A Multimode Multi-Band Excitation with Linear Prediction vocoder

This paper presents a Multimode Multi-Band Excitation with Linear Prediction model (MMBE-LP) at 2.35kbps. Unvoiced/Voiced (U/V) decisions and spectrum amplitudes estimations are improved in this vocoder compared with MBE vocoder. For better quantization results, different codebooks are used for different modes of U/V decision, so the number of sub-bands in a frame is fixed. Spectral amplitudes are estimated by Linear Prediction Mode and quantized by MSVQ (Multi-Stage Vector Quantization). Simulation results show that the unvoiced and voiced parts of synthetic speech are coherent with the parts of original speech obviously.