Frequency Scaling of Speech Signals by Transform Techniques

Abstract

The general framework of short‐time Fourier analysis, modification, and synthesis is used to describe in a unified way several known techniques for frequency scaling of speech signals. Subsequently, a frequency domain harmonic scaling technique is studied in detail with emphasis on improving its performance and its implementation efficiency. This technique is particularly attractive for 2:1 scaling by use of a sign tracking algorithm which avoids the need for explicit phase computation and unwrapping. The implementation efficiency is achieved by using the fast Fourier transform algorithm, embedded decimation and interpolation, and an extended version of a recently developed weighted overlap‐add synthesis scheme. The improvement in quality is achieved by improved sign tracking and elaborate design and selection of the analysis and synthesis prototype filters (data windows). Results of computer simulations, for a variety of adverse acoustical environment conditions, indicate that the system is highly robust but its quality for clean speech is lower than with a time domain harmonic scaling technique which uses pitch information. In applications which do not permit pitch transmission, a hybrid scheme which combines the two techniques is found to yield a better quality than either system alone.

Hybrid frequency-domain coding of speech signals

“…The spectral fine structure is dependent mainly on the-pitch period and the degree of voiced periodicity. Various techniques to exploit this quasi-periodic feature of voiced speech for speech coding have been investigated, a notable example being the Harmonic compression technique [7][8][9][10].…”

“…The harmonic compression (or expansion) can be realized either in the time domain by selectively eliminating (or inserting) pitch periods [8] or in the frequency domain by reducing (or increasing) the spectral gap between the pitch harmonics [9,10].…”

An algebraic approach to discrete short-time Fourier transform analysis and synthesis

“…In this case the analysis is done block by block, without overlap. C js nov a square matrix, A a square-diagonal matrix and P the identity matrix P'I (see (6)). The diagonal of A is periodic with period N and its elements ar matrices, either as C=FPA (3) and…”

“…In recent years the discrete short-time Fourier transform (DSTFT) has proved to be a useful tool for the analysis, modification, and synthesis of nonstationary signals, such as speech [1][2][3][4][5][6][7]. Portooff [4] has extended the earlier mathematical models and introduced the explicit use of a synthesis window.…”

“…The algebraic approach used in this paper can be applied also when the analysis window length N>fJ. Such windows are needed if the amount of overlap of adjacent analysis filters in the frequency domain is to be reduced [6,7]. In this paper we consider only the case NM, since if N>M the problem is more difficult.…”