This page contains LDM phonetic speech recognition experiment to evaluate the potential of LDM for speech classification.

Setup:
1) 13-dimensional Y[] and 13-dimensional X[]
2) features: 12MFCC + Energy
3) database: 2-speaker [tm] and [ss]. 3-sound /aa/, /m/, and /sh/ with 50 examples for each sound.

Experiment steps:
(1) train model /aa/ using 70 examples: tm_aa1 ... tm_aa35, ss_aa1 ... ss_aa35.
(2) train model /m/ using 70 examples: tm_m1 ... tm_m35, ss_m1 ... ss_m35.
(3) train model /sh/ using 70 examples: tm_sh1 ... tm_sh35, ss_sh1 ... ss_sh35.
(4) test sound /aa/ using 30 examples: tm_aa36 ... tm_aa50, ss_aa36 ... ss_aa50.
(5) test sound /m/ using 30 examples: tm_m36 ... tm_m50, ss_m36 ... ss_m50.
(6) test sound /sh/ using 30 examples: tm_sh36 ... tm_sh50, ss_sh36 ... ss_sh50.

EM training 50 iteration /aa/

EM training 50 iteration /m/

EM training 50 iteration /sh/

Confusion Matrix /aa/

Confusion Matrix /m/

Confusion Matrix /sh/

Result and Analysis

• As Dr. Picone suggested, we increaseded our database to 50 examples per sound and all the testing examples are recognized correctly. The recognition correct rate is 100%.
• This experiment demenstrates that LDM can be a good pattern classifier and has very good potential for speech classification.
• We are doing speech recognition experiment again with the same database using HMM model. It will work as our sustained-phone baseline.