A Modified LIMA Framework for Spectral Subtraction Applied to In-Car Speech Recognition

In noisy environments, speech recognition accuracy degrades significantly. Speech enhancement algorithms have been designed to overcome this, however solutions to date have not been optimal for speech recognition especially for non-stationary noise like that in a car. Recently, a likelihood-maximising (LIMA) criteria has been applied to speech enhancement techniques. This paper analyses the suitability of spectral subtraction for potential use under a modified version of this framework where direct access to and manipulation of speech recognition models is not available. Analysis shows spectral subtraction is suited to this holistic LIMA approach by confirming the cost surface is appropriate for gradient descent methods. It is also observed that there are regions on the cost surface where performance exceeds that achieved by parameter values traditionally selected for spectral subtraction.