多带激励.docVIP

多带激励语音编码算法 (MBE) Multi-Band Excitation Speech Coder (MBE) 1 Introduction 2 Multi-Band Excitation Analysis 3 MBE Synthesis ( 附录A：基音估值时域等效公式的推导 1. Introduction MPLPC, CELP etc. hybrid coding are successful at bit-rate 4.8-16 kb/s, and below, the limitation in the excitation accuracy become very dominant, resulting in rapid degradation of speech quality. MBE is proposed by Lincoln Lab. of MIT. Speech coding using MBE is successful at bit-rate about 4 - 8 kb/s. In the Multi-Band Excitation Speech Coder (MBE), a different approach is use for the representation of the excitation signal. The MBE speech coder replaces the single voiced/unvoiced classification of a classical vocoder with a set of such decisions over harmonic intervals in the frequency domain. This enable each speech segment to be a mixture of voiced and unvoiced. The MBE matching process is more perceptually biased where waveform similarity is not important. In fact, when speech is mainly unvoiced the original and synthesized speech segments do not have any similarities in time domain. As explained later, the speech envelope can still be represented by traditional methods. Fig. 1 Conventional and MBE Speech Model (2) (1) (4) (3) (6) (5) (7) Fig.2 MBE Excitation Model Original Spectrum, (2) Spectral Envelop, (3) Periodic(Pitch Period) Spectrum, (4) Multiple Voiced/Unvoiced Decisions, (5) White Noise Spectrum, (6) Excitation Spectrum, (7) Synthesized Spectrum MBE algorithm is based on minimizing the error between the original speech spectrum and the reconstructed speech spectrum using the squared error criterion: (1) where is the windowed original speech spectrum, is the reconstructed speech spectrum, is a weighted function. Or to minimize: (2) 2. Multi-Band Excitation Analysis The simplified block diagram of the MBE speech coder is shown in Fig. 2. In MBE speech analysis, before the analysis can begin, the sampled speech signal is high pass filtere