Enhanced Algoriths for Cockpit Voice Recognition Systems

In order to become a feasible human-interface technology suitable for operational deployment in high-noise environments, real-time speech recognition requires robust algorithms. Robust speech processing would function as an intuitive interface to relieve crew work load and enhance crew safety, particularly in hands-busy and high-noise environments, such as cockpits. We are currently developing an analog complementary metal-oxide-silicon (CMOS) silicon cochlea under an AF Phase II SBIR contract (AF41624-94-C-6004) specifically for noise reduction in cockpit speech-recognition applications. During Phase I, we propose to merge traditional techniques with biologically inspired neural techniques. The traditional technique uses a mel-cepstral coefficient (MCC) technique, with a vector quantizer (VQ) and a phoneme-based hidden Markov model (HMM). The biologically inspired technique uses a noise-insensitive cochlear model, with a trainable, time-invariant learning vector quantizer (LVQ). In the third algorithm, the traditional MCC front end passes its output to the LVQ classifier; for the fourth algorithm the cochlear model use the VQ codebook. All four algorithms send their outputs to word-based HMMs. In this way, we exploit the strengths of the front ends, as well as the strengths of the classifiers. Word-based predictions from all four classifiers are fed to a neural classifier that selects the appropriate word spoken.