Abstract
In this study, we propose a single microphone speech enhancement algorithm by combining harmonic structure estimation and traditional MMSE speech enhancement for a leveraged overall solution. Traditional single channel speech enhancement methods are usually based on the statistic characteristics of noise signals which are effective only for stationary noise, but not for non-stationary noise. In our study, we attempt to estimate noise by exploring the harmonic structure of the target speech combined with temporal noise tracking. In voiced segments, since speech energy is sparsely carried by harmonic partials, the spectrum located between adjacent harmonic partials are considered as noise. We assume that the speech spectrum distributes continuously along the frequency-dimension. Thus, the noise overlapped with speech harmonics can be estimated with an interpolation technique. Next, the estimated noise is incorporated into a traditional MMSE framework for speech enhancement. A listening test is carried out with 6 cochlear implant recipients to evaluate the proposed speech enhancement algorithm. The experimental results show that the proposed algorithm is able to improve the speech intelligibility in terms of word recognition rate for CI listeners.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call