As was mentioned before, lossy compression is frequently used as a front end to a lossless compressor. In the case of COMPAND.C, this is a very effective strategy. After the files have been processed, far fewer codes are present in the output file, which makes string matching more likely, such as that used by LZSS compressors. By compressing a file by 50 percent using the companding strategy, then by applying LZSS compression, we can frequently achieve upwards of 90 percent compression on sound samples.

Other Techniques

This chapter covered some of the simpler techniques used to compress sound samples. As the level of processing power available for processing goes up, far more complicated algorithms are being applied.

One of the most common compression algorithms in use today has been sanctioned by the CCITT in their recommendation G.721. The G.721 algorithm uses Adaptive Differential Pulse Code Modulation (ADPCM) to encode digital signals at 16Kbps or 32Kbps. This algorithm is commonly performed by digital signal processors, and it is generally applied to data that has already been digitized using standard codes.

The ADPCM algorithm combines two techniques. The first, delta pulse code modulation, encodes sound signals by measuring the difference between two consecutive samples, not their absolute values. The quantization level adapts itself to the changing input signals, so the size of the encoded voltage changes as the input signal changes. When the signal moves from a high voltage to a low voltage at a step rate, the encoded step value will be high. If a quiet input signal is being encoded, the step value will be low.

This becomes complicated because the ADPCM algorithm requires that the transmitter predict in advance where the input signal is headed. If this prediction is not made accurately, it is not possible to make good judgments about the size of the step defined by each code. The process of predicting where a waveform is headed occupies most of the processor’s time.

To compress sound samples to even lower bit rates, even more sophisticated techniques, such as Linear Predictive Coding (LPC), are used. Human speech can be compressed and replayed in a recognizable state with rates as low as 2,400 bits per second using LPC.

LPC attempts to compress human speech by modeling the vocal tract that produces the speech. Instead of storing thousands of samples per second. LPC instead attempts to determine just a few parameters that model the process used to create the sound. The success or failure of LPC hinges on the ability of the compressor to execute millions of instructions per second during the compression process.

Processes such as LPC and ADPCM represent the type of algorithms that will be used more and more frequently on the desktop. Unfortunately, the complexity of these algorithms are far beyond the scope of a sample program in this chapter.