Missing feature theory

1
Missing feature theory

• Statistical estimation of unreliable features for
  robust speech
• recognition
• 2) Missing feature theory and probabilistic
  estimation of clean
• speech components for robust speech
  recognition
• 3) State based imputation of missing data for
  robust speech
• recognition and speech enhancement
• 4) Missing data theory,spectral subtraction and
  signal-to-noise
• estimation for robust ASR an integrated
  study

2
Introduction
Parameters used in speech recognition can be
divided in two subsets

1) reliable or present parameters 2)
unreliable or missing parameters
3
Introduction
There are 2 problems in the application of
missing data in robust ASR 1) identification of
the reliable regions 2) recognition techniques
that can deal with incomplete data
4
Detection of unreliable feature
Method
(1) negative energy criterion
(2) SNR criterion
or
5
Detection of unreliable feature
(3) statistical approach noise is considered as
normally distributed
6
Noise estimation method in 4

1. simple estimation
2. weighted average estimation

C) second order method
D) Histogram method
7
Accuracy for the three detection methods
8
Recognition with incomplete data
Method (1) Marginalization unreliable data are
ignored for a single state model ,the
probability to emit vector is
9
Marginalization
10
Marginalization
1
bounded marginalization
11
Marginalization
In Philippes another paper 2 ,the clean
parameters are represented as pdfs and missing
parameters are considered as being uniformly
if 0ltxltY(w)
otherwise
12
Recognition with incomplete data
Method (2) GMM based Imputation unreliable
data are estimated advantages of the approach are
that can be followed by conventional techniques
like cepstral,RASTA
In the estimation process,the GMM means are used
to replace the unreliable features
the means and variances of GMM are
compensated with the additive noise,as in PMC
13
Imputation
using inverse log-normal approximation
14
Imputation
transformed into log-spectral domain
15
Imputation
using the noisy GMM,the weighting factor
associated with each distribution is computed as
follows
16
Imputation
Finally,the reliable data are enhanced using
a spectral subtraction and the unreliable data
are replaced by a weighted sum of the GMM means
17
features spetra
18
Discussion
Why using GMM in this paper? A HMM based data
imputation has been proposed in 3, when using
time-dependent statistical models,if an error in
the decoding sequence occurs,it can influence
the recognition in the second feature
domain therefore , GMM instead of HMM,but
sufficient and computationally efficient for data
imputation
19
Experiments results
20
Experiments results
21
Experiments results
22
Experiments results
23
Experiments results
24
Experiments results
25
Experiments results
26
Experiments results