Speech coding is the compression of speech (into a code) for transmission. Generally, the spectral envelope of the input signal is represented by an all-pole filter which is excited by a pulse train. The most common filter generation method is linear predictive coding (LPC) by the autocorrelation method. However, the filter coefficients are sensitive to errors and their range is largely unknown. The coefficients are therefore coded into some other representation, which is more tolerant to errors. Such representations are, among others, line spectrum pair (LSP), log-area ratios (LAR) and reflection coefficients (related to lattice filters and Levinson-Durbin recursion). The most widely used of these is the LSP, which is used for example in the GSM standard.
Speech coding methods apply theory from audio compression and audio signal processing, by concentrating only on information of the signal that is audible. For example, in narrow-band speech coding, only information in the frequency band 400Hz to 3500Hz is transmitted but the reconstructed signal is still adequate for illegbility.