CELP

Code-Excited Linear Prediction (CELP) is a family of speech coding algorithms used to compress human speech for digital transmission. The method models the speech signal as the output of a linear predictive analysis filter that is excited by a coded sequence of pulses or filtered noise. In a CELP encoder, the current speech frame is represented by a fixed-predictor filter whose coefficients are estimated from the signal. The encoder searches a codebook of possible excitation signals to find the one that, when passed through the synthesis filter, best reproduces the original frame. The index of the chosen excitation, and often a few additional parameters, are transmitted to the decoder, which uses the same predictive filter and the received excitation index to synthesize the speech.

CELP is notable for delivering good perceived speech quality at relatively low bitrates. Variants include algebraic CELP (ACELP) and Conjugate-Structure ACELP (CS-ACELP), which optimize the excitation codebook and search process for efficiency. CELP-based codecs have been standardized and deployed in telecommunications and VoIP systems, with implementations spanning a range of bitrate options and latency requirements.

Advantages of CELP include efficient use of bandwidth, robustness to channel noise, and scalability of bitrate through changes in codebook size and predictor order. Limitations can involve coding artifacts if the codebook or predictor parameters are not well matched to the signal, and computational complexity that is higher than some simpler waveform-based approaches. Despite advances in newer codecs, CELP remains a foundational technique in low-to-mid bitrate speech coding for telecommunication applications.