EP0815553B1 - Method of detecting a pause between two signal patterns on a time-variable measurement signal - Google Patents

Description

Pattern recognition is achieved in many technical processes is becoming increasingly important as this increases the degree of automation can be achieved. Let pattern recognition processes are usually reduced to a time-varying measurement signal, which from the patterns to be recognized in a suitable manner is derived. In the automatic analysis of these measurement signals However, the problem arises that these measurement signals not in a pure form, but often from stationary ones or are overlaid by transient interference signals. When examining measurement signals, which of course spoken language can be derived, these interference components the measurement signal, for example by background noise, Breathing noises, machine noises, or even through the recording medium and the transmission path become. Because the measurement signal is never in pure form it is particularly important between the components of the measurement signal, which contain the pattern to be recognized and between other proportions in which there is no pattern differentiate. So it is for better recognition of the patterns especially important to know exactly when pattern in the measurement signal and when there are no samples, i.e. not from the pattern originating signals are present as pause signals in the measurement signal are.

A pause detection is e.g. also important to a reduction the amount of data transmitted, for example in the case of voice communication channels and also in satellite broadcasting to achieve the general useful interference signal decision signal processing, or the end of an utterance to be found in automatic speech recognition systems. Here A robust pause detector serves to improve performance of voice-controlled systems. Especially applies this for speech recognition systems since it is about a spoken utterance as a pattern with an existing one Compare version. The problem of determining breaks is particularly detailed in automatic speech recognition described by Rabiner [1]. It also has an algorithm specified for break detection. There are pause detection Information taken into account which comes straight from the sampled time signal (energy, Zero crossing rate ETC.). This approach is everyone known pause detectors together [2]. You use in the Rule a more or less complicated set of rules to the Classification of the break from the calculated characteristics perform. Statistical classifiers were also used as an alternative used [3]. Because of this procedure, all these procedures only work up to a certain noise level. The limit depends on the type of disturbance. she can no longer be used with low signal-to-noise ratios be because pause detectors are usually threshold controlled. In noisy environments at very However, the common signal-to-noise ratios fail decision criteria with thresholds. Join in non-stationary disturbances with a signal-like character, the are barely detectable.

Use previous approaches for language break regulations for example a local, i.e. based on a temporal or spectral frame information Detection of signal and non-signal areas [4,5]. Also recent works published on this are based on the first Line on modifications or extensions of this work. Further procedures for pause detection in time variants Measurement signals are not known.

The object underlying the invention is a improved method for pause detection between patterns specify which are present in a measurement signal and which were modeled with the help of hidden Markov models.

This object is achieved in accordance with the features of patent claim 1 solved.

Further developments of the invention result from the subclaims.

An advantage of the method according to the invention is that that for the first time information in different signal processing levels be won and the one after the other occur for pause detection. The means the pause information is compared by comparing one special pause model with the feature vectors of the measurement signal won in a comparison level and at the feature extraction level the pattern recognition, so that in another time slice in the feature extraction level the pause status is taken into account in the measurement signal analysis can be.

The method according to the invention advantageously uses the Information that certain sample groups belong together, for example, in the case of words, these are phoneme pattern groups it is ensured that at least according to the sample group there must be a pause. This information will follow advantageously first in the feature extraction level Processing stage of the process exploited.

Advantageously, the method according to the invention also ensured that one to be recognized before arrival Sequence must have been a break. This fact is also used in pattern recognition.

The method according to the invention can advantageously also be used known methods for pause detection can be combined, what properties of the measurement signal in the time domain and in Evaluate the spectral range. This way, a higher one Detection rate in pattern recognition can be achieved.

Can be particularly advantageous with the inventive method Speech patterns, writing patterns or signaling patterns are analyzed as they are used in a wide range of technical applications occur and be appropriately modeled can.

The method according to the invention can advantageously be used ensure that if no patterns are detected, there must be a pause, this will increase Detection rate achieved in the pattern recognition, since with it the Feature extraction level makes pause information even more reliable can be made available.

In the following, the invention will be further explained with reference to figures explained.

FIG. 1 shows a schematic example of a speech recognition system equipped with pause recognition.
Figure 2 illustrates the pause detection process using various hidden Markov models.

a) durch eine gemeinsame Bewertung vieler Analyseintervalle und

b) durch eine Erkennung der Nutzsignale, wobei alle Signale, die nicht als Nutzsignal erkannt werden, beispielsweise dem Rauschen zugeordnet werden können. Vorteilhafterweise kann die Erfindung bei allen bekannten Mustererkennungsverfahren angewendet und mit diesem kombiniert werden.

FIG. 1 shows, using an example that is implemented here as a speech recognition system, how the pause information is detected and passed on, that is to say returned, using the method according to the invention. The measurement signal here as a speech signal Spr first reaches a feature extraction stage Merk which corresponds to the first signal processing stage in the method according to the invention. In this first signal processing stage, the spectral features of the speech signal or the measurement signal Spr are usually analyzed. These features, which are subsequently output by the feature extraction level, are designated here by m in FIG. 1. The spectral features m then arrive, for example, as feature vectors in a classification level Klass, in which they are compared with the hidden Markov models HMM. This is where the method according to the invention comes in, by comparing the feature vectors obtained from the measurement signals in special hidden Markov models for individual phonemes or for pause states. In the training phase of the hidden Markov models, for example, typical feature vectors for the background noise and for the useful signal are estimated. This makes it possible for a continuous pattern comparison to distinguish between useful and noise signals in every analysis interval. You get an even higher robustness with a very bad signal-to-noise ratio

(a) by evaluating many analysis intervals together and

b) by recognizing the useful signals, it being possible for all signals that are not recognized as useful signals to be assigned to the noise, for example. The invention can advantageously be applied to and combined with all known pattern recognition methods.

The method according to the invention is based in particular on that the signal states and the feature vectors of one time slice to the other time slice of the analysis interval do not change excessively. Thus, information can is obtained in the classification class Klass by, for example it is found that when comparing the hidden Markov models more likely to pause than for a pattern to be recognized, to the feature extraction level are forwarded as pause information Pa. It’s very likely that the time slice in which the pause is detected, another time slice is included Follow pause. This procedure allows the measurement signal existing undesirable disturbances in the formation of the feature vectors even with a low signal-to-noise ratio great security can be suppressed. Will be advantageous through the inventive method in the recognition stage in a second time slice of knowledge about the break a first signal processing stage is transmitted. This Knowledge can, for example, from a speech signal about the acoustic phonetic modeling level (hidden Markov models), who already have a lot of training data for speech recognition has been trained. In Phonebased systems is the pause as a model Phonems trained and thus includes the statistics of Training data. Modeling is more refined and therefore better taking into account the phoneme context, i.e. the Know which phoneme is following another. One links for example the pause decision of the acoustically phonetic Modeling level with common criteria for the Break estimation is an improvement in the break decision achievable.

In Figure 2, the various Viterbi paths V1 to V3 are for different hidden Markov models are shown. Here will over time the relationship between pattern recognition and the presence of a break between different ones Patterns shown. First, the measurement signal, which for example a voice signal, a write signal, or a Signal that is emitted by signaling methods via a suitable signal transformation or several signal transformations transformed into a feature vector space. In a training phase of the pattern recognition process become typical models for background noise and also estimated for the useful signal that followed to be used in the recognition process. For the method according to the invention can be used for training, for example realized with the method of the hidden Markov models become. However, the pause detection method can be equally with other pattern recognition methods, such as e.g. dynamic programming, or neural networks carry out. If with the method according to the invention Hidden Markov models can be applied, e.g. for example the distribution functions of the feature vectors be estimated for each recognition unit. With recognition units are in this context in the automatic Speech recognition Speech sounds (phonemes) meant. The invention The procedure was, for example, automatic Realized speech recognition, but it is conceivable that it can be used for any kind of pattern recognition can. Just make sure that signal patterns are provided and that there are pause conditions in which the interference signals can be determined in order to to train the hidden Markov models for break states. Some such examples of other pattern recognition methods, are, for example, the patterns used when signing a document in the form of pressure or time-dependent Write signals occur, or signal sequences that occur with automatic telecommunications signaling method applied become.

When carrying out the method according to the invention, in the recognition phase, for example, a continuous pattern comparison in every analysis interval or time slice the probability of generation for each recognition unit to calculate. An easy solution is to evaluate this Probabilities. Is the probability of pause, So for the hidden Markov model for break or its Correspondence highest, so the relevant analysis interval to re-estimate the distribution functions, or can be used to filter out noise suppression.

The method according to the invention becomes even more robust if the result of a pattern recognizer is taken into account as an additional source of knowledge. Assuming, for example, that the pattern recognizer is able to recognize every possible useful signal, the method according to the invention can take advantage of this and define all other analysis intervals, which are not classified as useful signals, as pauses. Such a time period is designated T _{p in} FIG. If there is no requirement for real-time processing with respect to the method, as is the case, for example, in simulations, the method according to the invention can hereby already be considered sufficient for pattern recognition. In practice, real-time criteria are to be used in the applications mentioned and the earliest possible assignment to the useful or noise signal. The method must therefore be integrated into the recognition process itself, for example. The recognition method is thus expanded in accordance with the invention in such a way that after each analysis step, for example, it is evaluated which of the patterns, for example words, composed of the recognition units is the most likely. In addition, the probability that it contains a signal pause is calculated over a larger analysis interval, for example. For example, the analysis interval is dimensioned such that it is in any case longer than short pauses, for example plosive pauses, in the useful signal. This probability is then compared with that of the most probable pattern, and they are related to an equally long time interval. The result of this comparison can already be used as a decision.

For example, even higher demands are placed on speech recognition systems. With them it must be avoided that the recognizer switches off prematurely and, as a result, outputs an incorrect word. In Figure 1, the recognizer is designated by Klass. These cases occur particularly with transient noise. For example, this can be prevented by an additional condition. For example, a signal pause is only recognized as the end of a word if, in addition to the criterion described above, the most likely word has always been the most likely word over a certain period of time. This time period is designated T _{ST in} FIG. The combination of these two criteria described provides a high level of reliability in pause recognition, which is important for the reliable functioning of a speech recognizer.

• Eigenschaften des Signals im Zeitbereich, wie z.B. Nulldurchgangsrate und Pegel, sowie
• im Spektralbereich, z.B. die Leistung und das Korrelationsmaß einschließlich des logarithmischen und/oder Merkmalsbereiches.
• Zusätzlich wird durch das erfindungsgemäße Verfahren die Pause detektiert, indem eine Rückführung von der Erkennungsstufe zur Merkmalsextraktionstufe realisiert wird.

The basic idea in a pattern recognition system is to use the knowledge sources available at different levels in signal processing stages to detect a pause. These extend to, for example

• Properties of the signal in the time domain, such as zero crossing rate and level, as well
• in the spectral range, for example the power and the correlation measure including the logarithmic and / or feature range.
• In addition, the pause is detected by the method according to the invention by realizing a return from the recognition stage to the feature extraction stage.

As a result, the existing in the different time slices Information about the presence of a pause in the classifier Class of the feature extraction level Merk supplied. For example, a dynamic one takes place during the recognition Pattern comparison, in which on the basis of the feature vectors in one Analysis window or a time slice an assignment to the pre-trained models is accomplished. A global one Search strategy, e.g. realized by the Viterbi algorithm, finds the most likely sequence of pre-trained Model states representing the incoming sequence of feature vectors reproduces [6].

In every time window, the classifier can be used for the classifier Information about pause / non-pause can be tapped and be fed to a pause detector in another stage. This is the case, for example, in the method according to the invention realizes that in the classifier a special hidden Markov model for pause with the incoming feature vectors is compared if a higher probability for Pause occurs than for other patterns, so it becomes a pause information for example to the feature extraction level Merk passed on and leads there to the decision that currently there is a pause. That means with this pause information can also be an existing one in the extraction stage Pause detector can be controlled to set pause. This pause decision can be probability-weighted, for example and is based on a decision that other sources of knowledge within the inventive method considered. Such other sources of knowledge are for example statistics of the measurement signal and phoneme context the Viterbi process. Because of the sequential structure a recognizer must, for example, when the Information on a pause detection level for the suppression of noise, e.g. the delay by an analysis window be taken into account. If you link the pause decision the acoustic phonetic modeling level speech recognition with common criteria for pause estimation, so is an improvement in the break decision achievable. Take the frame-by-frame detection, for example the breaks completely, so there is another source of knowledge in the detection system for the pause estimation.

1. Berücksichtigung der Signalcharakteristika im Zeitbereich (z.B. Nulldurchgangssrate, Pegel);

2. Zusätzliche Berücksichtigung der Eigenschaften im Spektralbereich (z.B. Leistung, Korrelationsmaß) einschließlich des logarithmischen und/oder Merkmalsbereiches;

3. Zusätzliche Berücksichtigung des frameweisen Mustervergleichs mit vortrainierten-Pausenmodellen;

4. Zusätzliche Berücksichtigung der Rückführung der Entscheidung des in der globalen Erkennung integrierten Pausedetektors.

For example, various coherent and also related patterns can be detected as a whole and conclusions can be drawn from them about pauses present in the measurement signal. For example, such a global pause detector can provide its information about the entire pattern to be recognized or the pattern sequence. In the case of speech recognition, such a pattern sequence would be a word to be recognized, for example. All areas except this pattern sequence can thus be recognized as a break, for example. This has the advantage that even current disturbances are included in the pause detection. The method according to the invention thus also works at very high interference levels, and is therefore more robust. Due to the principle, a larger time delay must be taken into account until a decision is made. This global pause detection stage is therefore particularly useful in connection with intermediate signal storage. It is particularly suitable for the processing of the measurement signal and can be used in particular to identify the breaks between individual words or to recognize pattern sequences. In summary, a pattern recognition and pause recognition system according to the invention can be described in the following stages.

1. Consideration of the signal characteristics in the time domain (eg zero crossing rate, level);

2. Additional consideration of the properties in the spectral range (eg power, correlation measure) including the logarithmic and / or feature range;

3. Additional consideration of the frame-wise pattern comparison with pre-trained break models;

4. Additional consideration of the return of the decision of the pause detector integrated in the global detection.

For example, one embodiment of the method according to the invention is described by the pseudocode shown in Table 1.

For example, the inventive method is in one Main program that is limited by main and end. This main program essentially contains a do-loop as a time loop. With a signal_analysis procedure a transformation of the measurement signal into a feature area carried out. For example, a special time slice of the measurement signal is analyzed and feature vectors from this time slice created. The created feature vectors are then analyzed in a subroutine calculate_word_wk. For example, the for each reference word Probability, e.g. with hidden Markov models and with Calculated using Viterbi decoding. For example the association probability that all previous Feature vectors were emitted, calculated. In a Another subroutine calculate_pause_wk becomes the probability calculated for pause for the last P time steps.

Here, too, the association probability is calculated that the last P feature vectors from the model for pause were issued. In another subroutine pause detector pause information is generated if the probability for pause is higher than for the best word, otherwise the pause information is not generated. For example here a standardization of the probability to be taken into account performed for the same period of time P. In a another query if (pause && word_stabil> x) break, will the procedure was terminated when paused by pause detector was recognized and the best word at least has been stable for x periods (word_stable). With the subroutine output the recognized pattern sequence, in speech recognition a word, spent.

literature

[1] Rabiner, L.R. and M. Sambur (1975). An algorithm for determining the endpoints of isolated utterances. The Bell System Technical Journal, 54 (2): 297-315

[2] Hansen, J.H. (1991). Speech enhancement employing boundary detection and morphological based spectral constraints. IEEE International Conference on Acoustics, Speech and Signal Processing, pp 901-904, Toronto. ICASSP.

[3] Katterfeldt, H. Language determination with polynomial classifiers. Proceedings pattern recognition 7, DAGM symposium, Erlangen S 180 - 184.

[4] Boll, S. (1979). Suppression of acoustic noise in speech using spectral subtraction. IEEE Transactions on Acoustics, Speech and Signal Processing, 31 (3): 678-684

[5] Widrow, B., J.Glover, J.McCool, J.Kaunitz (1975). Adaptive noise canceling: Principles and applications. Proceedings of the IEEE, 63 (12): 1692-1716.

[6] Rabiner, L.R. and B.H. Juang (1986). An introduction to hidden markov models. IEEE Transactions on Acoustics, Speech and Signal Processing, (1): 4-16.

Claims (11)

1. Method for identifying a signal pause between two patterns which are present in a measurement signal that varies with time and are identified with the aid of hidden Markov models, having the following features:

   a) in a first signal processing stage (Merk), feature vectors (m) are formed periodically for pattern identification, which feature vectors (m) describe the signal waveform of the measurement signal (Spr) within a time slice, and a pause detector contained therein does not detect any speech pause in a first time slice on the basis of features of a first feature vector which are present,

   b) in a second signal processing stage (Klass) of the method, and in a second time slice which follows the first time slice, the first feature vector is compared with at least two hidden Markov models (HMM), at least one of which has been trained for a pattern which is to be identified and a further one of which has been trained for a pattern which is characteristic of a pause,

   c) if the comparison of the first feature vector (m) with the hidden Markov models (HMM) results in a greater probability of the presence of a pause, then the information about the presence of a pause, the pause information (Pa), is passed to the pause detector in the first signal processing stage, and the measurement signal is dealt with as a signal pause there, at least in the second time slice.
2. Method according to Claim 1, in which a defined sequence of patterns, a pattern sequence, is to be identified and in which the pause information is passed on after identifying the pattern sequence over a number of time slices, so that the measurement signal is dealt with as a signal pause, and not as a pattern to be identified, in the first signal processing stage, at least in the time slice following the pattern sequence.
3. Method according to Claim 2, in which feature vectors are buffer-stored until a pattern sequence has been identified and in which the pause information is passed on after identifying the pattern sequence, so that the measurement signal is dealt with as a signal pause, and not as a pattern to be identified, in the first signal processing stage, at least in the time slice before the pattern sequence.
4. Method according to one of the preceding claims, in which characteristics of the measurement signal are evaluated in the time domain in the first signal processing stage for pause identification.
5. Method according to one of the preceding claims, in which characteristics of the measurement signal are evaluated in the frequency domain in the first signal processing stage for pause identification.
6. Method according to one of the preceding claims, in which context-modelled hidden Markov models are used.
7. Method according to one of the preceding claims, in which the measurement signal represents spoken voice information.
8. Method according to Claim 7, in which disturbances in the feature extraction stage of a voice processing system are suppressed.
9. Method according to one of Claims 7 or 8, in which channel adaptation of a voice channel is carried out.
10. Method according to one of Claims 1 to 6, in which the measurement signal represents writing movements on a surface.
11. Method according to one of Claims 1 to 6, in which the measurement signal represents signal sequences of an information signalling method.

Images