Model-based speech separation and enhancement with single -microphone input.

摘要

This thesis focuses on speech source separation problem in a single-microphone scenario. Possible applications of speech separation include recognition, auditory prostheses and surveillance systems. Sound signals typically reach our ears as a mixture of desired signals, other competing sounds and background noise. Example scenarios are talking with someone in crowd with other people speaking or listening to an orchestra with a number of instruments playing concurrently. These sounds are often overlapped in time and frequency. While human attends to individual sources remarkably well under these adverse conditions even with a single ear, the performance of most speech processing system is easily degraded. Therefore, modeling how human auditory system performs is one viable way to extract target speech sources from the mixture before any vulnerable processes.;Our approach is based on the findings of psychoacoustics. To separate individual sound sources in a mixture signal, human exploits perceptual cues like harmonicity, continuity, context information and prior knowledge of familiar auditory patterns. Furthermore, the application of prior knowledge of speech for top-down separation (called schema-based grouping) is found to be powerful, yet unexplored. In this thesis, a bi-directional, model-based speech separation and enhancement algorithm is proposed by utilizing speech schemas, in particular. As model patterns are employed to generate subsequent spectral envelopes in an utterance, output speech is expected to be natural and intelligible.;The proposed separation algorithm regenerates a target speech source by working out the corresponding spectral envelope and harmonic structure. In the first stage, an optimal sequence of Wiener filtering is determined for subsequent interference removal. Specifically, acoustic models of speech schemas represented by possible line spectrum pair (LSP) patterns, are manipulated to match the input mixture and the given transcription if available, in a top-down manner. Specific LSP patterns are retrieved to constitute a spectral evolution that synchronizes with the target speech source. With this evolution, the mixture spectrum is then filtered to approximate the target source in an appropriate signal level. In the second stage, irrelevant harmonic structure from interfering sources is eliminated by comb filtering. These filters are designed according to the results of pitch tracking.;Experiments were carried out for continuous real speech mixed with either competitive speech source or broadband noise. Results show that separation outputs bear similar spectral trajectories as the ideal source signals. For speech mixtures, the proposed algorithm is evaluated in two ways: segmental signal-to-interference ratio (segSIR) and Itakura-Saito distortion ( dIS). It is found that (1) interference signal power is reduced in term of segSIR improvement, even under harsh condition of similar target speech and interference powers; and (2) dIS between the estimated source and the clean speech source is significantly smaller than before processing. These assert the capability of the proposed algorithm to extract individual sources from a mixture signal by reducing the interference signal and generating appropriate spectral trajectory for individual source estimates.