Digital Signal Processing Reference
In-Depth Information
ψ
⎯ →
X
priority
Oo o
=
{
,
,...,
o
}
AVLI
12
p
O
ψ
Where
is neurons with different priority number, X is ALI samples.
priority
AVLI
is the output of PONN.
o
is the output of neural network for special language
p
.
For a period Voice, we will split it into some little pieces voice and transform it in-
to eigenvector in feature space. The ultimate identify formula is written as
OOOO
=
{
,
,...,
}
AVLI
AVLI
(1)
AVLI
( 2)
AVLI k
(
)
O
is the
i
piece time quantum. The last result is written as
Where
AVLI i
()
max{
count i O
( ) |
=
O
,
i
≠
j
}
AVLI i
()
AVLI
( )
j
count i
()
Where
is a function of taking sum number of same class in distin-
guished sequence.
4
Experience Results
Our experiment uses data of Chinese, English and Japanese language voice. The voice
is partitioned into same time pieces and been translated into MFCC features [7]. The
experiment is performed using methods of PONN and SVM in different dimension
feature space [8]. The samples number is 223, 474, and 107 respectively in Table 1.
Table 2 is the comparative results.
Table 1.
Training Number of different language
Training1 Training2 Training3 Training4 Training5
Time
0.512s
1.008s
1.504s
2.000s
2.496s
Training
Number
of sam-
ples
Chinese
5301
2323
1325
827
567
English
9452
4254
2449
1611
1119
Japanese
2126
1007
623
427
315
Total
16879
7584
4397
2865
2001
Table 2.
Comparative results of using PONN and SVM models
Dimension Size
496
992
1488
1984
2480
Chinese
94.170
91.031
91.928
89.238
77.130
English
88.608
84.177
84.177
76.160
86.709
PONN
Japanese
98.131
95.327
88.785
80.374
83.178
AVG
91.418
87.562
86.940
80.348
83.831
