Abstract: This paper addresses the tak of automatic speech recognition (ASR) in adverse condition, in particular from within a domestic car. Variations of non-linear spectral subtraction (NLSS) are assessed in terms of ASR relevant spectral matching in an attempt to find optimum conditions invariant across typical car noises. Lockwood reports the benefits of an SNR-dependent scaler in subtracting over-estimates of the noise spectrum. In this paper the idea is extended to frequency-dependent noise floors. Frequency-dependent noise floors, reflecting speech spectral distributions are proposed. With a simple linear approximation these are shown to offer small but finite improvements at no extra computational cost.

Download: