
Dear interested person,

It is my "messy" Matlab+Kaldi code for uncertainty propagation on DNN+HMM:

Karan Nathwani, Juan A. Morales-Cordovilla, Sunit Sivasankaran, Irina Illina and Emmanuel Vincent. "An extended experimental investigation of DNN uncertainty propagation for noise robust ASR". Hands-free Speech Communication and Microphone Arrays (HSCMA) . ISSN: 978-1-5090-5925-6. San Francisco, March, 2017.


The most important files are:

MatlabFE/JuanFE/Recognizer/AllRec.mat and /Fun/RecDir.mat: they start the Feature Extraction for Training and Test (FETrTeDir), the Kaldi HMM Training (TrRec), the Test or HMM speech recognition with uncertainty (FETrTeDir) and the final ASR Word-Error-Rate performance (RecPerfDirKald).

MatlabFE/JuanFE/UncFE/DnnUncert/Unc2LikGramDnn.mat: that takes a Kaldi featuregram (Fmllr_time matrix representation of speech) plus the uncertaintygram and produces a Kaldi likelygram (hmmStateLikelihood_time representation) based on difference propagations (Monte Carlo, Entire,...).

You need to explore the code and modify its paths until it runs. Unfortunately I am now on Bioinformatics and I do not have time to "clean" it.
Pleas, ask me any doubt.

Best,

Juan A. Morales-Cordovilla
jamc@ugr.es




