WO2002045002A1

WO2002045002A1 - Method and system for reducing the error rate in pattern recognitions

Info

Publication number: WO2002045002A1
Application number: PCT/EP2001/012454
Authority: WO
Inventors: Alfred Hauenstein
Original assignee: Siemens Aktiengesellschaft
Priority date: 2000-11-28
Filing date: 2001-10-26
Publication date: 2002-06-06

Abstract

The invention relates to a method and system for reducing the error rate in pattern recognitions, in particular, voice recognitions, handwriting recognitions, gesture recognitions and biometric recognitions. According to the invention, patterns with code characters are input by using a combination of different input devices, which are independent of one another, into data processing devices and/or communications devices.

Description

description

Method and system for reducing the error rate in pattern recognition

The invention relates to a method and a system for reducing the error rate in pattern recognition, such as, for example, voice recognition, font recognition, gesture recognition and biometric recognition, according to the preamble of patent claim 1 and the preamble of patent claim 8.

Pattern recognition is possible in a variety of ways when communicating with data processing devices and / or communication devices. For example, communication with a computer by voice recognition in the form of voice input or character recognition by means of input using a keyboard or a mouse is possible. A type recognition method is also known, in which handwriting is entered, for example by means of a pen, over a correspondingly pressure-sensitive area and recognized as characters.

Voice input makes it easier for the user to use computers when communicating with the computer and often increases the speed at which data or commands can be input. However, the method of voice input as a pattern recognition-based input method has a certain error rate, which is the case, for example, with a dictation system , as used in communication with PCs, is approx. 1 - 5%. Such dictation systems therefore often use a list of word alternatives in the event of an incorrectly recognized word. The incorrectly recognized word is first selected by the user using the mouse or a voice input, whereupon the PC then displays a list of alternative words from which the user uses mouse or voice input to select the correct word, ie the result. Such a subsequent correction of the words falling under the error rate has the disadvantage that the entire speech input Gabeverfahren is slow by this subsequent correction and it requires increased attention by the user to find the incorrect word and correct the incorrect word. Furthermore, the scope of the alternative words provided is only limited, which is why it is not always ensured that the correct word result is available in the list of alternative words. Then there is the possibility that the user enters the missing word himself, but this causes a further slowdown of the entire speech input process.

Such speech recognition methods also have the disadvantage that, in particular in the case of mobile end devices, such as mobile telephones, electronic organizers, PDAs and webpads or also mobile data acquisition terminals, often the display device in the form of a screen for the often complex representation of word results from speech recognition and its alternatives is not sufficient due to their size.

Another major disadvantage of such speech recognition or writing recognition methods is that despite constant further developments and optimizations of these methods, the accuracy of these pattern recognition-based input methods has limits due to the error rates mentioned above.

It is therefore the object of the present invention to provide a method and a system for reducing the error rate in pattern recognition, which not only enables increased accuracy in the recognition of the patterns, but also faster and simpler recognition of the patterns.

This object is achieved in terms of method by the features of claim 1 and for a system according to the features of claim 7. An essential point of the invention is that the patterns to be recognized are input simultaneously by means of at least two different mutually independent input devices, such as, for example, a voice input device and a font input device, which are patterns with the same content or at least with the same key characters. The formula then calculates the minimum error rate

arg opt = arg (min (P _g Z, "•)) (1)

instead of where

n error _ ^" * ~ ι p error, ~ \ ^r total, '~ yyy,' \ l,

with PX ° ^r = error probability for a calculation result i assigned to a key character when input by means of input device j, Cj = weight factor (lb, 2b, 3b) of input by means of input device j.

Such a combination of several different input methods makes it possible to use the above-mentioned formula (2) to form a sum of the error probabilities for the calculation result i from the error probabilities of the individual input devices with a different weighting Cj and from this the key character with the minimum accumulated error rate to be calculated using the formula (1). By summing up the error probabilities of the individual methods assigned to the input devices and a minimum cumulative error rate calculated therefrom, it is possible to reduce the error rate of the overall system consisting of the various input devices and thus to achieve an increased accuracy of the system. This is particularly useful if, for temporary high accuracy requirements (e.g. numerical a) fast, inaccurate procedure would be supplemented by a more precise but slow one.

In particular, compact terminals with small display devices (displays) often do not have the option of subsequently correcting pattern recognition by means of voice input using a keyboard or a mouse, since such terminals do not have such input devices. In this case, the input of patterns is possible quickly and easily by combining, for example, a biometric input method and a voice input to reduce the error rate.

The weight factors of the individual input methods can advantageously be predetermined, so that the user does not have to make any further settings before using the pattern recognition method.

Alternatively, in a preferred embodiment, the user can use the weighting factors of the individual input methods in

Depending on the desired accuracy of the method or the assigned input device, set individually beforehand (“enrollment”) or train depending on the accuracy that occurs with it.

Through the simultaneous and parallel use of a plurality of input devices which are independent of one another for inputting patterns, a comparison of the error rates of the different input devices for the pattern with the same key characters can take place and a minimum cumulative error rate can thus be calculated therefrom. This results in an increase in the accuracy of the entire system consisting of all input devices.

The results of the calculated error rates for each individual input device can advantageously be displayed separately on a display device or by means of an output device. device, such as a voice output device, are output to the user for control.

Further embodiments result from the subclaims. An embodiment of the invention is explained in more detail below with reference to the figure, which shows a schematic representation of an embodiment of the method according to the invention.

As can be seen in the figure, several input methods and associated input devices 1a, 2a, 3a are used to carry out a pattern recognition with a reduced error rate. A voice input with a microphone la and a font input with a stylus 2a and a pressure-sensitive surface can thus be combined with one another. In addition, the pattern can also be entered using a keyboard 3a.

An electronic telephone book is stored in a mobile, compact terminal, such as a cell phone 5, from which the selected telephone number can be automatically selected by selecting a telephone number. The terminal is also used to enter appointments in the electronic telephone book. Such a terminal can also be, for example, a conventional telephone, a cordless telephone, an electronic organizer, a PDA, a webpad or a mobile data acquisition terminal.

A selection of a telephone number from this telephone book via a name assigned to the telephone number takes place by means of a voice input and a written input. For example, the user speaks the words "Please choose Möller" or "New appointment tomorrow 5:00 p.m. with Müller". These sentences contain certain key characters in the form of keywords, namely the word "dial", which stands for an action to be carried out by the terminal and the word "Möller", which stands for the destination, ie the person to be called. If such a selection from the telephone book were to take place exclusively by means of a voice input via a microphone 1 a, then, in particular in the case of very extensive electronic telephone books, a confusion of similar terms such as “Möller” and “Müller” could take place due to the expected error rate. For this reason, the keywords "select" and "Möller" are written down with the pen 2a at the same time as the voice is input and recognized by a subsequent automatic font recognition. Different patterns with the same key words, namely the interchangeable words, are thus input via two different input devices. Alternatively, the entire sentence can be entered using language and writing. After the input has taken place, a weighting 1b, 2b, 3b of the results of the different pattern recognition methods that are assigned to the input devices j takes place, the different results i being assigned to the different keywords, that is to say the actions and / or goals.

The cumulative error rate is calculated using n input devices j from the following formula:

p - error X ¹ s ~ ι total error, i ~ 2-ι 1 JJ (2)

/ = 1

If only two input devices j = 1 and j = 2 are used for pattern recognition, the following formula results from the formula (2):

p - error s ~ ι r> error, * ~ t rfβrror total, i ^"L l ^X X, i + 2 ^X * 2ji (3).

Here p ^{* r} ° ^r ^ _ _{e is the} probability of error for the result i in the case of an input via the input device j, and C _{j is} the weight factor for an input with the input device j. The following also applies:

(4)

There is therefore a calculation of the probability of error for the result i for the input via each input device j. This probability of error for a specific input device is then combined with a weighting factor Ci or C _{2 in} order to subsequently calculate a sum of the error probabilities, also called classifier 4. It should be taken into account that the sum of all weight factors C _j results in exactly 1. Depending on the method associated with the respective input device j, a different weighting can be carried out by individually setting the weighting factors C _x and C ₂ in advance or by the user for each pattern recognition to be carried out.

With the formula

arg opt = arg (min (P _{mt t} )) ₍₁₎

the result is now calculated and selected with the minimum cumulative error rate. This results in an increase in the accuracy of the pattern recognition by combining different input devices.

In an alternative embodiment, the result with the minimum cumulative error rate according to (1) is not chosen, but the result with the minimum error rate for a given method j. This results in

arg opt = arg (min (P ™)) _{with fixed j (lb} , Method (lb) is expediently used if, as described above, an exact method j, which is slower than the standard method, is only used temporarily for keywords.

Possible input devices can be voice input devices, keyboard input devices, mouse input devices, handwriting input devices, gesture input devices and biometric input devices, such as for example for inputting brain waves, fingerprints, etc.

Claims

claims

1. A method for reducing the error rate in pattern recognition, in particular voice recognition, speech script recognition, gesture recognition and biometric recognition, in which patterns with key characters are entered into data processing devices and / or communication devices (5) by means of input devices (la, 2a, 3a), characterized in that different patterns having the same key character by means of at least two different, mutually independent input means (la, 2a, 3a) are entered simultaneously and the key character with the minimum error rate by means of the formula arg opt = arg (min (P _g _™,)) is calculated, where: nn error _ V " ¹ - < _v p error ^r total, ^_ YYY, <

with P _j ^β ° ^r = error probability for a calculation result i assigned to a key character when input by means of the input device j,

Cj = weight factor (lb, 2b, 3b) of the input using the input device j.

2. The method according to claim 1, which also means that, alternatively, the result with the minimum cumulative error rate is not chosen, but the result with the minimum error rate is to be used for a predetermined method j, where:

arg opt = arg (min (Pjy ^r )) with fixed j.

3. The method according to claim 1, characterized in that: _cj = l •

4. The method of claim 1 or 3, d a d u r c h g e k e n n z e i c h n e t that the weight factors (lb, 2b, 3b) are predetermined.

5. The method according to any one of the preceding claims, in particular according to claim 1 or 3, characterized in that the weight factors (lb, 2b, 3b) are determined by a user who enters the pattern, which can be done in advance or a learning step is carried out in which the accuracy of the procedures for the specific user is to be determined.

6. The method according to any one of the preceding claims, characterized in that the calculation results of the error rates as a function of an input method assigned to the respective input device (1a, 2a, 3a) are displayed on a display device of the data processing device and / or communication device (5) at the same time or with a time shift become.

7. The method according to any one of the preceding claims, that the calculation results of the error rates are output as a function of an input method assigned to the respective input device (la, 2a, 3a) with a speech output device.

8. System for reducing the error rate in pattern recognition, in particular voice recognition, voice script recognition, gesture recognition and biometric recognition, with input devices (la, 2a, 3a) in data processing devices and / or communication devices (5) for entering patterns with key characters, g e k e n n z e i c h n e t d u r c h

- at least two different, independent from each other Input devices (la, 2a, 3a) for the simultaneous input of different patterns with the same key characters, and - a computing device for calculating the key characters with minimal error rates using the formula arg opt = arg (min (P ° ™ _t ,,)) <where _p enor _ _rp error with PT? ^* = Error probability ^r total, 1 - ° j ^{X r} Jl ^J ''

for a calculation result i assigned to a key character when input is made using input device j, Cj = weight factor (lb, 2b, 3b) is input using input device j.

9. System according to claim 8, characterized in that the input devices (la, 2a, 3a) devices for voice input and / or devices for entering printed characters and / or devices for entering handwriting characters and / or devices for entering gestures and / or Includes facilities for entering biometric data.