A new speech coding scheme is introduceed which processes very Long, variable length, speech segments that approximately coincide with natural "bursts" of the speech sound. A new segmentation process distinguishes between steadystate and transient sections of the speech waveform, thereby enhancing the correlation within segments to be coded. Pitch information is used to obtain a synchronously structured two dimensional array of discrete cosine transform coefficients. This array is then quantized via a hierarchical vector quantization scheme capable of hand]ing large variable size blocks. The combination of these features results in a speech coder yielding SNR of 15 db at 9.6 Kb/s and perceptually good communication quality.
展开▼