A Joint Enhancement-Decoding Formulation for Noise Robust Phoneme Recognition
P. M. Nazreen, A. G. Ramakrishnan, Prasanta Kumar Ghosh · 2017
We consider a dictionary based speech enhancement in the context of automatic recognition of noisy speech. Speech in each analysis frame is denoised as a front-end processing using a class-specific (e.g. phoneme) dictionary selected based on the estimated class label. However, when the estimated label is erroneous, a wrong class model is chosen for many frames. We propose a Joint Enhancement-Decoding (JED) algorithm to overcome this issue by jointly optimizing for labels of all the frames and the decoding path. The algorithm optimizes over multiple enhanced versions of each frame using different phoneme specific dictionaries and gives the maximum likelihood path of state sequences as well as the best (in the maximum likelihood sense) choice of the enhanced observation sequence as its output. The number of phoneme-specific dictionaries used for enhancement in an analysis frame is varied from 1 to 5 based on the phoneme confusion matrix and the recognition results are reported for each case. Experiments with TIMIT corpus and five different noises at 0, 5 and 10 dB SNRs show that the recognition performance varies with the number of dictionaries, and in most of the cases, is the best when two or three dictionaries are employed.