This paper addresses the problem of musical noise in transform-based audio codecs. This artifact occurs when encoding audio segments with a noise-like spectrum, at low bit rates where signal quantization results in significant zero-valued coefficients. Due to the quantization commonly used, bands containing several non-zero coefficients are quantized to only one or two, giving rise to a musical artifact. This has been identified in other codecs, such as in CELT/OPUS, where a special transform is used prior to quantization to remedy the problem. In this paper, we provide a modified approach consisting of a Hadamard transform combined with an interleaving scheme. Simulation shows the proposed method has a lower complexity and yields improved perceptual scores as measured by PEAQ.
展开▼