14 Generalized Theory of Spectral Refinement and Application to Speech Enhancement for In-Car Communication Systems

Abstract

Speech enhancement algorithms usually operate on subband representations of the signals. These are obtained by transforming overlapped, segmented (windowed) time frames of the signals into the frequency domain using filterbanks. The parameters for the subband transformation are generally selected to offer a good trade-off between frequency resolution and output latency. Such a compromise is usually detrimental to the performance of algorithms exploiting the fundamental frequency (F0) of speech, especially when F0 is low, as closely separated harmonics occurring in voiced segments cannot be easily separated. Especially for voiced speech enhancement for in-car communication (ICC), where the ambient noise is pronounced at the low frequencies, this is critical. A reliable detection of voiced/unvoiced segments and an accurate estimation of the fundamental frequency is of paramount importance, which necessitates an increased spectral resolution. Spectral refinement (SR) offers a computationally inexpensive means of generating a refined (higher resolution) signal spectrum by linearly combining the spectra of shorter, contiguous signal segments. In this chapter, a generalized solution for the SR of signals is presented. The method can be applied to refine either the frequencies of the short-term spectrum (or a subset thereof) while maintaining the frequency resolution or to refine the whole frequency range, introducing additional frequency supporting points. Last, some practical points are addressed when applying this to F0 estimation, such as how to address the proper detection of onsets by transitioning between longand short-time windows for the spectral analysis.