DNN Toolkit: LIBDNN

https://github.com/botonchou/libdnn

Let's take a DNN structured:

69-1024-1024-1024-48

as an example ...

run.sh: design

Begin Training ...

Display the result

We didn't plot the error bar since those values are relatively small.

Project #1

Feature Extraction

1. add-deltas

2. compute-cmvn-stats

3. apply-cmvn

MFCC Extraction

Result 1: Training set MFCC

Result 2: Testing set MFCC

Project #2

HMM Training

Work Flow

1. MFCC? Already done in proj #1 !

2. Train mono-phone model -- 03.mono.train.sh

    We train two mono-phone models, one with      

    numiters=40 (default by TA), and the other with

    numbers=60 (extend training). The rest of the

    parameters are fixed.

mono_iters40

mono_iters60

We continue training with both models to see their differences.

output model: mono

3. Build decision tree for both mono-phone models

    -- 05.tree-build.sh

    => tree_iters40 & tree_iters60

4. Train tri-phone models from tree models trained

    previously -- 06.tri.train.sh

tri_iters40

tri_iters60

output model: tri-phone

Project #3

LM training, WFST decoding, Viterbi decoding

Work Flow

mono_iters40_*

WFST

Viterbi

tri_iters40_*

WFST

Viterbi

Observation:

Decoding time of *_WFSTs are too long, and the accuracies are not higher than *_viterbis !

mono_iters60_viterbi

Observation:

Viterbi outperforms WFST, while increasing iterations of HMM training does not result in better performance, and even worser ...