(Code Excited Linear Predictive) A speech compression method that achieves high compression ratios along with toll quality audio. LD-CELP (Low-Delay CELP) provides near toll quality audio by using a smaller sample size that is processed faster, resulting in lower delays. LD-CELP was developed by Dr. Raymond Chen when he was at AT&T. Chen was also involved with CELP. CELP is pronounced "kelp."
A family of voice compression algorithms used in voice over frame relay (VoFR), voice over Internet Protocol (VoIP), and in digital cellular radio systems based on cdmaOne and PDC standards. A key element of CELP and its variants is the construction and maintenance of a codebook, which comprises binary descriptions of sets of voice samples. CELP gathers a set of 80 8-bit PCM voice samples, representing 10 milliseconds (ms) of a voice stream, in a buffer. CELP employs silence suppression to remove periods of silence and redundancy from the data set, normalizes the volume level, compares the resulting data set to a set of candidate shapes in the codebook, and selects the shapes that most closely match the actual data. (Note: CELP uses a stochastically overlapped codebook, with each entry sharing all but two samples with its neighboring entries.) CELP then transmits the index number of the selected code description and the average loudness level of the set of samples. Every 10 ms, the code is sent across the network in a block of 160 bits, yielding a data rate of 16 kbps, which is a compression ratio of 4:1, compared with toll quality PCM voice over the circuit-switched public switched telephone network (PSTN) at 64 kbps. At the receiving end of the transmission, the transmitted code is compared to the codebook, the PCM signal is reconstructed, and, eventually, the analog signal is reconstructed. See also algorithm, analog, binary, buffer, cdmaOne, circuit switching, compression, CS-ACELP, LDCELP, PCM.