Synergy of spectral and ear model features for neural speech recognition

Hybrid speech recognition systems combining hidden Markov models modeling of speech with neural network ability to discriminate patterns by computing posterior probabilities have reached very good performances. A promising direction to further improve performances is the use of a greater amount of input information. In doing this neural network based models are better than hidden Markov models as they are not constrained by the condition of stochastic independence of its input features and can freely employ heterogeneous input sources. In the present work we explore the integration of standard speech recognition parameters with input features extracted by an ear model. The obtained results demonstrate that the synergy of multiple input sources is a viable line to increase speech recognition accuracy for neural classifiers