Efficient implementation of a frequency‐domain technique for frequency scaling of speech signals

Abstract

Frequency scaling of speech signals is a useful technique for reducing bandwidth requirements in analog and digital speech transmission systems. A recently developed frequency domain scaling (FDS) technique is based on scaling the phase of the signal's short‐time Fourier transform (STFT), evaluated at a discrete set of equally spaced frequencies. For the particular scale factor of 2, a simple sign tracking algorithm is used which avoids explicit phase computation and unwrapping. A computationally efficient implementation of this technique is presented. It is based on performing the STFT analysis, modification and synthesis by means of the FFT algorithm, using specially chosen analysis and synthesis window functions. In addition, efficient decimation and interpolation of the complex bandpass signals, along with the required change in output sampling rate are embedded in the implementation scheme. Simulation results show that this FDS technique is particularly robust to environmental conditions such as background noise, room reverberation, and multiple speakers. Processed examples will be played along with results obtained with the earlier reported pitch‐adaptive, time‐domain harmonic scaling (TDHS) technique. A hybrid scheme, which combines the two techniques will also be discussed and demonstrated.