On pitch‐synchronous abridgement of the linear prediction residual

We developed an approach to abridging the short‐term LPC residual and using the abridged information to generate speech at 16.0 kilobits per second (kb/s). The abridgement logic requires an average of less than one multiply per speech input sample. It offers a simple way to add a high‐rate, improved quality capability to existing 2.4 kb/s LPC voice communication systems. Our simulated 16.0 kb/s residual‐excited linear prediction (RELP) transmission system using abridgement generates very good quality speech output. Our abridgement approach is independent of voiced/unvoiced decisions, restricts the influence of errors to a single frame, and strongly resists brief input anomalies. We designed this medium‐rate system as an alternative to the baseband residual‐excited coder currently used in the NRL Multirate Processor (MRP). Our (computationally) simpler approach scores lower overall on the Diagnostic Rhyme Test than the MRP baseband system at 16.0 kb/s: 87.4 vs 93.9. Listening suggests that the speech quality from our approach is closer to the MRP quality for continuous speech than for the DRT words. For some applications, the computational efficiency and stability of our approach could out‐weigh the quality advantage of the MRP.