| Summary: | 碩士 === 大同工學院 === 電機工程研究所 === 87 === In various speech coding methods, the code-excitation linear prediction (CELP) and its variants are most widely used and can produce good speech quality in medium to high bit rates. However, as the bit rate is lowered, the synthetic speech may sound buzzy and even with tonal noise. The degree of this problem increases rapidly when background noise is present. Therefore, the research of improving the synthetic speech quality in low bit rate is an important area for speech coding.
In order to solve the problem mentioned above, we propose an algorithm that can be added on CELP vocoders at low bit rate applications. Because of the characteristic of human speech, the ideal excitation signals of an LPC-based vocoder for voiced sound should be a quasi-periodic pulse train. Unfortunately, this characteristic may not be apparent enough because of the inherent ability of the CELP vocoder.
In this thesis, we propose an algorithm which tries to reconstruct the periodic characteristic for voiced sounds by detecting and enhancing the harmonics of the excitation signal in frequency domain. Two format were developed according to the different criteria: 1) if the SNR increases after processing; and 2) if the harmonics of synthesis speech is apparent enough. Our experiment results showed that this algorithm can produce more natural synthetic speech and suppress the background noise.
|
|---|
