On the fundamental limitations of spectral subtraction: An assessment by automatic speech recognition

Spectral subtraction is one of the earliest and longest standing, popular approaches to noise compensation and speech enhancement. A literature search reveals an abundance of recent research papers that report the successful application of spectral subtraction to noise robust automatic speech recognition (ASR). However, as with many alternative approaches, the benefits lessen as noise levels in the order of 0 dB are approached and exceeded. Previously published works relating to spectral subtraction provide a theoretical analysis of error sources. Recently the first empirical assessment showed that these fundamental limitations can lead to significant degardations in ASR performance. Results illustrate that under particularly high noise conditions these degradations are comparable to those caused by errors in the noise estimate which are widely believed to have by far the greatest influence on spectral subtraction performance. The original contribution made in this paper is the assessment of the fundamental limitations of a practical implementation of spectral subtraction under the European standard ETSI Aurora 2 experimental protocols. Results illustrate that, perhaps contrary to popular belief, as noise levels in the order of 0 dB are approached phase and cross-term error sources do indeed contribute non-negligible degradations to ASR performance. This is believed to be a new observation in the context of spectral subtraction and ASR.