Wavelet-based compression of speech signals on the TMS320C30 digital signal processor

Abstract

Lossless and lossy compression of speech signals using wavelet transforms is examined. Reversible wavelets based on integer arithmetic are used to calculate the transform coefficients. These coefficients are quantized and coded using a two-pass Huffman coder. Comparisons are made with other state-of-the-art coders such as the low-delay code excited linear predictive coder described in the G.728 standard. The proposed technique achieves an average lossless compression ratio of 1.8:1 and average lossy compression ratio of 3.2:1 with peak-signal-to-noise-ratio (PSNR) of more than 40 dB. The technique has also been successfully implemented on the TMS320C30 digital signal processor for real-time speech compression.