US5806022A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Method and system for performing speech recognition
Assignee: Nuance Communications Inc
•
Filing: 1995-12-20
•
Granted: 1998-09-08
•
Expires: 2015-12-20
Abstract
Abstract Speech recognition processing is compensated for improving robustness of speech recognition in the presence of enhanced speech signals. The compensation overcomes the adverse effects that speech signal enhancement may have on speech recognition performance, where speech signal enhancement causes acoustical mismatches between recognition models trained using unenhanced speech signals and feature data extracted from enhanced speech signals. Compensation is provided at the front end of an automatic speech recognition system by combining linear predictive coding and mel-based cepstral parameter analysis for computing cepstral features of transmitted speech signals used for speech recognition processing by selectively weighting mel-filter banks when processing frequency domain representations of the enhanced speech signals.
Inventors
- 👤 Mazin G. Rahim
- 👤 Jay Gordon Wilpon
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-08 05:50
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/02; G10L15/08; G10L21/00; G10L21/02; G10L25/00; G10L25/03; G10L25/12; G10L25/18
Claims Analysis
Total Claims: 55
Claim 1
Independent Claim
87 words
A method of compensating for enhancement of speech signals for improving speech recognition performance, the method comprising the steps of: receiving speech signals enhanced by an audio enhancer, said enhancer introducing frequency gain variations to the speech signals; converting said enhanced speech signals to frequency domain representation; compensating said gain variations introduced to said speech signals by said enhancer using said frequency domain representation of said enhanced speech signals; and, computing cepstral features of the enhanced speech signals after they have been compensated by the compensating step.
Claim 2
Independent Claim
87 words
A method of compensating for enhancement of speech signals for improving speech recognition performance, the method comprising the steps of: receiving speech signals enhanced by an audio enhancer, said enhancer introducing frequency gain variations to the speech signals; converting said enhanced speech signals to frequency domain representation; compensating said gain variations introduced to said speech signals by said enhancer using said frequency domain representation of said enhanced speech signals; and, computing cepstral features of the enhanced speech signals after they have been compensated by the compensating step.
Claim 3
Dependent Claim
21 words
The method of claim 1, further comprising the step of: comparing said features to recognition models stored in a recognizer memory.
Claim 4
Dependent Claim
20 words
The method of claim 2 wherein said recognition models have been trained using speech signals not enhanced by said enhancer.
Claim 5
Dependent Claim
95 words
The method of claim 1, wherein the step of compensating further comprises the step of computing compensated mel-LPC coefficients, said step of computing compensated mel-LPC coefficients further comprising the steps of: preemphasizing the enhanced speech signals for spectral flattening; blocking the enhanced speech signals into consecutive frames of speech samples; multiplying each of the frames by a Hamming window; transforming each of the Hamming window frames from time to frequency domain representations; filtering each of the frequency domain representations through a set of weighted mel-filter banks; and, generating autocorrelation coefficients from the filtered frequency representations.
Claim 6
Dependent Claim
26 words
The method of claim 4, further comprising the steps of: converting the autocorrelation coefficients to LPC coefficients; and, computing the cepstral parameters from the LPC coefficients.
Claim 7
Dependent Claim
15 words
The method of claim 4, wherein the mel-filter banks below a specific frequency are attenuated.
Claim 8
Dependent Claim
13 words
The method of claim 6, wherein the specific frequency is approximately 290 Hz.
Claim 9
Independent Claim
131 words
A system for compensating for enhancement of speech signals for improving speech recognition performances the system comprising: an enhancer for selectively varying the gain of select frequencies of time varying speech signals transmitted on a network path; a receiver for receiving said enhanced speech signals; a frequency domain converter for converting the enhanced speech signal received at the receiver to frequency domain representations; a compensator for receiving the frequency domain representations of the speech signals enhanced by said enhancer, wherein said compensator introduces gain variations to the frequency domain representations of the speech signals transmitted on the path for compensating for gain variations introduced to the speech signals by said enhancer; and, a cepstral feature computer for computing cepstral features from the compensated, frequency domain representations of the enhanced speech signals.
Claim 10
Independent Claim
131 words
A system for compensating for enhancement of speech signals for improving speech recognition performances the system comprising: an enhancer for selectively varying the gain of select frequencies of time varying speech signals transmitted on a network path; a receiver for receiving said enhanced speech signals; a frequency domain converter for converting the enhanced speech signal received at the receiver to frequency domain representations; a compensator for receiving the frequency domain representations of the speech signals enhanced by said enhancer, wherein said compensator introduces gain variations to the frequency domain representations of the speech signals transmitted on the path for compensating for gain variations introduced to the speech signals by said enhancer; and, a cepstral feature computer for computing cepstral features from the compensated, frequency domain representations of the enhanced speech signals.
Claim 11
Dependent Claim
19 words
The system of claim 8, further comprising: a recognizer for receiving said cepstral features from said cepstral feature computer.
Claim 12
Dependent Claim
35 words
The system of claim 9 wherein said recognizer compares said features to recognition models trained on the path using speech signals not enhanced by said enhancer, said recognition models being stored in a recognizer memory.
Claim 13
Dependent Claim
25 words
The system of claim 8, wherein the compensator passes the frequency domain representations of the enhanced speech signals through a set of weighted mel-filter banks.
Claim 14
Dependent Claim
18 words
The system of claim 11, where the cepstral computer generates autocorrelation coefficients from the compensated frequency domain representations.
Claim 15
Dependent Claim
30 words
The system of claim 12, wherein the cepstral computer further performs the operations of: converting the autocorrelation coefficients to LPC coefficients; and, computing said cepstral features from the LPC coefficients.
Claim 16
Dependent Claim
15 words
The system of claim 12, wherein the mel-filter banks below a specific frequency are attenuated.
Claim 17
Dependent Claim
13 words
The system of claim 14, wherein the specific frequency is approximately 290 Hz.
Claim 18
Independent Claim
138 words
A method of computing compensated cepstral features of speech signals enhanced by a first enhancer using a computing system, for improving speech recognition performance, said computing system comprising a memory and at least one processor, said memory including mel-filter bank weighting values, the method comprising the steps of: receiving the enhanced speech signals at the processor; preemphasizing the enhanced speech signals for spectral flattening; blocking the enhanced speech signals into consecutive frames of speech samples; multiplying each of the frames by a Hamming window; transforming each of the Hamming window frames from time to frequency domain representations; filtering each of the frequency domain representations using the mel-filter bank weighting values stored in the memory; generating autocorrelation coefficients from the filtered frequency representations; converting the autocorrelation coefficients to LPC coefficients; and, computing the cepstral parameters from the LPC coefficients.
Claim 19
Independent Claim
138 words
A method of computing compensated cepstral features of speech signals enhanced by a first enhancer using a computing system, for improving speech recognition performance, said computing system comprising a memory and at least one processor, said memory including mel-filter bank weighting values, the method comprising the steps of: receiving the enhanced speech signals at the processor; preemphasizing the enhanced speech signals for spectral flattening; blocking the enhanced speech signals into consecutive frames of speech samples; multiplying each of the frames by a Hamming window; transforming each of the Hamming window frames from time to frequency domain representations; filtering each of the frequency domain representations using the mel-filter bank weighting values stored in the memory; generating autocorrelation coefficients from the filtered frequency representations; converting the autocorrelation coefficients to LPC coefficients; and, computing the cepstral parameters from the LPC coefficients.
Claim 20
Dependent Claim
21 words
The method of claim 16, further comprising the step of: comparing said features to recognition models stored in a recognizer memory.
Claim 21
Dependent Claim
19 words
The method of claim 17 wherein said recognition models have been trained using speech signals not having been enhanced.
Claim 22
Dependent Claim
18 words
The method of claim 16, wherein the values of the mel-filter banks below a specific frequency are attenuated.
Claim 23
Dependent Claim
13 words
The method of claim 19, wherein the specific frequency is approximately 290 Hz.
Claim 24
Dependent Claim
46 words
The method of claim 16 further comprising the step of: modifying the mel-filter bank weighting values stored in the memory for compensating for enhancement to the speech signals by a second enhancer, said second enhancer providing enhancement differing from said enhancement provided by said first enhancer.
Claim 25
Independent Claim
140 words
A method of computing compensated cepstral features of speech signals using a computing system, said computing system comprising a memory and at least one processor, said memory including mel-filter bank weighting values providing compensation for the computation of cepstral features for speech signal enhancement by an enhancer, the method comprising the steps of: receiving the speech signals at the processor; preemphasizing the speech signals for spectral flattening; blocking the speech signals into consecutive frames of speech samples; multiplying each of the frames by a Hamming window; transforming each of the Hamming window frames from time to frequency domain representations; filtering each of the frequency domain representations using the mel-filter bank weighting values stored in the memory; generating autocorrelation coefficients from the filtered frequency representations; converting the autocorrelation coefficients to LPC coefficients; and, computing the cepstral parameters from the LPC coefficients.
Claim 26
Independent Claim
140 words
A method of computing compensated cepstral features of speech signals using a computing system, said computing system comprising a memory and at least one processor, said memory including mel-filter bank weighting values providing compensation for the computation of cepstral features for speech signal enhancement by an enhancer, the method comprising the steps of: receiving the speech signals at the processor; preemphasizing the speech signals for spectral flattening; blocking the speech signals into consecutive frames of speech samples; multiplying each of the frames by a Hamming window; transforming each of the Hamming window frames from time to frequency domain representations; filtering each of the frequency domain representations using the mel-filter bank weighting values stored in the memory; generating autocorrelation coefficients from the filtered frequency representations; converting the autocorrelation coefficients to LPC coefficients; and, computing the cepstral parameters from the LPC coefficients.
Claim 27
Dependent Claim
21 words
The method of claim 22, further comprising the step of: comparing said features to recognition models stored in a recognizer memory.
Claim 28
Dependent Claim
19 words
The method of claim 23 wherein said recognition models have been trained using speech signals not having been enhanced.
Claim 29
Dependent Claim
68 words
A feature extractor apparatus for generating speech feature signals characterizing enhanced speech signals comprising: a frequency domain converter for generating a set of spectral samples representing the speech signals; a weighting unit for selectively weighting the set of spectral samples said weightings being determined by characteristics of said enhanced speech; and, a feature computer for generating speech characterizing feature signals based on the weighted set of spectral samples.
Claim 30
Dependent Claim
68 words
A feature extractor apparatus for generating speech feature signals characterizing enhanced speech signals comprising: a frequency domain converter for generating a set of spectral samples representing the speech signals; a weighting unit for selectively weighting the set of spectral samples said weightings being determined by characteristics of said enhanced speech; and, a feature computer for generating speech characterizing feature signals based on the weighted set of spectral samples.
Claim 31
Dependent Claim
23 words
The apparatus of claim 25, further comprising a receiver, wherein the receiver is connected to the frequency domain converter through a preprocessing circuit.
Claim 32
Dependent Claim
13 words
The apparatus of claim 26, wherein the preprocessing circuit comprises a preemphasizer circuit.
Claim 33
Dependent Claim
13 words
The apparatus of claim 26, wherein the preprocessing circuit comprises a frame blocker.
Claim 34
Dependent Claim
13 words
The apparatus of claim 26, wherein the preprocessing circuit comprises a Hamming multiplier.
Claim 35
Dependent Claim
18 words
The apparatus of claim 25, wherein the weighting unit comprises a set of M selectively weighted mel-filter banks.
Claim 36
Independent Claim
57 words
A method for generating speech feature signals characterizing enhanced speech signals, the method comprising the steps of: generating a set of spectral parameters characterizing speech signals; weighting the set of spectral parameters said weightings being determined by characteristics of said enhanced speech signals; and generating speech characterizing feature signals based on the weighted set of spectral samples.
Claim 37
Independent Claim
57 words
A method for generating speech feature signals characterizing enhanced speech signals, the method comprising the steps of: generating a set of spectral parameters characterizing speech signals; weighting the set of spectral parameters said weightings being determined by characteristics of said enhanced speech signals; and generating speech characterizing feature signals based on the weighted set of spectral samples.
Claim 38
Independent Claim
59 words
A system for recognizing speech signals and for compensating for network enhancement of said signals comprising: a filter for compensating for a network enhancement component of enhanced speech signals; a feature extractor for extracting features based on filtered speech signals from said filter; and a speech recognizer for recognizing speech signals based on extracted features from the feature extractor.
Claim 39
Independent Claim
59 words
A system for recognizing speech signals and for compensating for network enhancement of said signals comprising: a filter for compensating for a network enhancement component of enhanced speech signals; a feature extractor for extracting features based on filtered speech signals from said filter; and a speech recognizer for recognizing speech signals based on extracted features from the feature extractor.
Claim 40
Dependent Claim
16 words
The system of claim 32, wherein the filter comprises a set of selectively weighted mel-filter banks.
Claim 41
Dependent Claim
22 words
The system of claim 32, wherein the feature extractor comprises a combination of a linear predictive coding analyzer and a mel-filter analyzer.
Claim 42
Independent Claim
72 words
A method of generating feature signals from speech signals having a network enhancement comprising the following steps: receiving the speech signals; blocking the speech signals into frames; forming frequency domain representations of said blocked speech signals; passing said frequency domain representations through mel-filter banks to compensate for the network enhancement; and performing in combination linear predictive coding and cepstral recursion analysis on the blocked speech signals to produce mel-LPC cepstrum feature signals.
Claim 43
Independent Claim
72 words
A method of generating feature signals from speech signals having a network enhancement comprising the following steps: receiving the speech signals; blocking the speech signals into frames; forming frequency domain representations of said blocked speech signals; passing said frequency domain representations through mel-filter banks to compensate for the network enhancement; and performing in combination linear predictive coding and cepstral recursion analysis on the blocked speech signals to produce mel-LPC cepstrum feature signals.
Claim 44
Dependent Claim
19 words
The method of claim 35 further comprising the step of comparing the mel-LPC cepstrum features to a recognition model.
Claim 45
Dependent Claim
39 words
The method of claim 35 further comprising the step of: utilizing a mel-filter bank to filter the blocked speech signals and produce mel-filtered signals which are then analyzed by performing linear predictive coding and cepstral recursion analysis in combination.
Claim 46
Dependent Claim
21 words
The method of claim 35 further comprising the step of preemphasizing the speech signals to achieve spectral flattening of said signals.
Claim 47
Dependent Claim
19 words
The method of claim 38 further comprising the step of multiplying each of the frames by a Hamming window.
Claim 48
Dependent Claim
26 words
The method of claim 39 further comprising the step of transforming each of the Hamming window frames from a time representation to a frequency domain representation.
Claim 49
Dependent Claim
23 words
The method of claim 40 further comprising filtering each of the frequency domain representations using mel-filter bank weighting values stored in a memory.
Claim 50
Dependent Claim
17 words
The method of claim 41 further comprising the step generating autocorrelation coefficients from the filtered frequency representations.
Claim 51
Dependent Claim
34 words
The method of claim 42, wherein the linear predictive coding analysis operates to convert the autocorrelation coefficients to LPC coefficients and the cepstral recursion analysis operates to compute cepstral parameters from the LPC coefficients.
Claim 52
Independent Claim
88 words
A method of compensating for enhancement of speech signals to improve speech recognition performance, the method comprising the steps of: receiving speech signals enhanced by an audio enhancer, said enhancer introducing frequency gain variations to the speech signals; converting said enhanced speech signals to a frequency domain representation; compensating said gain variations introduced to said speech signals by said enhancer using said frequency domain representation of said enhanced speech signals; and generating feature signals characterizing the enhanced speech signals after they have been compensated by the compensating step.
Claim 53
Independent Claim
88 words
A method of compensating for enhancement of speech signals to improve speech recognition performance, the method comprising the steps of: receiving speech signals enhanced by an audio enhancer, said enhancer introducing frequency gain variations to the speech signals; converting said enhanced speech signals to a frequency domain representation; compensating said gain variations introduced to said speech signals by said enhancer using said frequency domain representation of said enhanced speech signals; and generating feature signals characterizing the enhanced speech signals after they have been compensated by the compensating step.
Claim 54
Dependent Claim
21 words
The method of claim 44 further comprising the step of: comparing said features to recognition models stored in a recognizer memory.
Claim 55
Dependent Claim
16 words
The method of claim 45, wherein said recognition models have been trained using enhanced speech signals.
References Cited (Backward Citations)
US5195132A
Cited
US5195132B1
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2001-09-07 | 2002-03-09 | 2002-09-08 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2005-09-07 | 2006-03-09 | 2006-09-08 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2009-09-07 | 2010-03-09 | 2010-09-08 | $7700.00 | Window Closed / Lapsed |