High performance telephony speech recognition via cascade HMM/ANN hybrid
Iman Gholampour, K. Nayebi · 2003
A new formulation for discriminative training of HMMs is introduced as a solution to telephony speech recognition problem. This formulation uses a properly trained MLP in a simple interconnection with HMMs called "cascade HMM/ANN hybrid". Our training algorithm has a simple realization in comparison with other discriminative training for HMMs such as MDI and MMI. We also present a rigid mathematical proof of its convergence. We found that using cascade HMM/ANN for telephony isolated word recognition results in increasing the recognition accuracy from 88.1% in classic HMMs to 98.1% using a two layer multilayer perceptron (MLP). This structure also reveals better robustness to ending point positions of the words in the presence of background noise, particularly for mobile telephone calls. No significant increase in computational requirements is needed in the recognition phase and the recognition task can still be performed in real-time. Both theoretical and experimental results are included in the paper.