JPH0962292A - Reject method in time series pattern recognition processing and time series pattern recognition device mounting it - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the processing amount by finding the optimum path obtained from a state having the maximum probability among states terminated with an HMM network and a state having the maximum probability among all states, and judging as rejection when the maximum probability state is not on the optimum path. SOLUTION: An optimum path calculating means 103 finds the probability which is obtained from a state becoming a termination among probabilities obtained from a reference means 101 and is maximum among the probabilities, and finds the optimum path having the maximum probability. A maximum likelihood state calculating means 104 finds the probability which is maximum among all probabilities obtained from a reference means and finds a state having the maximum probability. A judging means 105 judges whether the state obtained from the maximum likelihood state calculating means 104 exists on the optimum path obtaining from the optimum path calculating means 103 or not, and if it exists, outputs a pattern represented by the optimum path obtaining from the calculating means 103 as the recognition result, and if not, outputs rejection as the rejection result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a time-series pattern recognition device, and more particularly to a reject method capable of detecting a pattern which is not a recognition target and a time-series pattern recognition having the pattern rejected. It relates to the device.

The reject method or rejection is a function in which the recognition processing outputs "input of non-recognition target data" or "(no high accuracy) recognition result" as a recognition result.

[0003]

2. Description of the Related Art Generally, a recognition process outputs, as a recognition result, reference data that is most similar to input data among reference data given in advance as recognition targets.

Voice, music, written characters, sign language, gestures, moving images, etc. are converted into a time series pattern, which is, for example, a statistical method HMM (Hidden Markov Model).
Can be recognized by using. Rejection is an essential function when incorporating these recognition technologies into various application systems and putting them into practical use. This is because, for example, in the case of a voice recognition application system, a system that incorporates rejection and has a mechanism that encourages the user to re-voice rather than cause the system to malfunction as a result of erroneously recognizing the user's utterance content. Will be easier to use.

As a method of realizing rejection, it is conceivable to set an absolute threshold value in advance for the probability value and perform rejection. However, even if the recognition result is the same, if the input data generation environment is different, the probability value itself obtained by the HMM also changes, so it is difficult to set an absolute threshold value.

On the other hand, there is rejection that sets a relative threshold value for the probability value. Examples of speech recognition include "Watanabe et al .: Rejection of unknown word utterance by likelihood correction using syllable recognition, IEICE Transactions, Vol. J75-D-II, No. 12, pp. 2.
002-2009 ". This method collates a first HMM network (corresponding to a standard pattern) that represents a word or a sentence to be recognized regarding an application. Separately, matching (also called syllable typewriter) with a second HMM network representing a syllable sequence and an unconstrained syllable string is also performed, where the probability value of the second HMM network matching result is relatively referred to. Achieve rejection.

[0007]

In the conventional reject method in the time-series pattern recognition processing as described above, the processing amount and the storage capacity necessary for the processing in the second HMM network matching processing increase. There were challenges. If the amount of calculation increases, there is a problem that the response speed becomes slow. Also, in an environment where computer resources are limited, it is necessary to save the storage capacity used as much as possible. For example, when an application such as voice recognition is considered as one of the input means of the mobile terminal, it is required that the functions of the same level reduce the calculation amount and the storage capacity as much as possible from the viewpoint of power consumption.

An object of the present invention is to recognize a time series pattern based on a probabilistic model including an automaton,
An object of the present invention is to provide a reject method that does not increase the processing amount and the storage capacity. Especially, H which is a kind of automaton
A reject method using MM is provided.

Another object of the present invention is an HM representing a recognition target.
Another object of the present invention is to provide a reject method that does not increase the processing amount and the storage capacity other than the matching process of the M network, and a time-series pattern recognition device that implements the reject method. In particular, the present invention provides a reject method for preventing a malfunction due to an erroneous input of a sound or voice input that is not a recognition target, or a voice in which a part of the voice is missing (unvoiced voice), and a voice recognition device equipped with the reject method.

[0010]

In order to achieve the above object, according to the present invention, when the probability calculation of each state on the HMM network with respect to a time series input pattern is completed, the HMM network that represents the recognition target terminates. The optimum path obtained from the state having the maximum probability among the possible states and the state having the maximum probability among all the states in the HMM network representing the recognition target are obtained, and the state having the maximum probability among all the states of the HMM network is Rejection method that detects pattern input that is not a recognition target by judging rejection when it is not on the optimal path and prevents malfunction,
And a time series pattern recognition device implementing the reject method.

In the concrete configuration of the present invention, in the time series pattern recognition processing based on the probabilistic model including the automaton and the like, when the probability calculation of each state on the probabilistic model for the time series input pattern is completed, the recognition target is recognized. The optimal path obtained from the state that has the maximum probability among the states that can be the terminal in the probabilistic model that is represented, and the state that has the maximum probability of all the states in the probabilistic model that represents the recognition target are obtained, and the maximum probability of all states is obtained. If the state is not on the optimum path, it is determined as rejection.

In another configuration of the present invention, in a time series pattern recognition process based on a probabilistic model including an automaton or the like, the recognition target is determined at the time when the probability calculation of each state on the probabilistic model for the time series input pattern is completed. The state having the maximum probability among the states that can be the terminal in the probabilistic model that is represented and the state that has the maximum probability of all the states in the probabilistic model that represents the recognition target are obtained, and the probability values of the states having the two maximum probabilities are compared. , If the difference between the probability values is larger than a predetermined threshold value, it is determined as rejection.

[0013] In the concrete configuration when using the HMM,
In the time-series pattern recognition process based on HMM, when the probability calculation of each state on the HMM network with respect to the time-series input pattern is completed, it is obtained from the state having the maximum probability among the states that can be terminated in the HMM network that represents the recognition target. The optimal path that is obtained and the state that has the maximum probability of all states in the HMM network that represents the recognition target,
Rejection is determined when the state having the maximum probability among all the states is not on the optimum path.

In another specific configuration when using the HMM, in the time series pattern recognition processing based on the HMM,
The HM representing the recognition target at the time when the probability calculation of each state on the HMM network for the time series input pattern is completed.
The state having the maximum probability among the states that can be the termination in the M network and the state having the maximum probability among all the states in the HMM network representing the recognition target are obtained, and the probability values of the states having the two maximum probabilities are compared. If the difference between the probability values is larger than a predetermined threshold value, it is determined as rejection.

The voice recognition device of the present invention represents a voice input means for capturing a voice as an electric signal, a voice analysis means for converting the features of the input voice into a time series pattern and outputting the time series pattern, and a voice pattern to be recognized. Storage means for holding the HMM network, and HMM stored in the storage means
HMM by matching the network and the time series pattern
Collation means for calculating the probability of each state on the network,
An optimal path calculating means for obtaining an optimal path obtained from a state having a maximum probability among states that can be terminated in the HMM network; a maximum likelihood calculating means for obtaining a state having a maximum probability of all states in the HMM network; It has a determining unit that determines whether or not the state having the maximum probability is on the optimal path, and outputs a recognition result of "not applicable" when the state is not on the optimal path.

The determining means may be a determining means for obtaining the optimum path probability, comparing it with the maximum probability, and outputting a recognition result "not applicable" when the comparison result is larger than a predetermined threshold value. .

[0017]

According to the present invention, the pattern not to be recognized can be detected only by the matching process of the HMM network representing the object to be recognized, so that the rejection can be realized with almost no increase in processing amount and storage capacity.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. A detailed description of pattern recognition by HMM is given below.
For example, "Seiichi Nakagawa: Speech recognition by probabilistic model, Institute of Electronics, Information and Communication Engineers, 1988" and the like are not described here in detail. The present invention does not limit the type of recognition unit model such as HMM.

FIG. 1 is a flow chart showing an embodiment of the processing procedure of the present invention. FIG. 2 is a diagram showing an example of the HMM network.

The processing procedure of the present invention will be described with reference to the flowchart of FIG. First, the probability on the HMM network is calculated for the input pattern (100
1). The input pattern is a time-series pattern. For example, when represented by a symbol string such as an alphabet, "aaaabcccc"
It becomes a time series pattern like "cdd". In the initial setting, a probability of 1 is given to the starting state (Ss in FIG. 2), and a probability of 0 is given to the other states. After that, for each unit time, the state (S1) having a probability transits to the state (S2) connected by an arc, and the probability that the input symbol at that time appears is calculated, and the new probability of the subsequent state S2 is calculated. And At the end,
Each state of the transition on the HMM network has a probability value. Usually at this final point HMM
A state that can be a terminal on the network (Sf1, S in FIG. 2)
f2, Sf3, ...) are compared only for the magnitude of the probability, and the state with the highest probability among them is sought.
A path on the MM network is selected as a recognition result candidate (1002). In the example of FIG. 2, it is “ABC”. The state obtained here is Sbest, and the selected path is Rbest. In the present invention, the state with the highest probability among all states on the HMM network is obtained (10
03). The state obtained here is Smax. Next, it is checked whether the state Smax is one of the states on the path Rbest (1004). Here, if Smax is one of the states on the path Rbest, the pattern ("AB
C ”) is output as the recognition result (1005), and if Smax is not in the state on the path Rbest,“ Not applicable ”is output as the recognition result (1006). The function of outputting“ Not applicable ”is rejection. is there.

Here, there may be a case where a plurality of paths or states having the same probability are obtained for Rbest and Smax, respectively. However, if there is a combination in which Smax is in a state on Rbest, the result is obtained. If it is not output, it may be determined as rejection if it is not output.

If the input pattern is very likely as the data to be recognized, the collation is successful and Sbe
st and Smax should match. However, in the case of an input that is not a recognition target, these do not always match. This means that the recognition result of a time-series pattern that is not a recognition target, which is input from a path that reaches the state Smax having the maximum probability, is forcibly obtained as an approximate solution in a given HMM network. . HMM
The larger the network, the easier it is to obtain an approximate solution that is close to the true solution due to the breadth of the representation of the network, and the easier it is to detect inputs that are not recognition targets. Taking the case of a voice recognition device as an example, if the network is sufficiently large, a result sufficiently similar to that in the case of recognizing syllables one by one can be obtained. Further, it is possible to reject short noises such as breath, cough, and throat clearing as well as unknown words based on the same principle as in the case of detection of an unrecognized input. Further, it is possible to detect an error in the voice section detection process for unvoiced voice and the like.

The processing amount of 1003 and 1004 in FIG. 1 is so small that it can be ignored in comparison with the probability calculation (1001) of the HMM network for the input pattern. Further, this method has a feature that it is not necessary to set a complicated threshold value.

FIG. 3 shows another embodiment of the processing procedure according to the present invention. The present embodiment is different from the previous embodiments in that a threshold value is set for the probability value and reject determination is performed.

The processing procedure of the present invention will be described with reference to the flowchart of FIG. First, the probability on the HMM network is calculated for the input pattern (200
1). Usually, the probability is compared only between states that can be terminations on the HMM network at this final point,
The state with the highest probability among them is sought, and the path on the HMM network that has reached that state is selected as a recognition result candidate (2002). The state obtained here is Sbest, and the selected path is Rbest. Up to this point,
This is the same as the first embodiment described using. Further, the probability at Sbest is held as Pbest (2002).
Next, the highest probability Pmax among all the states on the HMM network is calculated (2003). Next, the difference between Pmax and Pbest is divided by the length of the input pattern (the number of frames is N) to obtain a normalized value as a reference probability value (ΔP) (2
004). Next, ΔP and a preset threshold value (Pth) are compared and inspected (2005). Here, if ΔP is smaller than Pth, the pattern corresponding to Pbest is output as the recognition result (2006), and if ΔP is larger than Pth, “not applicable” is output as the recognition result (200
6).

FIG. 4 is a block diagram showing the structure of the pattern recognition apparatus according to the present invention.

In FIG. 4, the HMM network storage means 102 stores the HMM network expressing the recognition target. In the present invention, the HMM network storage means 1
The storage method of the HMM network in 02 is not limited. The collating means 101 uses the HMM as the input pattern.
The probability of each pattern being expressed in the network is calculated. The optimum path calculation unit 103 obtains the maximum probability obtained from the probable states of the probabilities obtained from the matching unit 101, and obtains the optimum path having the probability. Maximum likelihood state calculation means 104
Calculates the maximum probability among all the probabilities obtained from the matching means, and calculates the state having that probability. The determination means 105
It is determined whether or not the state calculated by the maximum likelihood state calculation unit 104 exists on the optimum path obtained by the optimum path calculation unit 103, and if it exists, the state is expressed by the optimum path obtained by the optimum path calculation unit 103. The pattern to be output is output as the recognition result, and if it does not exist, the rejection is output as the recognition result.

FIG. 5 is a block diagram showing an embodiment of the voice recognition apparatus according to the present invention.

In FIG. 5, the voice input means 206 comprises a microphone and an analog-digital converter, receives a voice waveform of air vibration as an input, and outputs a data string of amplitude values sampled at a constant sampling period. The voice analysis unit 207 analyzes the data sequence output from the voice input unit 206 by a constant time interval or a fixed number of units, and converts the features of the input data sequence into a feature vector sequence containing some parameters. Output. The feature vector sequence here corresponds to the symbol sequence in the description relating to FIG. The present invention does not limit the types of parameters of this analysis process. For example, the LPC cepstrum coefficient, short-term power value, etc. can be used, and details including other parameters can be found in "Furui: Acoustic and Speech Engineering, Modern Science Co., Ltd., 1992". Subsequently, the HMM network storage unit 202 stores the HMM network expressing the recognition target (sentence or word). In the present invention, HM
The method of storing the HMM network in the M network storage unit 202 is not limited. The matching means 201 calculates the probability that the feature vector sequence, which is an input pattern, will be an individual pattern expressed in the HMM network. The optimum path calculation means 203 is the matching means 2
Among the probabilities obtained from 01, the maximum probability obtained from the state that can be the terminal end and among them is obtained, and the optimal path having that probability is obtained. The maximum likelihood state calculation means 204
Find the maximum probability among all the probabilities obtained from the matching means,
Find the state with that probability. The determining unit 205 determines whether or not the state obtained by the maximum likelihood state calculating unit 204 exists on the optimum path obtained by the optimum path calculating unit 203, and if there is, obtains it by the optimum path calculating unit 203. The pattern expressed by the optimum path is output as the recognition result, and if it does not exist, the rejection is output as the recognition result. The judging means 205 may use the reject method shown in FIG. 1 and its description.

Further, in the judging means 205, as shown in FIG. 3, the probability of the state having the highest probability (Pbes) among the probabilities of the states which can be the terminal on the HMM network.
t) and the highest probability among all states on the HMM network may be compared and inspected, and if the difference is larger than a predetermined threshold value, “not applicable” may be output as the recognition result.

[0031]

According to the present invention, the effect that the rejection can be realized with almost no increase in the processing amount and the storage capacity can be obtained. For example, when applied to voice recognition, short noises such as breath, cough, and throat clearing can be rejected, and unknown words can also be rejected. Further, it is possible to obtain the effect of being able to detect a mistake in the voice section detection process for unvoiced speech. Therefore, in the application system in which the rejection method of the present invention is implemented, it is possible to prevent malfunction of the system due to erroneous input by the user or erroneous recognition of recognition processing.

The rejection according to the present invention is H
The larger the MM network is, the more arbitrary input patterns can be expressed, and the more effective the effect becomes.

[0033]

[Brief description of drawings]

FIG. 1 is a flowchart showing an example of a processing procedure of the present invention.

FIG. 2 is a block diagram showing an example of an HMM network according to the present invention.

FIG. 3 is a flowchart showing another example of the processing procedure of the present invention.

FIG. 4 is a block diagram showing an embodiment of a configuration of a time series pattern recognition device according to the present invention.

FIG. 5 is a block diagram showing an embodiment of the configuration of a voice recognition device according to the present invention.

[Explanation of symbols]

101 ... Collation means, 102 ... HMM network storage means, 103 ... Optimal path calculation means, 104 ... Maximum likelihood state calculation means, 105 ... Judgment means.

Claims (6)

[Claims] 1. A time series pattern recognition process based on a probabilistic model including an automaton, etc., when the probability calculation of each state on the probabilistic model for a time series input pattern is completed, the probabilistic model representing the recognition target can be terminated. The optimal path obtained from the state having the maximum probability among the states, and the state having the maximum probability among all the states is obtained by the stochastic model representing the recognition target, and the state having the maximum probability among all the states is on the optimal path. Rejection method characterized by determining rejection if not in a state. 2. A time series pattern recognition process based on a probabilistic model including an automaton and the like, at the time when the probability calculation of each state on the probabilistic model for a time series input pattern is completed, the probabilistic model representing the recognition target can be terminated. The state having the maximum probability among the states and the state having the maximum probability among all states are obtained by the probability model representing the recognition target, and the probability values of the states having the two maximum probabilities are compared. A reject method, characterized in that it is judged to be rejected when it is larger than a predetermined threshold value. 3. In the time-series pattern recognition processing based on HMM, at the time when the probability calculation of each state on the HMM network for the time-series input pattern is completed, the maximum probability among the states that can be terminated in the HMM network representing the recognition target. The optimal path obtained from the states having the maximum probability and the state having the maximum probability of all states in the HMM network that represents the recognition target, and rejected if the state having the maximum probability of all states is not on the optimal path. Rejection method, which is characterized by determining that it is an action. 4. In a time series pattern recognition process based on HMM, at the time when the probability calculation of each state on the HMM network with respect to the time series input pattern is completed, the maximum probability among states that can be terminated in the HMM network representing the recognition target. And a state having the maximum probability among all states in the HMM network representing the recognition target, the probability values of the states having the two maximum probabilities are compared, and the difference between the probability values is a predetermined threshold value. Rejection method characterized by determining rejection if larger. 5. A voice input means for taking a voice as an electric signal, a voice analysis means for converting the features of the input voice into a time series pattern and outputting the same, and a memory for holding an HMM network representing a voice pattern to be recognized. Means, collating means for collating the HMM network stored in the storage means with the time series pattern to calculate the probability of each state on the HMM network, and the maximum probability of states that can be terminated in the HMM network. Optimal path calculating means for obtaining an optimal path obtained from a state having
Maximum likelihood calculating means for obtaining a state having the maximum probability among all states in the network, and determining whether or not the state having the maximum probability is a state on the optimal path, and when it is not on the optimal path, Is a voice recognition device having a determination means for outputting a recognition result of "not applicable". 6. A voice input means for taking a voice as an electric signal, a voice analysis means for converting a feature of the input voice into a time series pattern and outputting the same, and a memory for holding an HMM network representing a voice pattern to be recognized. Means, collating means for collating the HMM network stored in the storage means with the time series pattern to calculate the probability of each state on the HMM network, and the maximum probability of states that can be terminated in the HMM network. Optimal path calculating means for obtaining an optimal path obtained from a state having
Maximum likelihood calculating means for obtaining a state having the maximum probability among all states in the network, and obtaining the optimum path probability and comparing it with the maximum probability. If the comparison result is larger than a predetermined threshold value, "not applicable". A voice recognition device having a determination means for outputting the recognition result of