Frame erasure concealment technique for a bitstream-based feature extractor
View Patent ↗A frame erasure concealment technique for a bitstream-based feature extractor in a speech recognition system particularly suited for use in a wireless communication system operates to “delete” each frame in which an erasure is declared. The deletions thus reduce the length of the observation sequence, but have been found to provide for sufficient speech recognition based on both single word and “string” tests of the deletion technique.
1. A method of generating speech coding parameters of an erased frame in a bitstream-based front end of a speech recognition system, the method comprising the steps of:
detecting an erased frame;
computing the probability of a correct observation sequence from the group O=(O 0 ,O m ), where the probability is defined as:
P
(
O
c
|
λ
)
=
∑
q
1
,
…
,
q
N
π
q
1
b
q
1
(
o
1
)
a
q
1
q
2
b
q
2
(
o
2
)
…
a
q
1
-
Nq
1
a
q
1
q
1
+
n
b
q
1
+
1
(
o
l
+
Γ
)
…
a
qN
-
1
qN
b
qN
(
o
N
)
deleting an observation vector of the erased frame from the observation sequence; and
decoding the observation sequence with a speech decoder using a standard hidden Markov model process.
2. The method as defined in claim 1 wherein in performing the detection, the following steps are performed:
measuring the Euclidean distance between the line spectrum pairs of contiguous frames;
defining a steady-state threshold T; and
deleting one frame of the contiguous frames when the Euclidean distance is less than the threshold.
3. The method as defined in claim 1 wherein in detecting a frame erasure, an erasure is declared when the bits most sensitive to error within a frame are determined to be in error.
4. The method as defined in claim 3 wherein the bits most sensitive to error in a frame in a bitstream-based speech recognition system include the line spectrum pair information bits and the gain information bits.
5. A system of generating speech coding parameters of an erased frame in a bitstream-based front end of a speech recognition system comprising:
means for detecting an erased frame;
means for computing the probability of a correct observation sequence;
means for deleting an observation vector of the erased frame from the observation sequence; and
means for decoding the observation sequence with a speech decoder using a standard hidden Markov model process.
6. The system of claim 5 wherein the correct observation sequence comprises the group O=(O 0 ,O m ), and the probability of a correct observation sequence is defined as:
P
(
O
c
❘
λ
)
=
∑
q
1
,
…
,
γ
N
π
q
1
b
q
1
(
o
1
)
a
q
1
q
2
b
q
2
(
o
2
)
…
a
q
1
-
Nq
1
a
q
1
q
1
+
n
b
q
1
+
1
(
o
l
+
Γ
)
…
a
qN
-
1
aN
b
qN
(
o
N
)
.
7. The system of claim 5 wherein the means for detecting an erased frame further comprises:
means for measuring the Euclidean distance between the line spectrum pairs of contiguous frames;
means for defining a steady-state threshold T; and
means for deleting one frame of the contiguous frames when the Euclidean distance is less than the threshold.
8. The system of claim 5 wherein the means for detecting an erased frame further comprises means for declaring an erasure when the bits most sensitive to error within a frame are determined to be in error.
9. The system of claim 8 wherein the bits most sensitive to error in a frame in a bitstream-based speech recognition system include the line spectrum pair information bits and the gain information bits.
10. A system of generating speech coding parameters of an erased frame in a bitstream-based front end of a speech recognition system comprising:
means for detecting an erased frame;
means for measuring the Euclidean distance between the line spectrum pairs of adjacent frames (n−1) and n;
means for defining a steady-state threshold T associated with an acceptable difference between the line spectrum pairs of the adjacent frames;
means for deleting the line spectrum pairs of the n th frame in an observation sequence if the measured distance is less than or equal to T; and
means for generating the speech coding parameters with a standard hidden Markov model process.
11. The system of claim 10 wherein the means for detecting an erased frame further comprises means for declaring an erasure when the bits most sensitive to error within a frame are determined to be in error.
12. The system of claim 11 wherein the bits most sensitive to error in a frame in a bitstream-based speech recognition system include the line spectrum pair information bits and the gain information bits.