US5274739A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Product code memory Itakura-Saito (MIS) measure for sound recognition
Assignee: Boeing Co
•
Filing: 1992-04-15
•
Granted: 1993-12-28
•
Expires: 2012-04-15
Abstract
Abstract A SOUND RECOGNITION PROCESS comprises training a sound recognition system to a shape Hidden Markov Model (HMM) from a shape codebook of shape codewords, and to a separate gain HMM from a gain codebook of gain codewords. An unknown token is assigned to the sound category associated with the shape HMM and gain HMM with the greatest probability product. Shape and gain codewords replacements for incoming Linear Predictive Coded (LPC) vectors are selected by choosing the codeword with minimum distortion between the codeword and a combination of the vector and some of the preceding codewords, exponentially weighted by time.
Inventors
- 👤 Jeffrey P. Woodard
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 13:09
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/10; G10L15/14; G10L15/142; G10L15/144
Claims Analysis
Total Claims: 20
Claim 1
Independent Claim
340 words
A method of recognizing a sound token, the method comprising the steps of: (a) converting an acoustic signal from the token into an analog electric signal; (b) converting the analog electrical signal into a digital electrical signal; (c) windowing the digital signal into short-time frames by Hamming windows; (d) providing a sound recognition system of the Linear Predictive Coding (LPC)/Hidden Markov Model (HMM) type with a shape codebook of shape codewords in a memory, a plurality of shape HMMs, a gain codebook of gain codewords in a memory, and a plurality of gain HMMs, each combination of shape HMM and gain HMM being associated with a sound category; (e) Linear Predictive Coding the windowed digital signal into an LPC vector sequence, one LPC vector for each window; (f) providing the LPC vector sequence to the sound recognition system; (g) determining, for each LPC vector of the LPC vector sequence, a shape codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding shape codewords, and replacing the respective vector with the determined codeword, thereby producing a shape codeword sequence; (h) determining, for each LPC vector of the LPC vector sequence, a gain codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding gain codewords, and replacing the respective vector with the determined codeword, thereby producing a gain codeword sequence; (i) comparing each shape HMM with the shape codeword sequence and determining the shape probability of the compared shape HMM; (j) comparing each gain HMM with the gain codeword sequence and determining the gain probability of the compared gain HMM; (k) determining a probability product of each shape HMM probability according to step (i) multiplied by each gain HMM probability according to step (j); (l) associating the probability product with the sound category associated with that combination of shape HMM and gain HMM; and (m) displaying the sound category with the greatest probability product.
Claim 2
Independent Claim
340 words
A method of recognizing a sound token, the method comprising the steps of: (a) converting an acoustic signal from the token into an analog electric signal; (b) converting the analog electrical signal into a digital electrical signal; (c) windowing the digital signal into short-time frames by Hamming windows; (d) providing a sound recognition system of the Linear Predictive Coding (LPC)/Hidden Markov Model (HMM) type with a shape codebook of shape codewords in a memory, a plurality of shape HMMs, a gain codebook of gain codewords in a memory, and a plurality of gain HMMs, each combination of shape HMM and gain HMM being associated with a sound category; (e) Linear Predictive Coding the windowed digital signal into an LPC vector sequence, one LPC vector for each window; (f) providing the LPC vector sequence to the sound recognition system; (g) determining, for each LPC vector of the LPC vector sequence, a shape codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding shape codewords, and replacing the respective vector with the determined codeword, thereby producing a shape codeword sequence; (h) determining, for each LPC vector of the LPC vector sequence, a gain codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding gain codewords, and replacing the respective vector with the determined codeword, thereby producing a gain codeword sequence; (i) comparing each shape HMM with the shape codeword sequence and determining the shape probability of the compared shape HMM; (j) comparing each gain HMM with the gain codeword sequence and determining the gain probability of the compared gain HMM; (k) determining a probability product of each shape HMM probability according to step (i) multiplied by each gain HMM probability according to step (j); (l) associating the probability product with the sound category associated with that combination of shape HMM and gain HMM; and (m) displaying the sound category with the greatest probability product.
Claim 3
Dependent Claim
31 words
The method of claim 1, further comprising the step of displaying a sound category associated with the shape HMM and gain HMM producing each probability product, together with the probability product.
Claim 4
Dependent Claim
96 words
The method of claim 1, wherein: (a) the step of determining the probability product of each shape HMM probability according to step (i) of claim 2 multiplied by each gain HMM probability according to step (j) of claim 2 further comprises the step of appending such probability product to a sequence of such probability products; and (b) the method of claim 2 further comprises the steps of: (1) determining a greatest such probability product in the sequence of such probability products; and (2) displaying a representation of the sound category associated with such greatest probability product.
Claim 5
Dependent Claim
159 words
The method of claim 1, wherein each of the steps of replacing each LPC vector with the respective shape and gain codewords closest to a combination of that LPC vector and a weighted selection of respective preceding shape and gain codewords, thereby producing respective shape and codeword sequences, further comprises: (a) comparing a codeword candidate with the LPC vector and producing an LPC vector distortion measure; (b) adding the LPC vector distortion measure to a sum; (c) comparing the codeword candidate with the next preceding codeword and producing a next codeword distortion measure; (d) multiplying the next codeword distortion measure by a next preselected constant to produce a next product; (e) adding the next product to the sum; (f) repeating steps (c) through (e) a preselected number of times; (g) noting the sum for the codeword candidate under consideration; (h) repeating steps (a) through (g) for each codeword candidate; and (i) determining a codeword candidate with the lowest sum.
Claim 6
Dependent Claim
30 words
The method of claim 4, wherein the step (f) preselected number of times for shape codewords is the same as the step (f) preselected number of times for gain codewords.
Claim 7
Dependent Claim
29 words
The method of claim 4, wherein the step (f) preselected number of times for shape codewords is different from the step (f) preselected number of times for gain codewords.
Claim 8
Dependent Claim
32 words
The method of claim 4, wherein each next preselected constant equals the previously calculated preselected constant multiplied by a fixed constant which lies between zero and one, thereby producing an exponential weighting.
Claim 9
Dependent Claim
22 words
The method of claim 7, wherein the fixed constant for shape codewords is the same as the fixed constant for gain codewords.
Claim 10
Dependent Claim
21 words
The method of claim 7, wherein the fixed constant for shape codewords is different from the fixed constant for gain codewords.
Claim 11
Independent Claim
364 words
An apparatus for recognizing a sound token the apparatus comprising: (a) means for converting an acoustic signal from the token into an analog electric signal; (b) means for converting the analog electrical signal into a digital electrical signal; (c) means for windowing the digital signal into short-time frames by Hamming windows; (d) a sound recognition system of the Linear Predictive Coding (LPC)/Hidden Markov Model (HMM) type comprising a shape codebook of shape codewords in a memory, a plurality of shape HMMs, a gain codebook of gain codewords in a memory, and a plurality of gain HMMs, each combination of shape HMM and gain HMM being associated with a sound category; (e) means for Linear Predictive Coding the windowed digital signal into an LPC vector sequence, one LPC vector for each window; (f) means for providing the LPC vector sequence to the sound recognition system; (g) means for determining, for each LPC vector of the LPC vector sequence, a shape codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding shape codewords, for replacing the respective vector with the determined codeword, and for thereby producing a shape codeword sequence; (h) means for determining, for each LPC vector of the LPC vector sequence, a gain codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding gain codewords, for replacing the respective vector with the determined codeword, and for thereby producing a gain codeword sequence; (i) means for comparing each shape HMM with the shape codeword sequence and determining the shape probability of the compared shape HMM; (j) means for comparing each gain HMM with the gain codeword sequence and determining the gain probability of the compared gain HMM; (k) means for determining a probability product of each shape HMM probability according to step (i) multiplied by each gain HMM probability according to step (j); (l) means for associating the probability product with the sound category associated with that combination of shape HMM and gain HMM; and (m) means for displaying the sound category with the greatest probability product.
Claim 12
Independent Claim
364 words
An apparatus for recognizing a sound token the apparatus comprising: (a) means for converting an acoustic signal from the token into an analog electric signal; (b) means for converting the analog electrical signal into a digital electrical signal; (c) means for windowing the digital signal into short-time frames by Hamming windows; (d) a sound recognition system of the Linear Predictive Coding (LPC)/Hidden Markov Model (HMM) type comprising a shape codebook of shape codewords in a memory, a plurality of shape HMMs, a gain codebook of gain codewords in a memory, and a plurality of gain HMMs, each combination of shape HMM and gain HMM being associated with a sound category; (e) means for Linear Predictive Coding the windowed digital signal into an LPC vector sequence, one LPC vector for each window; (f) means for providing the LPC vector sequence to the sound recognition system; (g) means for determining, for each LPC vector of the LPC vector sequence, a shape codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding shape codewords, for replacing the respective vector with the determined codeword, and for thereby producing a shape codeword sequence; (h) means for determining, for each LPC vector of the LPC vector sequence, a gain codeword having the least distortion, according to a pre-selected Itakura-Saito measure, to a combination of that LPC vector and a weighted selection of preceding gain codewords, for replacing the respective vector with the determined codeword, and for thereby producing a gain codeword sequence; (i) means for comparing each shape HMM with the shape codeword sequence and determining the shape probability of the compared shape HMM; (j) means for comparing each gain HMM with the gain codeword sequence and determining the gain probability of the compared gain HMM; (k) means for determining a probability product of each shape HMM probability according to step (i) multiplied by each gain HMM probability according to step (j); (l) means for associating the probability product with the sound category associated with that combination of shape HMM and gain HMM; and (m) means for displaying the sound category with the greatest probability product.
Claim 13
Dependent Claim
75 words
The apparatus of claim 10, wherein: (a) the means for determining the probability product of each shape HMM probability multiplied by each gain HMM probability further comprises means for appending such probability product to a sequence of such probability products; and (b) the apparatus of claim 10 further comprises means for displaying a representation of the sound category associated with the shape HMM and gain HMM producing each probability product, together with the probability product.
Claim 14
Dependent Claim
94 words
The apparatus of claim 10, wherein: (a) the means for determining the probability product of each shape HMM probability according to step (i) of claim 10 multiplied by each gain HMM probability according to step (j) of claim 10 further comprises means for appending such probability product to a sequence of such probability products; and (b) the apparatus of claim 10 further comprises means for: (1) determining a greatest such probability product in the sequence of such probability products; and (2) displaying a representation of the sound category associated with such greatest probability product.
Claim 15
Dependent Claim
179 words
The apparatus of claim 10, wherein each of the means for replacing each LPC vector with the respective shape and gain codewords closest to a combination of that LPC vector and a weighted selection of respective preceding shape and gain codewords, thereby producing respective shape and codeword sequences, further comprises: (a) means for comparing a codeword candidate with the LPC vector and producing an LPC vector distortion measure; (b) means for adding the LPC vector distortion measure to a sum; (c) means for comparing the codeword candidate with the next preceding codeword and producing a next codeword distortion measure; (d) means for multiplying the next codeword distortion measure by a next preselected constant to produce a next product; (e) means for adding the next product to the sum; (f) means for repeatedly activating elements (c) through (e) a preselected number of times; (g) means for noting the sum for the codeword candidate under consideration; (h) means for repeatedly activating elements (a) through (g) for each codeword candidate; and (i) means for determining a codeword candidate with the lowest sum.
Claim 16
Dependent Claim
30 words
The apparatus of claim 13, wherein the element (f) preselected number of times for shape codewords is the same as the element (f) preselected number of times for gain codewords.
Claim 17
Dependent Claim
29 words
The apparatus of claim 13, wherein the element (f) preselected number of times for shape codewords is different from the element (f) preselected number of times for gain codewords.
Claim 18
Dependent Claim
32 words
The apparatus of claim 13, wherein each next preselected constant equals the previously calculated preselected constant multiplied by a fixed constant which lies between zero and one, thereby producing an exponential weighting.
Claim 19
Dependent Claim
22 words
The apparatus of claim 16, wherein the fixed constant for shape codewords is the same as the fixed constant for gain codewords.
Claim 20
Dependent Claim
21 words
The apparatus of claim 16, wherein the fixed constant for shape codewords is different from the fixed constant for gain codewords.
References Cited (Backward Citations)
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 1996-12-27 | 1997-06-28 | 1997-12-28 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2000-12-27 | 2001-06-28 | 2001-12-28 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2004-12-27 | 2005-06-28 | 2005-12-28 | $7700.00 | Window Closed / Lapsed |