ASR for mixed speech using SNMF based separation algorithm

Combined experiment for separation and recognition of the mixed speech signal of two speakers using different features are reported in this work. For separation of speech signals, sparse NMF is used after wavelet decomposition of mixed speech signal. Kaldi toolkit with different features is used for speech recognition having 4-gram language model. For experiment purpose, phonetically balanced 1000 Hindi sentences from AMUAV corpus are used. Simulation is done for different target to mixed speech mixing level. Results show that recognition of speech signals after separation using SNMF based algorithm perform better as compared to the recognition of mixed speech when the target to mixed signal ratio is lower than 20dB. It is also observed that Δ + Δ - Δ feature-based speech recognition model perform better as compare of other models.