dictionary entry are marked as critical.

It is then checked whether the intra-HMM distance already exists for all dictionary entry HMMs and thus for all dictionary entries in the electronic dictionary 200 has been determined (step 520).

If this is not the case, then another one Dictionary entry HMM as the first dictionary entry HMM selected, for the an intra-HMM distance has not yet been determined (step 521). For the as the new first dictionary entry HMM selected Dictionary entry HMM steps 503 to 520 are carried out again.

In other words, this means that an intra-HMM distance for each dictionary entry HMM and, in addition, a respective inter-HMM distance to all other dictionary entry HMMs of the electrical dictionary 200 has been determined.

The result of the method described above is that in 4 Confusion matrix created according to the invention 400 ,

The most important results of the determination of the confusion matrix according to the invention 400 versus the confusion matrix 600 according to the state of the art are briefly summarized below:

• 1. The inventive, on the Monte Carlo process based procedure for determining a mix-up matrix also represents an elevated Risk of confusion between the two digits "zero" and "nine" and also reflects correctly, that the number "zero" is often confused with the number "nine" conversely, the number "nine" with the number "zero".
• 2. According to the invention, according to the confusion matrix determined 400 also correctly stated that the number "two" is confused with the number "three" more often than with the number "one". Furthermore, it is correctly stated that the number "three" is usually mistaken for the number "two" However, according to this exemplary embodiment, there are two discrepancies with the confusion matrix 600 According to the prior art: According to the exemplary embodiment of the invention, it is indicated that the number "two" is confused with the number "two", which is not the case, however. Furthermore, the number "one" is allegedly confused with the number "eight", which is also not the case.
• 3. The confusion matrix determined according to the invention also indicates 400 assumes that the number "two" is very often confused, which is not the case in reality.

In summary, it can be stated that in some small aspects the Monte Carlo method according to the invention is not completely reliable, but does essentially reflect correct modeling. The one in the confusion matrix 400 discrepancies to be found can be attributed to simplified assumptions in the model calculation within the inter-HMM distance: the global variance was set to the value σ = 0, so that only the prototypes were used in the creation of realizations and no Gaussian sampling in the Probability distribution was applied. Furthermore, the mixed weights assigned to the prototypes were ignored, which leads to an overrepresentation of the unlikely prototypes and an underrepresentation of the probable prototypes. Furthermore, according to this exemplary embodiment, the number of states of all HMMs was predefined, ie each dictionary entry HMM had 18 States (3 × 3 × 2), where the first factor describes the number of phonemes, the second factor the number of considered phoneme segments and the third factor a doubling of the states in each Hidden Markov model. Furthermore, the distances were not calculated using the DTW method because this was not necessary due to the same number of states in the Hidden Markov models.

In a subsequent step (step 522), it is now checked whether individual or all critical dictionary entries from the electronic dictionary 200 allowed to be removed (step 522).

If this is the case, the or will be allowed dictionary entries to be removed and the associated dictionary entry HMMs from the electronic dictionary removed (step 523) and, if necessary, by others, phonetically more suitable dictionary entries and with associated Dictionary entry HMM replaced.

Subsequently or in the event that no dictionary entry from the electronic dictionary marked as a critical dictionary entry 200 may be removed, it is checked whether the threshold value for the relative minimum separability distance Θ _S should be changed (step 524).

If this is the case, then in one subsequent step (step 525) the threshold for the relative Minimum distance Trennbarkeits changed and the process begins again in step 503.

Are neither dictionary entries from the electronic dictionary marked as critical 200 If the threshold value for the relative minimum separability distance is still to be changed, the method is ended (step 526).

100

Voice recognition system

101

utterance

102

microphone

103

analog signal

104

preamplification

105

amplified analog signal

106

Analog / digital converter

107

digital signal

108

Feature extraction unit

109

feature vectors

110

computer

111

Input interface

112

microprocessor

113

Storage

114

Output interface

115

computer bus

116

keyboard

117

computer mouse

118

electrical management

119

electrical management

120

connection

121

connection

122

speaker

123

actuator

200

electronic dictionary

201

word entry

202

Hidden Markov model

300

Compare room

301

Dictionary entry HMM

302

Fault word HMM

303

middle HMM-distance

400

confusion matrix

500

flow chart

501

begin

502

Put Threshold for relative minimum separability

distance

503

Enter or customize dictionary entries in

electronic dictionary

504

Select dictionary entry from dictionary

505

Form HMM for chosen Dictionary entry

506

Formed HMM for all dictionary entries?

507

Select the next dictionary entry from dictionary

508

Select dictionary entry HMM from dictionary as

first Dictionary entry HMM

509

To calculate Intra-HMM distance of the selected one first

Dictionary entry HMM

510

Choose another Dictionary entry HMM from dictionary

as second dictionary entry HMM

511

To calculate mean inter-HMM distance between

selected first Dictionary entry HMM and second

Dictionary entry HMM

512

For every second Dictionary entry HMM Inter-HMM

distance determined?

513

Select next other Dictionary entry HMM as

second Dictionary entry HMM

514

Determine the second dictionary entry HMM, that to

selected first Dictionary entry HMM the next

lies

515

Determine and save difference from mean inter-

HMM-distance between the determined second to the

selected first Dictionary entry HMM and the intra-

HMM-distance of the selected one first dictionary entry

HMM

516

Determine and storing relative separability distance

517

relative Separability distance> threshold?

518

Accept first dictionary entry HMM and related

Dictionary entry

519

To mark first dictionary entry HMM and related

Dictionary entry as critical

520

For everyone first Dictionary entry HMM Intra-HMM-distance

determined?

521

Select next dictionary entry HMM first

Dictionary entry HMM

522

May be critical Dictionary entries from dictionary

away become?

523

remove critical dictionary entries from dictionary

524

Change threshold for relative Minimum-

Trennbarkeits distance?

525

Change threshold for relative Minimum-

Trennbarkeits distance

526

The End

600

confusion matrix

Claims (20)

Method for computer-aided determination of the speech processing separability between a first dictionary entry, to which a first dictionary entry generation model is assigned, and a second dictionary entry, to which a second dictionary entry generation model is assigned, in an electronic dictionary for electronic speech processing, using a word a realization of the respective dictionary entry can be generated in the book entry generation model, - in which a plurality of implementations are generated by the first dictionary entry generation model, - in which, using the realized implementations of the first dictionary entry generation model, an intea generation model distance of the first dictionary entry Generation model is determined, in which an inter-generation model distance between the first dictionary entry generation model and the second dictionary entry generation model is determined using the generated implementations of the first dictionary entry generation model and a plurality of implementations of the second dictionary entry generation model, and - In which a separability value is determined using the intra-generation model distance of the first dictionary entry generation model and the inter-generation model distance, which does the language processing Characterized gs separability of the first word and the second word in the context of an electronic language processing. Method according to claim 1, in which the generation models are implemented under Generate use of spoken units. Method according to claim 1 or 2, in which at least one of the the following spoken units are used: - a phoneme segment, - a phoneme, - an allophone, - a diphone, - a half syllable, - A syllable, or - on Word. Procedure according to a of claims 1 to 3 in which the dictionary entry generation models Hidden Markov models are. Procedure according to a of claims 1 to 4, which for all in the electronic dictionary included dictionary entry generation models carried out becomes. Procedure according to a of claims 1 to 5, using a variety of realizations of the first dictionary entry generation model an average intra generation model distance of the first dictionary entry generation model is determined. Procedure according to a of claims 1 to 6, using a variety of realizations of the first dictionary entry generation model and a plurality of generated implementations of the second dictionary entry generation model an average inter-generation model distance is determined. Procedure according to a of claims 1 to 7, - at the checked becomes whether the separability value meets a predetermined separability criterion enough, and - at that for the case that the separability value meets the specified separability criterion not enough the first dictionary entry and / or the second dictionary entry from the electronic dictionary will be deleted. Procedure according to a of claims 1 to 7, - at the checked becomes whether the separability value meets a predetermined separability criterion enough, and - at that for the case that the separability value meets the specified separability criterion not enough that Separability criterion changed is, so that also a separability value representing poorer separability the changed Separability criterion is sufficient. Process for computer-aided speech processing under Use of a method according to one of claims 1 to 9 submitted electronic dictionary, used in speech recognition or in speaker recognition. Procedure according to a of claims 1 to 10, used in an embedded system. Device for determining speech processing separability between a first dictionary entry, a first dictionary entry generation model is assigned, and a second dictionary entry to which a second Dictionary entry-generation model is assigned to an electronic dictionary for electronic Speech processing using a dictionary entry generation model a realization of the respective dictionary entry can be generated can, with a processor unit that is set up such that the following process steps can be carried out or are being carried out: - of the first dictionary entry generation model a plurality of realizations are generated, - under Using the generated implementations of the first dictionary entry generation model becomes an intra-generation model distance of the first dictionary entry generation model determined, - under Using the generated implementations of the first dictionary entry generation model and a plurality of implementations of the second dictionary entry generation model becomes an inter-generation model distance between the first dictionary entry creation model and the second dictionary entry generation model determined, and - under Using the intra-generation model distance of the first dictionary entry generation model and the inter-generation model distance becomes a separability value that determines the speech processing separability of the first word and the second word in the context of electronic language processing characterized. Speech processing device with a processor unit to perform of speech processing and with a memory in which an electronic dictionary stored with dictionary entries, where each dictionary entry at least one dictionary entry creation model is assigned, using the implementation of the respective Dictionary entry can be generated, the electronic dictionary using the following method has undergone: - of the first dictionary entry creation model a plurality of realizations are generated, - under Using the generated implementations of the first dictionary entry generation model becomes an intra-generation model distance of the first dictionary entry generation model determined, - under Using the generated implementations of the first dictionary entry generation model and a plurality of implementations of the second dictionary entry generation model becomes an inter-generation model distance between the first dictionary entry creation model and the second dictionary entry generation model determined, and - under Using the intra-generation model distance of the first dictionary entry generation model and the inter-generation model distance becomes a separability value that determines the speech processing separability of the first word and the second word in the context of electronic language processing characterized. Speech processing device according to claim 13, set up as an embedded system. Speech processing device according to claim 13 or 14, set up as a speech recognition device. Speech processing device according to claim 13 or 14, set up as a speaker recognition device. Control device for controlling a technical System with a speech processing device according to one of claims 12 to 16, the electronic dictionary used to control the Technical commands provided control commands saved as dictionary entries are. Telecommunication device with a control device according to claim 17th Computer-readable storage medium in which a computer program is stored for determining the language processing separability between a first dictionary entry to which a first dictionary entry generation model is assigned and a second dictionary entry to which a second dictionary entry generation model is assigned in an electronic dictionary for electronic speech processing , wherein a realization of the respective dictionary entry can be generated using a dictionary entry generation model, which, if it is executed by a processor unit, has the following method steps: - a plurality of implementations are generated by the first dictionary entry generation model, - using the realizations of the first dictionary entry generation model, an intra-generation model distance of the first dictionary entry generation model is determined, Using the generated implementations of the first dictionary entry generation model and a plurality of implementations of the second dictionary entry generation model, an inter-generation model distance between the first dictionary entry generation model and the second dictionary entry generation model is determined, and - using the intra-generation model -Distance between the first dictionary entry generation model and the inter-generation model distance, a separability value is determined, which characterizes the speech processing separability of the first word and the second word in the context of electronic speech processing. Computer program element for determining the speech processing separability between a first dictionary entry, a first dictionary entry generation model is assigned, and a second dictionary entry to which a second Dictionary entry-generation model is assigned to an electronic dictionary for electronic Speech processing using a dictionary entry generation model a realization of the respective dictionary entry can be generated which, when executed by a processor unit, has the following process steps: - from the first dictionary entry creation model a plurality of realizations are generated, - under Using the generated implementations of the first dictionary entry generation model becomes an intra-generation model distance of the first dictionary entry generation model determined, - under Using the generated implementations of the first dictionary entry generation model and a plurality of implementations of the second dictionary entry generation model becomes an inter-generation model distance between the first dictionary entry creation model and the second dictionary entry generation model determined, and - under Using the intra-generation model distance of the first dictionary entry generation model and the inter-generation model distance becomes a separability value that determines the speech processing separability of the first word and the second word in the context of electronic language processing characterized.