The objective of this phase is to build a strong system for Kaggle ASR challenge.
Evolution of the System
15 Dec 2017
Built a baseline using Encoder Decoder framework with bidirectional LSTMs.
Features - MFCC(7 dimensional + delta + delta-delta )
Training and Validation Errors:
Built a baseline using Encoder Decoder framework with CNN as encoder and DNN as decoder.
Features - Magnitude and Phase spectrum from Short Term Fourier Transform.
Training and Validation Errors:
16 Dec 2017
No expt beats ConvEnc + DNN Dec. Need to experiment with different features. filterbanks ?
17 Dec 2017
No expt beats ConvEnc + DNN Dec.
Confusion Matrix shows highest confusion between the classes 'on' and 'off'.
Need to build tools to visualize better.
18 Dec 2017
Still no update
19 Dec 2017
No update. The clusters dont seem to have enough memory to get test output. So copying everything onto GPU. This is obviously a stupid thing to do.
20 Dec 2017
Wrapped this phase up. Waiting for Phase 02 to start in January.