A vast amount of research is going on for design of robust speech recognition in to alleviate speech variability conditions. One of the variability aspects is the difference between telephony speech and direct speech (recorded in noise free conditions). In this paper by More
A vast amount of research is going on for design of robust speech recognition in to alleviate speech variability conditions. One of the variability aspects is the difference between telephony speech and direct speech (recorded in noise free conditions). In this paper by using a set of experiments, it is shown that LHCB parameters are superior to traditional MFCCs for speech recognition applications when they are used in a neural network based speech recognition system for both direct and telephony speech. Then by extraction of LHCBs from direct and telephony speech, and training of a MLP based speech recognition model, a direct and telephony speech recognition system is developed. Using a neural network inversion based on gradient descent method, the telephony speech feature vectors are modified toward to the direct speech feature vectors and by training a second network on modified telephony and direct speech feature vectors a 1.4% enhancement on speech recognition was achieved. Later, using general inversion method of neural networks both telephony and direct speech feature vectors are modified in a manner which mainly contains phonetic information and not other speech variations. Then by the training of the second neural network on this dataset, the system achieved 2.98% and 1.68% higher recognition rate for direct and telephony speech, respectively.
Manuscript profile
