Improved Prediction of Nearly-Periodic Signals

Abstract:
We present methods that are relevant for the commonly used pitch predictors in speech/audio coders. We first discuss optimal pre- and post-filtering and derive a general result that post-filtering is more effective than pre-filtering. This, together with the perceived nature of the distortion, suggests the omission of the pre-filter if system delay is critical. We propose a practical paired-zero filter design for the low-rate regime. We extend this design to handle frequencydependent periodicity levels. We also provide a general performance measure for a post-filter that only uses information available at the decoder. This criterion allows the optimization or selection of a postfilter without increasing the rate. Our experiments show that the resulting methods provide state-of-the-art performance both for objective measures and in terms of quality as perceived by test subjects.