Pitch synchronous modulated lapped transform of the linear prediction residual of speech

In most of the prevalent low-bit-rate speech coding schemes, lossy speech models are used, which makes the perfect reconstruction of speech impossible even at higher bit rates. In this paper, a lossless model of speech is presented. It performs a pitch synchronous modulated lapped transform (MLT) on the linear prediction residual of speech. The residual signal can be reconstructed without distortion from its transform when the quantization error is neglected. The pitch synchronous process exploits the quasi-periodicity of the speech. The MLT converts the 'periodic' signal into the frequency domain. Efficient coding similar to that used in the waveform interpolation speech coder can be achieved on the transform result. The combination of pitch synchronization and MLT provides a promising basic model for speech coding and processing.