Sadegh Masjoodi
Mansour Vali

Despite one decade of the missing feature theory application in the domain of Robust Automatic SpeechRecognition (ASR), this field is still an active area for researchers. In this report using fuzzy concepts, we will present a method for modifying the cluster-based reconstruction of unreliable components of the noisy speech spectrogram. In this simple but effective method using a fuzzy membership function the feature vector component reliability is fuzzified. In the next stage this new parameter is applied as a weighting parameter for summing new reconstructed components and their old noisy values. Experiments were done on the FarsDat database usingtwo recognition models, a Neural Network (NN) and a Hidden Markova Model (HMM). The improvements in the recognition results using this new reconstruction method in low SNRs for the frame-based neural network was approximately 5% and for the phoneme-based HMM was between one and two percent.