Data processing: speech signal processing – linguistics – language – Speech signal processing – Application
Reexamination Certificate
2000-01-18
2004-02-24
Dorvil, Richemond (Department: 2654)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Application
C704S231000
Reexamination Certificate
active
06697782
ABSTRACT:
The present invention relates to a method for recognizing speech commands, in which a group of command words selectable by speech commands are defined, a time window is defined, within which the recognition of the speech command is performed, and a first recognition stage is performed, in which the recognition result of the first recognition stage is selected.
The present invention also contemplates a speech recognition device in which a vocabulary of selectable command words is defined. The device includes means for measuring the time used for recognition and comparing it with a predetermined time window, and means for selecting a first recognition result.
The present invention also includes a wireless communication device to be controlled by speech comprising means for recognizing speech commands, in which a vocabulary of selectable command words is defined, the means for recognizing speech commands comprising means for measuring the time used for recognition and comparing it with a predetermined time window, and means for selecting a first recognition result.
For facilitating the use of wireless communication devices, so-called hands free devices have been developed, whereby the wireless communication device can be controlled by speech. Thus, speech can be used to control different functions of the wireless communication device, such as turning on/off, transmission/reception, control of sound volume, selection of telephone number, or answering a call, whereby particularly in the use in a vehicle, it is easier for the user to concentrate on the driving.
One drawback in a wireless communication device controlled by speech is that speech recognition is not fully faultless. In a car, the background noise caused by the environment has a high volume, thereby making it difficult to recognize speech. Due to the unreliability of the speech recognition, users of wireless communication devices have so far shown relatively little interest in the control by speech. The recognition capability of present speech recognizers is not particularly good, especially under difficult conditions, such as in a moving car, where the high volume of background noise hampers reliable recognition of words substantially. Incorrect recognition decisions cause most problems usually in the implementation of the user interface, because incorrect recognition decisions may start undesired functions, such as terminating a call during a call, which is naturally particularly disturbing to the user. One result of an incorrect recognition decision may be that a call is connected to an incorrect number. For this reason, the user interface is designed in such way that the user usually is asked to repeat a command if the speech recognizer does not have sufficient certainty of a word uttered by the user.
Almost all speech recognition devices are based on the functional principle that a word uttered by the user is compared, by an usually rather complicated method, with a group of reference words previously stored in the memory of the speech recognition device. Speech recognition devices usually calculate a figure for each reference word to describe how much the word uttered by the user resembles the reference word. The recognition decision is finally made on the basis of these figures so that the decision is to select the reference word which the uttered word resembles most. The best known methods in the comparison between the uttered word and the reference words are dynamic time warping (DTW) and the statistical hidden Markov model (HMM).
In both the DTW and the HMM methods, an unknown speech pattern is compared with known reference patterns. In dynamic time warping, the speech pattern is divided into several frames, and the local distance between the speech pattern included in each frame and the corresponding speech segment of the reference pattern is calculated. This distance is calculated by comparing the speech segment and the corresponding speech segment of the reference pattern with each other, and it is thus a kind of numerical value for the differences found in the comparison. For speech segments close to each other, a smaller distance is usually obtained than for speech segments further from each other. On the basis of local distances obtained this way, a minimum path between the beginning and end points of the word are sought by using a DTW algorithm. Thus, by dynamic time warping, a distance is obtained between the uttered word and the reference word. In the HMM method, speech patterns are produced, and this stage of speech pattern generating is modelled with a state change model according to the Markov method. The state change model in question is thus the HMM. In this case, speech recognition on received speech patterns is performed by defining a observation probability on the speech patterns according to the hidden Markov model. In speech recognition by using the HMM method, an HMM model is first formed for each word to be recognized, i.e. for each reference word. These HMM models are stored in the memory of the speech recognition device. When the speech recognition device receives the speech pattern, a observation probability is calculated for each HMM model in the memory, and as the recognition result, a counterpart word is obtained for the HMM model with the greatest observation probability. Thus for each reference word the probability is calculated that it is the word uttered by the user. The above-mentioned greatest observation probability describes the resemblance of the received speech pattern and the closest HMM model, i.e. the closest reference speech pattern.
Thus, in present systems the speech recognition device calculates a certain figure for the reference words on the basis of the word uttered by the user. In the DTW method, the figure is the distance between the words, and in the HMM method, the figure is the probability for the equality of the uttered word and the HMM model. When the HMM method is used, the speech recognition devices are usually set a certain threshold probability which the most probable reference word must achieve to make the recognition decision. Another factor affecting the recognition decision can be e.g. the difference between the probabilities of the most probable and the second probable word, which must be sufficiently great to make the recognition decision. Thus, it is possible that when the background noise has a high volume, on the basis of a command uttered by the user, the reference word in the memory, e.g. the reference word “yes”, obtains at each attempt the greatest probability in relation to the other reference words, e.g. the probability 0.8. If the threshold probability is for example 0.9, the recognition is not accepted and the user may have to utter the command several times until the recognition probability threshold is exceeded and the speech recognition device accepts the command, even though the probability may have been very close to the acceptable value. This is very disturbing to the user.
Furthermore, the speech recognition is hampered by the fact that different users utter the same words in different ways, wherein the speech recognition device works better when used by one user than when used by another user. In practice, it is very difficult with the presently known techniques to adjust the certainty levels of speech recognition devices to consider all users. When adjusting the required certainty level e.g. for the word “yes” in speech recognition devices of prior art, the required threshold is typically set according to so-called worst speakers. Thus, the problem emerges that words close to the word “yes” also become incorrectly accepted. The problem is aggravated by the fact that in some situations, mere background noise may be recognized as command words. In speech recognition devices of prior art, the aim is to find a suitable balance in which a certain part of the users have great problems in having their words accepted and the number of incorrectly accepted words is sufficiently small. If the speech recognition device is adjusted in a way
Iso-Sipilä Juha
Laurila Kari
Dorvil Richemond
Harper V. Paul
Nokia Mobile Phones Ltd.
Perman & Green LLP
LandOfFree
Method in the recognition of speech and a wireless... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method in the recognition of speech and a wireless..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method in the recognition of speech and a wireless... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3344269