A self adaptive filter of environmental noises from speech and its evaluation

This paper is concerned with the design principles, testing and evaluation of a self-adaptive filter to filter environmental noises from speech when noise and speech parameters may arbitrarily vary (as they do) and are a-priori unknown and when no access is available to noise alone or to speech alone, but only the sum of speech plus noise is available. The above situations are typical to what exists in hearing aids. The study considers two concrete realizations, one for a filter that has now been incorporated in single CMOS chip form in behind ear hearing aids and involves considerable design short cuts. The other concerns a bread-board realization not yet reducible to satisfy actual hearing aids, but a version of which should be reducible to fit hearing aids in a few years. Performance tests on improvement via filtering in signal-to-noises-ratio (SNR) and in intelligibility (for mono-syllable standard word lists) in a wide range of noises, all point out the superiority of the frequency-domain design, whose real-time on-line realization is also the simplest. Indeed, performance tests, as presented in the paper, show SNR improvements via the frequency-domain design from 6.7 to 25.5 dB for a multitude of noises throughout the spectrum, and where the input SNR (prior to filtering) varies from-5 to-20 dB. The 10dB or higher improvement was the more common one, even in cafeteria/cocktail situations. Intelligibility-scores were equally impressive, as is tabulated, and reached, in cafeteria situations, an improvement from a 32% correct score prior to filtering to a 90% score after frequency-domain filtering, the average improvement being 21% (for single-syllable words). Listening comfort, though not quantitatively measurable, was most impressive, this being a major aspect in fatigue when listening in a noisy environment. The performance of the resulting filter thus points to providing considerable intelligibility, SNR and listener-comfort improvements under the widest pos- ible noise situations, when noise is apriori unknown and inaccessible.