Study On Artificial Bandwidth Extension Of Speech Signal

In order to obtain wideband speech quality without changing the existing speech communication network, artificial bandwidth extension technology, which only uses the narrowband speech information to reconstruct the wideband speech, has been developed. Since the traditional bandwidth extension method must fully decode the narrowband bit-stream to speech and encode the extend speech after the bandwith extension, its computational complexity is fairly high. To remedy this problem, a novel speech bandwidth extension method based on bit-stream is proposed in this thesis. Considering the differences among different codes in bit-stream formats, we take the bandwidth extension methods from G.729to G.729.1and from AMR to AMR_WB as examples to describe the proposed method. Evaluations show that, compared with the traditional bandwidth extension methods, under the premise of ensuring the extended speech quality, the computational complexity of the proposed methods is significantly reduced.The main tasks of this paper are as follows:(1)Introduce the digital model of speech, linear prediction analysis, vector quantization, dynamic clustering, G.729/G.729.1and AMR/AMR_WB coding standards. Review the existing bandwidth extension methods. In the feature parameters extraction part, we explain in detail the process of extracting the line spectrum pair (LSP) or the immittance spectral pair(ISP), the pitch, the fixed-codebook, and the gain;.(2)In the training stage, the C-Means clustering algorithm and a weighted clustering algorithm are adopted to establish the one-to-one correspondence envelope codebooks from G.729to G.729.1.(3) Support vector regression (SVR) model is adopted to extend the AMR_WB ISP from AMR LSP.(4) Elaborate the training and extending process of other parameters and the synthesis of the G.729.1/AMR WB wideband bit-stream.