Toward an interpretive framework of two-dimensional speech-signal processing

Abstract

Traditional representations of speech are derived from short-time segments of the signal and result in time-frequency distributions of energy such as the short-time Fourier transform and spectrogram. Speech-signal models of such representations have had utility in a variety of applications such as speech analysis, recognition, and synthesis. Nonetheless, they do not capture spectral, temporal, and joint spectrotemporal energy fluctuations (or "modulations") present in local time-frequency regions of the time-frequency distribution. Inspired by principles from image processing and evidence from auditory neurophysiological models, a variety of twodimensional (2-D) processing techniques have been explored in the literature as alternative representations of speech; however, speech-based models are lacking in this framework. This thesis develops speech-signal models for a particular 2-D processing approach in which 2-D Fourier transforms are computed on local time-frequency regions of the canonical narrowband or wideband spectrogram; we refer to the resulting transformed space as the Grating Compression Transform (GCT). We argue for a 2-D sinusoidal-series amplitude modulation model of speech content in the spectrogram domain that relates to speech production characteristics such as pitch/noise of the source, pitch dynamics, formant structure and dynamics, and offset/onset content. Narrowband- and wideband-based models are shown to exhibit important distinctions in interpretation and oftentimes "dual" behavior. In the transformed GCT space, the modeling results in a novel taxonomy of signal behavior based on the distribution of formant and onset/offset content in the transformed space via source characteristics. Our formulation provides a speech-specific interpretation of the concept of "modulation" in 2-D processing in contrast to existing approaches that have done so either phenomenologically through qualitative analyses and/or implicitly through data-driven machine learning approaches. One implication of the proposed taxonomy is its potential for interpreting transformations of other time-frequency distributions such as the auditory spectrogram which is generally viewed as being "narrowband"/"wideband" in its low/high-frequency regions. The proposed signal model is evaluated in several ways. First, we perform analysis of synthetic speech signals to characterize its properties and limitations. Next, we develop an algorithm for analysis/synthesis of spectrograms using the model and demonstrate its ability to accurately represent real speech content. As an example application, we further apply the models in cochannel speaker separation, exploiting the GCT's ability to distribute speaker-specific content and often recover overlapping information through demodulation and interpolation in the 2-D GCT space. Specifically, in multi-pitch estimation, we demonstrate the GCT's ability to accurately estimate separate and crossing pitch tracks under certain conditions. Finally, we demonstrate the model's ability to separate mixtures of speech signals using both prior and estimated pitch information. Generalization to other speech-signal processing applications is proposed.