AR過程の逐次適応同定による音声の高能率符号化

概要

論文の詳細を見る
The coding principle of the research is an application of recursive and adaptive identification process of AR (auto-regressive) process. The system can reproduce speech wave only by the transmission of the difference (error) signal between input speech and predicted one at sending end, with two identical linear predictors and Kalman Filters at sending and receiving end respectively. At sending and receiving end, predictors are controlled recursively and adaptively, in the same way, by the difference signal with Kalman Filters. The final output is the sum of received error signal and predicted one at the receiving end. Several techniques are developed to reduce a bit rate. A bit rate can be reduced to 16kbps with an average SNR better than 30 dB, and a quality of reproduced speech is good enough for an ordinary communication use. A bit rate can be reduced to 8kbps by an alternate use of control and free running samples, and voice quality is definitely improved by the addition of one bit transmission of the sign of difference signal at free running interval with a slight increase of bit rate to 12kbps.
社団法人日本音響学会の論文