| Summary: | 碩士 === 國立臺北科技大學 === 電機工程系研究所 === 99 === Speech signals are tend to decrease the speech quality when corrupted by background noises. The aim of speech enhancement is to reduce the background noise from a noisy speech signal while keeping the speech distortion as low as possible. Speech enhancement systems could also be a pre-processer for speech processing systems such as speech recognizer, speech coder, and so on. There are three categories for speech enhancement including filtering techniques, spectral restoration techniques, and speech model techniques.
In this thesis five speech enhancement methods based on spectral restoration techniques are investigated, including minimum mean-square error (MMSE), minimum mean-square error log-spectral amplitude (LSA), maximum a posteriori spectrum (MAP), maximum-likelihood spectral amplitude (MLSA), and maximum-likelihood spectral power (MLSP). The Wiener filter (WF) method is also included for comparison. Each method incorporates with three well-known noise tracking algorithms, including minimum statistics (MS), minima controlled recursive averaging (MCRA), and improved minima controlled recursive averaging (IMCRA) for recovering clean speech. The experimental results show that compared with the Wiener filter, all the five spectral restoration techniques provide better performance. Among all, the MMSE method incorporated with MCRA achieves the most significant enhancement performance. If apply it prior to a speech recognizer, the experimental results show that both word accuracy rate and sentence correct rate are increased obviously.
|
|---|
