Speech enhancement by combination of transient emphasis and noise cancelation

Abstract

This paper evaluates the effectiveness of combining speech modification techniques that enhance transition components with active noise cancellation to improve the intelligibility of speech in noise. Two speech modification techniques were considered. One is based on wavelet-packet analysis, and the second uses a fixed filter, derived from timefrequency analysis, that emphasizes high frequencies.Active noise cancellation was provided by Bose noisecancelling headphones. The test noise was real, generated by a ground auxiliary generator on the tarmac at an Air National Guard facility. The test signals were speech tokens from the modified rhyme (psycho-acoustic) test, recorded by a male speaker. This paradigm was used to measure word recognition rates at various signal-to-noise ratios. Active noise cancellation by itself provided over 40% increase in word recognition, while the modified speech and fixed filter techniques alone provided up to 20% improvement, depending on the signal-to-noise ratio. In combination, the speech modification approaches provided over 15% additional improvement in intelligibility over noise cancellation alone.

The listening talker: A review of human and algorithmic context-induced modifications of speech

“…6.2.7. Boost transients [247] found transient-boosting methods can improve intelligibility of speech in noise (recordings of an aircraft auxiliary power unit) presented at various SNRs from -30 dB to -10 dB, and that gains were still seen when combining this with active noise-cancelling headphones. Although the best-performing method was inspired by a dynamic timevarying filter approach, in order to obtain real-time performance this was approximated as a simple fixed filter, which amounts to attenuation below 700Hz and mid-and high-frequency boosting.…”

“…Distortions introduced by that method may have counteracted the benefits of transient boosting. From the results available in [247] and from simple spectral shaping, it is not possible to separate out the effects of a general high-frequency boost from specific localised transient boosting. [322] also found that increasing the intensity of transients relative to steady-state regions increased intelligibility.…”

Speech Enhancement System Based on Lyapunov FXLMS Algorithm

“…In speech enhancement systems (SES), a similar approach is active noise cancellation, in which the noise signal is directly measured and then removed form the noisy speech signal. Algorithm noise cancellation requires a separate microphone to record the noise alone, it is successfully implemented to improve speech intelligibility and is now widely availably commercially [1]- [3]. A conventional approach in SES system is to use FIR filters as adaptive controllers where the filter coefficients are adjusted according to the least mean square (LMS) algorithm to minimize errors between output of the filters and the desired response.…”