1 Introduction
-
Automatically handle the issue of speech length variability
-
Conserve processing time on the mobile device
2 Feature extraction
2.1 Experimentation data
3 Automatic speech recognition classifiers
3.1 Standard Multi-layer perceptron
3.2 Proposed Dynamic Multi-layer Perceptron implementation
3.3 Dynamic MLP experimentation
Dynamic MLP algorithm forward Pass
|
---|
For every D with F
N
frames
|
For the N
th
F of D
|
Arrange elements in ascending numbers of IU
|
End
|
Pass IU to X
|
Propagate X to H using only connected W
in
|
For every h
n
ϵ H
|
End
|
Propagate Net1
i
to O using W
out
|
For every o
n
ϵ O
|
End
|
End
|
3.4 Hidden Markov model implementation
4 Results
4.1 Dynamic Multi-layer Perceptron results
4.2 HMM results
4.3 Memory and time comparison between DMLP and HMM
Device | Processor | RAM | HMM 4 | HMM 8 | HMM 12 | DMLP 50 | DMLP 100 | DMLP 200 | DMLP 300 |
---|---|---|---|---|---|---|---|---|---|
Galaxy S3 | 1.4 GHz (quad-core) | 1 GB | 66.1204 | 139.3366 | 201.893 | 0.5664 | 1.2288 | 2.3012 | 3.7752 |
Galaxy S4 | 1.9 GHz (quad-core) | 2 GB | 19.9746 | 38.8962 | 59.53 | 0.3326 | 0.6814 | 1.989 | 5.1172 |
Galaxy S5 | 2.5 GHz (quad-core) | 2 GB | 12.895 | 26.1608 | 39.593 | 0.1512 | 0.3376 | 0.6988 | 1.3744 |
Note 2 | 1.6 GHz (quad-core) | 2 GB | 84.6602 | 106.9112 | 221.2478 | 0.3426 | 0.7134 | 1.4506 | 2.3222 |
Note 3 | 2.3 GHz (quad-core) | 3 GB | 12.1884 | 26.7236 | 35.592 | 0.207 | 0.441 | 1.2524 | 2.3436 |
HMM 4 | HMM 8 | HMM 12 | DMLP 50 | DMLP 100 | DMLP 200 | DMLP 300 |
---|---|---|---|---|---|---|
752 KB | 805 KB | 861 KB | 3.25 MB | 4.25 MB | 6.28 MB | 8.34 MB |