| Summary: | 碩士 === 國立臺灣師範大學 === 電機工程研究所 === 94 === The practical thrust of this thesis is to propose a new speech enhancement algorithm based on noise detection/cancellation and depress the remnant noise by a linear smooth filter.
In this thesis, we put special emphasis on that the frequency distribution of noise we want to cancel is the same as the frequency distribution of speech signal in our algorithm. The new algorithm based on the concept of active noise control and speech enhancement filter, but has the originality in noise corrupted point detection and spectral-subtraction process. First of all, we explain how to win through the difficulty of signal drift by using the method of construct a stable noise template with statistics. Second, we develop a noise detection algorithm named "Diff-RMS algorithm" which analyze the energy of noise and noise corrupted signal. After we find the correct noise corrupted point, we adaptively subtract the coefficients of discrete cosine transform of noise template form an input signal. When subtraction is done, we convert the entire processed coefficients form frequency domain back to time domain by inverse discrete cosine transform. In the end, we apply a linear smooth filter to depress the remnant noise to an inaudible degree.All of the thetechnological processes contains acoustics concepts of digital signal processing in both time domain and frequency domain and that will be discussed in the following chapter in the thesis.
Finally, we implement the new noise cancellation algorithm on embedded platform to verify the stability and future practicability.
|
|---|
