Speech Spectrum Coding Based on KLT and Power Series
碩士 === 國立臺北科技大學 === 電機工程系研究所 === 97 === Most of low bit-rate speech coders based on linear predictive coding (LPC) model use line spectrum frequencies (LSF) to represent the short-time spectrum of speech. The Karhunen-Loeve transform power-series vector quantizer (KLT-PSQ) method is proposed in the...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | zh-TW |
| Published: |
2009
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/9aa5hb |
| Summary: | 碩士 === 國立臺北科技大學 === 電機工程系研究所 === 97 === Most of low bit-rate speech coders based on linear predictive coding (LPC) model use line spectrum frequencies (LSF) to represent the short-time spectrum of speech. The Karhunen-Loeve transform power-series vector quantizer (KLT-PSQ) method is proposed in the thesis to efficiently decrease the bit rates. The multi-stage vector quantization (MSVQ) that is adopted by mixed excitation linear prediction (MELP) coder needs 24-25 bits/frame to reach an spectral distortion of 1 dB. KLT-PSQ1 and KLT-PSQ2 can maintain the same spectral transparency at only 20-21 bits/frame. The scheme that includes both KLT-PSQ1 and a fixed memoryless VQ controlled by a switch is denoted as KLT-PSQ3. KLT-PSQ3 is shown to be able to significantly decrease error propagation. Over the memoryless binary symmetric channel (BSC) with error rates from 0 % to 0.5 %, the experimental result shows that the performance of KLT-PSQ3 is more robust than that of MSVQ. However, the BSC with error rate from 1 % to 10 %, the advantage is obviously canceled out by error propagation.
|
|---|