Patent US4597274A successfully refreshed from Google Patents.
EP0240329A2
Expired / Lapsed
EP
⏱️ 0.0 years remaining
Noise compensation in speech recognition
Assignee: National Research Development Corp UK
•
Filing: 1987-04-01
•
Granted: —
•
Expires: 2007-04-01
Abstract
Abstract In speech recognition it is advantageous to take account of noise levels both in recognition and training. In both processes signals reaching a microphone 10 are digitised and passed through a filter bank 13 to be separated into frequency channels. In training a noise estimator 18 and a masker 15 are used with a processor 17 to prepare template signals which are held in a store 16. In each region the template signals are made equal to the highest noise level in that region during the whole training process if below this level. In recognition the input signals in each region are made equal to the highest noise level in that region if below this level. "Distance" measurements for each region on which recognition is based depend on the levels of the template and input signals unless either is represented by noise when the relative levels determine how the distance shall be calculated.
Inventors
- 👤 Nigel Charles Sedgwick
- 👤 John Nicholas Holmes
Bibliographic Details
Application Number
—
Jurisdiction
EP
Data Source
google_patents
Last Synchronized
2026-11-04 04:29
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/20
Claims Analysis
Total Claims: 10
Claim 1
Independent Claim
132 words
A method of training a sound recognition system comprising preparing masked template signals from example sounds by representing sounds to be recognized by first or second types of signals in each region of the frequency spectrum at successive times, the first type being a signal level representative of the example sounds in each region where the signal is above a noise level and the second type being a masking level representative of the said noise level where the signal in the region is below the said noise level, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level encountered in that region during the time in which active preparation of the masked template signals takes place.
Claim 2
Independent Claim
132 words
A method of training a sound recognition system comprising preparing masked template signals from example sounds by representing sounds to be recognized by first or second types of signals in each region of the frequency spectrum at successive times, the first type being a signal level representative of the example sounds in each region where the signal is above a noise level and the second type being a masking level representative of the said noise level where the signal in the region is below the said noise level, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level encountered in that region during the time in which active preparation of the masked template signals takes place.
Claim 3
Dependent Claim
360 words
Apparatus for use in processing signals derived from speech comprising means for storing a plurality of masked template signals, each representing the frequency spectrum of a sound to be recognized and each represented by a masking level where a noise level found in preparing the template signals was greater than the input signals used in preparation, means for estimating the input noise level, means for deriving masked input signals during recognition by representing any region of the frequency spectrum of an input signal below noise level with a masking level which is representative of the noise level in that region, and means for deriving a distance measure for each region of the frequency spectrum of input signals during recognition by comparing each region of each masked input signal with the corresponding region of the masked template signals, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level for that region found in preparing the template signals, and the means for deriving a distance measure for each region does so according to the first of the following rules and at least one other of them, considering corresponding regions:- (a) if both the masked input and template signals are represented by unmasked signal levels, the distance measure is derived from those unmasked levels, (b) if the masked input and template signals are represented by a masking level and an unmasked signal level, respectively, and if the masking level is greater than the unmasked level then the distance measure is given a predetermined value, otherwise the distance measure is derived from the masking and unmasked levels, (c) if the marked input and template signals are represented by an unmasked signal level and a masking level, repsectively, the distance measure is derived from the unmasked and masking levels, and (d) if both the masked and template signals are represented by masking levels, and if the input masking level is greater than the template masking level then the distance measure is given the same, or a different, predetermined value, otherwise the distance measure is derived from the masking levels.
Claim 4
Dependent Claim
360 words
Apparatus for use in processing signals derived from speech comprising means for storing a plurality of masked template signals, each representing the frequency spectrum of a sound to be recognized and each represented by a masking level where a noise level found in preparing the template signals was greater than the input signals used in preparation, means for estimating the input noise level, means for deriving masked input signals during recognition by representing any region of the frequency spectrum of an input signal below noise level with a masking level which is representative of the noise level in that region, and means for deriving a distance measure for each region of the frequency spectrum of input signals during recognition by comparing each region of each masked input signal with the corresponding region of the masked template signals, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level for that region found in preparing the template signals, and the means for deriving a distance measure for each region does so according to the first of the following rules and at least one other of them, considering corresponding regions:- (a) if both the masked input and template signals are represented by unmasked signal levels, the distance measure is derived from those unmasked levels, (b) if the masked input and template signals are represented by a masking level and an unmasked signal level, respectively, and if the masking level is greater than the unmasked level then the distance measure is given a predetermined value, otherwise the distance measure is derived from the masking and unmasked levels, (c) if the marked input and template signals are represented by an unmasked signal level and a masking level, repsectively, the distance measure is derived from the unmasked and masking levels, and (d) if both the masked and template signals are represented by masking levels, and if the input masking level is greater than the template masking level then the distance measure is given the same, or a different, predetermined value, otherwise the distance measure is derived from the masking levels.
Claim 5
Independent Claim
354 words
A method of processing signals derived from speech comprising the steps of storing a plurality of masked template signals, each representing the frequency spectrum of a sound to be recognized and each represented by a masking level where a noise level found in preparing the template signals was greater than the input signals used in preparation, estimating the input noise level, deriving masked input signals during recognition by representing any region of the frequency spectrum of an input signal below noise level with a masking level which is representative of the noise level in that region, and deriving a distance measure for each region of the frequency spectrum of input signals during recognition by comparing each region of each masked input signal with the corresponding region of the masked template signals, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level for that region found in preparing the template signals, and the step of deriving a distance measure for each region does so according to the first of the following rules and at least one other of them, considering corresponding regions:- (a) if both the masked input and template signals are represented by unmasked signal levels, the distance measure is derived from those unmasked levels, (b) if the masked input and template signals are represented by a masking level and an unmasked signal level, respectively, and if the masking level is greater than the unmasked level then the distance measure is given a predetermined value, otherwise the distance measure is derived from the masking and unmasked levels, (c) if the marked input and template signals are represented by an unmasked signal level and a masking level, respectively, the distance measure is derived from the unmasked and masking levels, and (d) if both the masked and template signals are represented by masking levels, and if the input masking level is greater than the template masking level then the distance measure is given the same, or a different, predetermined value, otherwise the distance measure is derived from the masking levels.
Claim 6
Independent Claim
354 words
A method of processing signals derived from speech comprising the steps of storing a plurality of masked template signals, each representing the frequency spectrum of a sound to be recognized and each represented by a masking level where a noise level found in preparing the template signals was greater than the input signals used in preparation, estimating the input noise level, deriving masked input signals during recognition by representing any region of the frequency spectrum of an input signal below noise level with a masking level which is representative of the noise level in that region, and deriving a distance measure for each region of the frequency spectrum of input signals during recognition by comparing each region of each masked input signal with the corresponding region of the masked template signals, characterised in that a respective said noise level is provided for each region, and the said noise level for each region is the highest noise level for that region found in preparing the template signals, and the step of deriving a distance measure for each region does so according to the first of the following rules and at least one other of them, considering corresponding regions:- (a) if both the masked input and template signals are represented by unmasked signal levels, the distance measure is derived from those unmasked levels, (b) if the masked input and template signals are represented by a masking level and an unmasked signal level, respectively, and if the masking level is greater than the unmasked level then the distance measure is given a predetermined value, otherwise the distance measure is derived from the masking and unmasked levels, (c) if the marked input and template signals are represented by an unmasked signal level and a masking level, respectively, the distance measure is derived from the unmasked and masking levels, and (d) if both the masked and template signals are represented by masking levels, and if the input masking level is greater than the template masking level then the distance measure is given the same, or a different, predetermined value, otherwise the distance measure is derived from the masking levels.
Claim 7
Independent Claim
20 words
A method or apparatus according to any preceding claim characterised in that the spectral regions are channels derived by filters.
Claim 8
Dependent Claim
42 words
A method or apparatus according to Claim 2 or 3, or Claim 4 insofar as dependent on Claim 2 or 3, characterised in that each distance measure is derived by squaring the difference between the two levels from which it is derived.
Claim 9
Dependent Claim
38 words
Apparatus according to any of Claims 2, 4 or 5 characterised in that the means for estimating the noise input level includes a first microphone, and the means for deriving masked input levels includes a second separate microphone.
Claim 10
Dependent Claim
43 words
Apparatus according to any of Claims 2, 4 or 5 characterised in that the means for estimating the noise input level and the means for deriving masked input levels include a common microphone and common means for separating noise and noise plus speech.
References Cited (Backward Citations)
No citation records indexed.
USPTO Annuity & Maintenance Windows
Maintenance schedule is calculated for granted US utility patents.