Burst Packet Loss Concealment and Noise Robust for Distributed Speech Recognition

• Main Authors: Cheng-Chang Lee, 李政璋
• Other Authors: 廖元甫
• Format: Others
• Language: zh-TW
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/3dk6sm

碩士 === 國立臺北科技大學 === 電腦與通訊研究所 === 94 === The thesis discusses the burst packet loss concealment and noise environment mismatch compensation problem over the phase of distributed speech recognition. First, for burst packet loss, we use sub-frame interleaver to disperse burst length in the front-end, and reconstruct lost feature after space transformation in the back-end, finally apply ARMA filter to smooth the discontinuity due to the reconstruction. And, for noise environment mismatch, the a priori knowledge interpolation method is proposed to alleviate the problem of unseen environments. We evaluate our methodology on the Aurora2 noisy digits database. In the packet loss concealment experiment, average recognition rate on nine simulated channel conditions is improved from 76.16%(ETSI baseline) to 93.06%. In the noise environment mismatch compensation experiment, the performance of all sets were improved from MVA baseline to AKI method (setA : from 92.09% to 92.91%, setB : from 92.08% to 92.11%, setC : from 91.71% to 92.37%).