In this thesis work, a variable frame length(VFL) speech coding method is proposed and its performance is studied by computer simulation. The proposed speech coding method is based on the idea that the performance of speech coding would be improved by varying the tramission frame length according to the stationarly of input speech signal and vector quantizing the representative PARCOR coefficient of the transmission frame.
In the proposed speech coding method, input speech signal is first segmented into voice, unvoice, or silence regions. The feature vector sequence consists of PARCOR coefficient vectors obtained by analyzing input speech signal sampleby-sample using the recursive least square lattice algorithm. In the voice region of the input signal, a pitch period interval is taken to be a subsegment length. In the unvoice region of the input signal, a 20 sample interval is taken to be a subsegment length. by joining the consecutive speech frames of phonetically similar characteristics using likelihood ratio distortion measure, we can obtain transmission frames of variable lengths. In a transmission frame interval, a representative feature vector is selected and encoded, and then transmitted to the receiver. this results in reduced transmission bit rate while maintaining the same reproduced speech quality.
Transmission parameters are encoded in two ways: one using scalar quantization, the other using vector quantization(VQ). Simulation results show that the proposed VFL speech coding method with vector quantization at the transmission bit rate of about 1267 bits/s yields the better performance as that of block processing LPC at 3800 bits/s in which a segment is taken as a transmission frame.
본 논문에서는 저 전송 음성 부호화기를 제안하였고 computer simulation을 통하여 성능분석과 유용성을 입증하였다. 제안된 부호화 방식은 음성신호의 변화부분의 특성을 잘 나타내기 위하여 prewindowed RLS 알고리즘을 사용하였고 전송속도를 줄이기 위하여 음성신호를 무성음 영역에서는 20 sample을 subsegment로, 유음성 영역에서는 pitch 주기단위를 subsegment로 취하여 likelihood ratio distortion measure를 사용하여 유사도에 따라 subsegment들을 병합함으로써 전송 frame을 결정하였고 또 전송 frame의 대표 PARCOR 계수를 vector quantization으로 부호화 하였다.
제안된 부호화기의 성능을 음성 청취 실험을 통하여 고찰하여 보았는데 PRLS 알고리즘을 사용하여 필터 계수를 추출하고 VFL과 VQ를 적용했을 때가, 즉 1267 bit/sec인 경우가 block processing 으로 필터 계수를 추출하고 하나의 subsegment를 전송 frame으로 하였을 때보다 음질이 훨씬 좋다는 것을 알 수 있었고 VQ 부호화에 의한 block processing 방식에서, 즉 3800 bit/sec에서 얻어지는 음질을 VFL과 VQ 방식을 적용한 point-wise processing 방식에서는 560 bit/sec에서 얻었다.
그러므로 본 논문에서 제안한 저 전송 부호화기는 기존의 LPC 방식보다 동일한 전송속도에서는 음질을 크게 향상시키며 동일한 음질에서는 전송속도를 크게 줄일 수 있다.
