Data processing: speech signal processing – linguistics – language – Speech signal processing – Recognition
Reexamination Certificate
2000-02-11
2003-01-07
Dorvil, Richemond (Department: 2654)
Data processing: speech signal processing, linguistics, language
Speech signal processing
Recognition
C704S207000, C704S209000
Reexamination Certificate
active
06505154
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The invention relates to a method for comparing at least one acoustic input signal fed into an input device with at least one further acoustic reference signal stored in a memory, in accordance with which
a harmonic analysis of the input signal is carried out in order to produce a Fourier spectrum in a frequency analyzer connected to the input device and the memory, in accordance with which, furthermore,
the Fourier spectrum is used to define an n-dimensional input signal vector with n input (signal) coordinates, and in accordance with which
the input signal vector is checked for correspondence with at least one reference signal vector, which is defined in the same way and thereafter stored in the memory.
2. The Prior Art
In the case of a method of the configuration described at the beginning, a circuit arrangement for voice recognition is described which is equipped with an evaluation circuit for determining spectral feature vectors. In detail, an evaluation is made in this case of a time frame of a digital voice signal by means of a spectral analysis. The feature vectors obtained are compared with reference feature vectors. Before the comparison, the evaluation circuit performs a recursive high-pass filtering of the spectral feature vectors. The aim hereby is to enable recognition independent of speaker and to reduce the influence of interference on the result of recognition (compare DE 41 11 995 A1).
Moreover, there is a description in the prior art of a voice analysis system in which time segments of a voice signal are selected and a series of spectrum components of each voice segment are determined. These spectrum components form a discrete Fourier transform of samples of the voice signal. It is now possible to derive the position of significant peaks in each time segment from the series of relevant spectrum components.
Moreover, the procedure here is to select a value for the pitch. Intervals are defined around this initial value and a number of sequential integral multiples thereof. These intervals are regarded as openings in a mask, specifically in the sense that a frequency value which coincides with an opening is passed by the mask. Thus, in this sense the mask acts as a type of filter for frequency values. The aim in this way is for the described voice analysis system to be insensitive to interference signals and to require fewer calculations (compare DE 29 49 582 A1).
In a method known from practice, the general procedure is for an acoustic input signal, for example a tone, a tone sequence or else voice, to be compared with a stored acoustic reference signal and, in the event of correspondence, for a control device connected to the input device to be activated. This can relate to a door in connection with access control. Driving a machine is also conceivable. Corresponding attempts are also being made in the automotive sector in order, for example; to control individual functions by means of voice, for example to set flashers, switch lights on and off, etc.
The methods and devices previously considered are generally of very complex design and configuration, because in the end a complete analysis of the input signal is performed. The invention starts from this point.
The technical problem on which the invention is based is to develop such a method so as to achieve recognition with a simple means and quickly as well as with high accuracy. Moreover, an appropriately adapted device is to be specified.
In order to achieve this object, the invention proposes in the case of a method of the generic type and in accordance with a first alternative that the respective reference signal vector is flanked by a safety space or definition space, preferably adapted to the number of the reference signal vectors to be stored, and by an identity space, in which case the input signal vector is usually checked with the reference signal vector for correspondence of respectively corresponding coordinates within prescribed tolerances, and in which case, furthermore, the input signal vector is identified as being equal to the reference signal vector at least whenever it (the input signal vector) is situated inside the identity space and in which case respective reference signal vectors including the definition space have no overlap, that is to say the definition or safety spaces of all the reference signal vectors do not overlap one another. As a rule, recourse is made to at least two reference signal vectors. Otherwise, these reference signal vectors are, as mentioned, configured such that the intersections of respective reference signal vectors including the definition space is an empty set.
According to a preferred refinement, furthermore, the input signal vector to be compared with the reference signal vector is not detected outside the respective identity space. Alternatively, it is also possible to proceed so as additionally to define an option space which surrounds the identity space with a gray zone and serves, so to speak, as a collecting net for input signal vectors which cannot be assigned to an identity space of a reference signal vector.
The result in any case is that it is possible to compare the input signal vector and reference signal vector for correspondence in a way which is more accurate and simpler. The reason is that the safety space or definition space can be of variable configuration just like the identity space. Usually, the procedure here is that the reference signal vectors including the safety space exhibit no overlap. Since the configuration of the identity space is smaller than or equal to the safety space, input signal vectors can be interpreted as being equal or identical to the corresponding reference signal vector at least within the identity space, even if large tonal or voice deviations are to be noted in part.
Additionally, it is possible to add the described option space, which defines a type of gray zone around the identity space. This gray zone can be dimensioned such that an input signal vector can be assigned two (or more) reference signal vectors within this region—and only this region. This is generally not desirable, but offers advantages under some circumstances. In any case, owing to this property this gray zone acts, so to say, as a collecting net for input signal vectors which cannot be assigned to an identity space of a reference signal vector. This may be ascribed, inter alia, to the fact that input signal vectors are not detected as a rule outside the respective identity space.
According to a further proposal of the invention, which is of independent importance, in the case of a method of the generic type it is proposed that with the aid of an adjustable n-fold filter respectively preselectable characteristics of the Fourier spectrum are evaluated in the frequency analyzer and converted into the n input (signal) coordinates of the input signal vector. In this case, as well, respectively corresponding coordinates of the input signal vector and the reference signal vector are regularly checked here for correspondence within prescribed tolerances.
The prescribable characteristics can be the highest frequency, the maximum amplitude, the duration or relative gains of salient frequencies or the like of the input signal. Of course, a comparable statement also holds for the reference signal vector, which can likewise be defined via the abovementioned characteristics of highest frequency, maximum amplitude, duration and relative gains of salient frequencies.
Of course, it is also possible to use additional coordinates based on other characteristics. It is conceivable here to determine the number of individual peaks in the Fourier spectrum. It would also be possible to determine a coordinate as the sum of the individual amplitudes, and thus as the amplitude integral over the frequency.
In the final analysis, this depends on the number of stored reference signal vectors with which the input signal vector must be compared: the trend is that more coordinates of the individual vectors are required t
Bottenbruch Hermann
Mertens Michael
Collard & Roe P.C.
Dorvil Richemond
Primasoft GmbH
LandOfFree
Method and device for comparing acoustic input signals fed... 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 and device for comparing acoustic input signals fed..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and device for comparing acoustic input signals fed... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3034953