Research On Speech Coding Model Of LPMBE

Speech communication is a basic and useful way to get information, and the key technique of speech communication is speech coding. A good speech coding model will undoubtedly give directions to speech coding technique. Traditional codecs, include waveform codecs and parameter codecs, are not perfect, with high bit rate and the synthesized speech loss of naturalness separately. Multi-Band Excited Model (MBE) can give natural synthesized speech, and at the same time LPC Model can simulate Vocal trait Model correctly. So, in this paper, we proposed LPMBE Model. LPMBE attempts to combine the merits of LPC and MBE and synthesize natural speech.In LPMBE, Glottal Excitation and Vocal Trait are considered as the causes which led speech come into being, excitation is analyzed at large in Frequency Domain. And because warp exists in the pitch detection algorithm of MBE, Modified Short Time Average Magnitude Difference Function (MAMDF) is designed to complete the task of 'pre-fetch' of pitch in excitation (Time domain). Center Variance Function(CVF) is designed to get exact pitch, and reverse reasoning method is adopted in Center Variance Function and veracity of final pitch can be assured. LPMBE can reduce the probability of Voice/Unvoiced decision error to get more natural speech quality and at the same time Transition cue is maintained. The methods of how to cut down the numbers of spectral magnitude were studied in LPMBE. Speech communication is a mutual activity between speakers, so Voice Active Detection (VAD) was also researched.Finally, in order to measure the quality of LPMBE, we simulate LPMBE and an object/subject evaluation platform for speech codecs is built in MATLAB to examine all aspects of LPMBE.