A hybrid LP-harmonics model for low bit-rate speech compression with natural quality

We present a new LP-harmonic based speech codec. At the coder the speech signal is pre-processed, and an LP analysis is carried out together with pitch estimation and voicing decision. At the decoder and when the frame is voiced, the encoded parameters are used to estimate the spectrum envelope, and extract and classify the harmonies as either strong or weak depending on their relative distance from multiples of the fundamental frequency. Strong harmonics parameters are then used to generate pure sinusoids, while weak harmonics are used to generate a mixed signal of a pure sinusoid and a random like signal. For unvoiced frames, the excitation for the LP filter is expressed as a white noise signal. The proposed model allows for the mixing of strong and weak periodic signals together with random signals to produce an excitation input that results in natural speech. Informal testing of two versions of the coder operating at 1.82, and 2.56 kbit/s showed that the output speech has high intelligibility, with quality comparable to that of a 4 kbit/s sinusoidal codec.