An improved mixed excitation linear prediction (MELP) coder

This paper presents an improved Mixed Excitation Linear Prediction (MELP) coder. The MELP is the linearprediction-based speech coder that was recently chosen as the new 2400 bps U.S. Federal Standard. Even though the MELP is quite good, there are still some perceivable distortions, particularly around non-stationary speech segments and for some low-pitch male speakers. The key features of our new coder include a robust pitch detection algorithm, a new plosive analysis/synthesis method, and a post processor for the Fourier magnitude model. Formal quality tests are used to show that the new MELP improves the quality of the U.S. Federal Standard MELP coder while requiring only a small increase in algorithmic delay and while also retaining compatibility with the Federal Standard MELP bit-stream speci cation.