A High-Performance Auditory Feature for Automatic Speech Recognition

Feature extraction is the first crucial block in any automatic speech recognition system. To achieve better and more robust ASR performances, especially in adverse acoustic environments, a new feature extraction algorithm is desirable. Human auditory system converts speech signals to nerve firing rates in different frequency bands for auditory cognition in the brain. It consists of the following modules: outer ear, middle ear, cochlea, hair cells, and nerve system. In this study, we model each of modules in its function of information and signal processing to construct a new feature extraction algorithm, which is fast enough for a real-time AST in a software implementation.