Audio Signal Processing and Coding

7.3 PERCEPTUAL TRANSFORM CODER

While Brandenburg developed the OCF algorithm, similar work was simultaneously underway at AT&T Bell Labs. Johnston developed several DFT-based transform coders [John88a] [John89] for audio during the late 1980s that became an integral part of the ASPEC proposal. Johnston's work in perceptual entropy [John88b] forms the basis for a transform coder reported in 1988 [John88a] that achieves transparent coding of FM-quality monaural audio signals (Figure 7.2). A stereophonic coder based on similar principles was developed later.

Figure 7.2. PXFM encoder (after [John88a]).

7.3.1 PXFM

A monaural algorithm, the perceptual transform coder (PXFM), was developed first. The idea behind the PXFM is to estimate the amount of quantization noise that can be inaudibly injected into each transform domain subband using PE estimates. The coder works as follows. The signal is first windowed into overlapping (1/16) segments and transformed using a 2048-point FFT. Next, the PE procedure described in Section 5.6, is used to estimate JND thresholds for each critical band. Then, an iterative quantization loop adapts a set of 128 subband quantizers to satisfy the JND thresholds until the fixed bit rate is achieved. Finally, quantization and bit packing are performed. Quantized transform components are transmitted to the receiver along with appropriate side information. Quantization ...