A LOW DELAY, VARIABLE RESOLUTION, PERFECT RECONSTRUCTIONSPECTRAL ANALYSIS-SYNTHESIS SYSTEM FOR SPEECH ENHANCEMENT

摘要

The choice of the window function and window length inshort time analysis-synthesis (AS) systems based on the discreteFourier transform (DFT) has to balance conflictingrequirements: Long windows provide high spectral resolutionwhile short windows allow for high temporal resolution.Furthermore, for many applications a low algorithmic delayis desirable. Therefore, long standard windows such asthe Hann or Hamming windows cannot be used. In this contributionwe propose a novel AS system which achieves perfectreconstruction (PR) and a low delay by using a variablelength analysis window and a relatively short synthesis window.The variable length analysis windows allow a spectralanalysis that is adapted to the signals span of stationarity.The AS windows are designed such that they can be switchedat any time instant without violating PR. We show that thespectral representation of typical speech data is improved ascompared to AS systems with standard windows.