Patent US5193142A successfully refreshed from Google Patents.
US5193142A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Training module for estimating mixture gaussian densities for speech-unit models in speech recognition systems
Assignee: Panasonic Holdings Corp
•
Filing: 1990-11-15
•
Granted: 1993-03-09
•
Expires: 2010-11-15
Abstract
Abstract A model-training module generates mixture Gaussian density models from speech training data for continuous, or isolated word HMM-based speech recognition systems. Speech feature sequences are labeled into segments of states of speech units using Viterbi-decoding based optimized segmentation algorithm. Each segment is modeled by a Gaussian density, and the parameters are estimated by sample mean and sample covariance. A mixture Gaussian density is generated for each state of each speech unit by merging the Gaussian densities of all the segments with the same corresponding label. The resulting number of mixture components is proportional to the dispension and sample size of the training data. A single, fully merged, Gaussian density is also generated for each state of each speech unit. The covariance matrices of the mixture components are selectively smoothed by a measure of relative sharpness of the Gaussian density. The weights of the mixture components are set uniformly initially, and are reestimated using a segmental-average procedure. The weighting coefficients, together with the Gaussian densities, then become the models of speech units for use in speech recognition.
Inventors
- 👤 Yunxin Zhao
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-01 16:40
CPC Classifications
G; G10; G10L; G10L15/00; G10L15/08; G10L15/14; G10L15/142; G10L15/144
Claims Analysis
Total Claims: 32
Claim 1
Independent Claim
151 words
A speech recognition system including a training module for estimating Gaussian densities for use in speech-unit models of the speech recognition system, said speech recognition system comprising: means for receiving an input speech signal; segmentation means for segmenting said speech signal into a plurality of speech units, each speech unit having a plurality of states, each state of each speech unit being a segment of said signal, each segment including a plurality of frames; means for grouping all segments corresponding to like speech unit states for like speech units into respective sets of segments for all states of all speech units; means for estimating parameters defining a Gaussian density for each segment of each set of segments; means for merging said Gaussian densities of each set into a mixture Gaussian density for the set; and means for using said mixture Gaussian densities for identifying speech units of additional input speech signals.
Claim 2
Dependent Claim
151 words
A speech recognition system including a training module for estimating Gaussian densities for use in speech-unit models of the speech recognition system, said speech recognition system comprising: means for receiving an input speech signal; segmentation means for segmenting said speech signal into a plurality of speech units, each speech unit having a plurality of states, each state of each speech unit being a segment of said signal, each segment including a plurality of frames; means for grouping all segments corresponding to like speech unit states for like speech units into respective sets of segments for all states of all speech units; means for estimating parameters defining a Gaussian density for each segment of each set of segments; means for merging said Gaussian densities of each set into a mixture Gaussian density for the set; and means for using said mixture Gaussian densities for identifying speech units of additional input speech signals.
Claim 3
Dependent Claim
58 words
The speech recognition system of claim 1, wherein said means for merging comprises: means for calculating distances between all pairs of Gaussian densities in a set; means for sorting the calculated distances to find a Gaussian density pair separated by a minimum distance; and means for merging said minimum distance Gaussian density pair into a single Gaussian density.
Claim 4
Dependent Claim
34 words
The speech recognition system of claim 2, further comprising: means for iteratively carrying out a merging procedure of said merging means until the distance between the minimum distance pair is above a predetermined threshold.
Claim 5
Dependent Claim
89 words
The speech recognition system of claim 2 wherein said minimum distance pair is represented by indices i* and j*, and a said merged Gaussian density is represented by μ and C, and wherein said means for merging said minimum distance Gaussian pair comprises: means for calculating μ, wherein μ is given by: μ=λ.sub.i ·μ.sub.i ·+λ.sub.j ·μ.sub.j ·; and means for calculating C, wherein C is given by: C=λ.sub.i* C.sub.i +λ.sub.j* +λ.sub.i* λ.sub.j* (μ.sub.i* -μ.sub.j*)(μ.sub.i* -μ.sub.j*)'; wherein λ i* is given by: ##EQU18## wherein λ j* is given by: ##EQU19##
Claim 6
Dependent Claim
27 words
The training module of claim 2, wherein a distance d(i,j) between Gaussian density pairs i and j is given by: d(i,j)=(p.sub.i +p.sub.j)λ.sub.i λ.sub.j ∥μ.sub.i -μ.sub.j ∥.sup.2 ##EQU20##
Claim 7
Dependent Claim
16 words
The speech recognition system of claim 5, further comprising: means for smoothing the mixture gaussian densities.
Claim 8
Dependent Claim
27 words
The speech recognition system of claim 7 wherein said means for smoothing the mixture Gaussian densities includes smoothing said mixture Gaussian density by a fully-merged Gaussian density.
Claim 9
Dependent Claim
45 words
The speech recognition system of claim 7 wherein said fully-merged Gaussian density is estimated from all segments in the set by calculating parameters μ and C defining said fully-merged Gaussian density as follows: ##EQU21## wherein N is the number of Gaussian densities in the set.
Claim 10
Dependent Claim
50 words
The speech recognition system of claim 8 wherein said smoothing of a mixture Gaussian density by said fully-merged Gaussian density comprises: means for interpolating a covariance matrix C i of a mixture component by the covariance matrix C of the fully-merged Gaussian density, as follows: C.sub.i =λC.sub.i +(1-λ)C wherein 0≦λ≦1.
Claim 11
Dependent Claim
25 words
The speech recognition system of claim 1, further comprising: means for incorporating a fully-merged Gaussian density as an extra mixture component in the mixture density.
Claim 12
Dependent Claim
18 words
The training module of claim 10, further comprising: means for estimating the weights of the mixture Gaussian densities.
Claim 13
Dependent Claim
106 words
The speech recognition system of claim 11 wherein said means for estimating the weights of said mixture Gaussian densities comprises: means for initially setting the weights of the mixture Gaussian density to a uniform value; means for using the mixture Gaussian density and the uniform weights as models of speech units to segment a chosen training sequence into states; means for calculating the likelihood score of each mixture component on a per-frame basis in the associated labeled segments; and means for averaging each likelihood score over each state of the context environment for each speech unit, said averaged scores becoming the weight of the mixture density.
Claim 14
Dependent Claim
22 words
The speech recognition system of claim 12 further comprising: means for smoothing said weights of the mixture density by said uniform weights.
Claim 15
Dependent Claim
41 words
The speech recognition system of claim 13 further comprising: means for providing the Gaussian mixture densities for each state of each speech unit and the weights of said Gaussian densities to be used as reference models in a speech recognition process.
Claim 16
Dependent Claim
89 words
The speech recognition system of claim 1 wherein said parameters defining a Gaussian density for each segment comprise μ and C, and wherein said means for estimating said parameters comprises: means for calculating μ, wherein μ is given by: ##EQU22## wherein i is a segment index, L i is the number of frames in segment i, and x l .sup.(i) is the lth feature vector in segment i; and means for calculating C, wherein C is given by: ##EQU23## where the symbol ' denotes the transpose of a vector.
Claim 17
Independent Claim
130 words
A method for speech recognition, said speech recognition method comprising the steps of: receiving an input speech signal, segmenting the input speech signal into a plurality of speech units, each speech unit having a plurality of states, each state of each speech unit being a segment of said signal, each segment including a number of frames; identifying said speech units; grouping all speech segments corresponding to like speech unit states for like speech units into respective sets of segments for all states of all speech units; estimating parameters defining a Gaussian density for each segment of each set of segments; merging said Gaussian densities of each set into a mixture Gaussian density for the set; and using a mixture Gaussian densities for identifying speech units of additional input speech signals.
Claim 18
Independent Claim
130 words
A method for speech recognition, said speech recognition method comprising the steps of: receiving an input speech signal, segmenting the input speech signal into a plurality of speech units, each speech unit having a plurality of states, each state of each speech unit being a segment of said signal, each segment including a number of frames; identifying said speech units; grouping all speech segments corresponding to like speech unit states for like speech units into respective sets of segments for all states of all speech units; estimating parameters defining a Gaussian density for each segment of each set of segments; merging said Gaussian densities of each set into a mixture Gaussian density for the set; and using a mixture Gaussian densities for identifying speech units of additional input speech signals.
Claim 19
Dependent Claim
53 words
The method of speech recognition of claim 16, wherein said step of merging comprises: calculating distances between all pairs of Gaussian densities in a set; sorting the calculated distances to find a Gaussian density pair separated by a minimum distance; and merging said minimum distance Gaussian density pair into a single Gaussian density.
Claim 20
Dependent Claim
87 words
The method of speech recognition of claim 17 wherein said minimum distance pair is represented by indices i* and j*, and a said merged Gaussian density is represented by μ and C, and wherein said means for merging said minimum distance Gaussian pair comprises: calculating μ, wherein μ is given by: μ=λ.sub.i* μ.sub.i +λ.sub.j* μ.sub.j* ; calculating C, wherein C is given by: C=λ.sub.i* C.sub.i +λ.sub.j* C.sub.j* +λ.sub.i* λ.sub.j* (μ.sub.i* -μ.sub.j*)(μ.sub.i* -μ.sub.j*)'; wherein λ i* is given by ##EQU24## wherein λ j * is given by: ##EQU25##
Claim 21
Dependent Claim
32 words
The method of speech recognition of claim 17 further comprising the steps of: iteratively carrying out a merging procedure until the distance between the minimum distance pair is above a predetermined threshold.
Claim 22
Dependent Claim
31 words
The method of speech recognition of claim 17, wherein a distance d(i,j) between Gaussian density pairs i and j is given by: d(i,j) =(p.sub.i +p.sub.j)λ.sub.i λ.sub.j ∥μ.sub.i -μ.sub.j ∥.sup.2 where ##EQU26##
Claim 23
Dependent Claim
27 words
The method of speech recognition of claim 16, further comprising the step of: incorporating a fully-merged Gaussian density as an extra mixture component in the mixture density.
Claim 24
Dependent Claim
21 words
The method of speech recognition of claim 21, further comprising the step of: estimating the weights of the mixture Gaussian densities.
Claim 25
Dependent Claim
102 words
The method of speech recognition of claim 22 wherein said step of estimating the weights of said mixture Gaussian densities comprises: initially setting the weights of the mixture Gaussian density to a uniform value; using the mixture Gaussian densities and the uniform weights as models of speech units to segment a chosen training sequence into states; calculating the likelihood score of each mixture component on a per-frame basis in the associated labeled segments; and averaging likelihood scores of each mixture component over each state of the context environment for each speech unit, said averaged scores becoming the weights of the mixture density.
Claim 26
Dependent Claim
24 words
The method of speech recognition of claim 23 further comprising the step of: smoothing said weights of the mixture density by said uniform weights.
Claim 27
Dependent Claim
40 words
The method of speech recognition of claim 24 further comprising: providing the Gaussian mixture densities for all states of all speech units and the weights of said Gaussian densities to be used as reference models in a speech recognition process.
Claim 28
Dependent Claim
89 words
The method of speech recognition of claim 16 wherein said set of parameters defining a Gaussian density comprise μ and C, and wherein said step of estimating said parameters comprises: calculating μ, wherein μ is a vector given by: ##EQU27## wherein i is a segment index, L i is the number of frames in segment i, and x l .sup.(i) is the lth feature vector in segment i; and calculating C, wherein C is a matrix given by: ##EQU28## wherein the symbol ' denotes the transpose of a vector.
Claim 29
Dependent Claim
18 words
The method of speech recognition of claim 26, further comprising the step of: smoothing the mixture Gaussian densities.
Claim 30
Dependent Claim
28 words
The method of speech recognition of claim 27 wherein said step of smoothing the mixture Gaussian densities includes smoothing said mixture Gaussian density by a fully-merged Gaussian density.
Claim 31
Dependent Claim
57 words
The method of speech recognition of claim 28 wherein said fully-merged Gaussian density is estimated from all segments in the set by calculating parameters μ and C defining said fully-merged Gaussian density as follows: ##EQU29## wherein N is the number of Gaussian densities in th set and wherein the symbol ' denotes the transpose of a vector.
Claim 32
Dependent Claim
55 words
The method of speech recognition of claim 29 wherein said step of smoothing of a mixture Gaussian density by said fully-merged Gaussian density comprises the step of: interpolating the covariance matrix C i of a mixture component by the covariance matrix C of the fully-merged Gaussian density, as follows: C.sub.i =λC.sub.i +(1-λ) C, wherein 0≦λ≦1.
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) | 1996-03-08 | 1996-09-07 | 1997-03-09 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2000-03-08 | 2000-09-07 | 2001-03-09 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2004-03-08 | 2004-09-07 | 2005-03-09 | $7700.00 | Window Closed / Lapsed |