CELPbased

CELPbased refers to codecs and systems that implement Code-Excited Linear Prediction (CELP) for speech compression. CELP-based codecs model the vocal tract with a linear predictive filter and use a coded excitation signal to drive that filter. They are typically implemented on a frame-by-frame basis, with Hop or frame lengths of about 10 to 20 milliseconds.

In a CELP-based encoder, the speech signal is analyzed to produce a target representation of the vocal tract and the excitation needed to recreate the speech. The encoder searches a fixed codebook for an excitation pattern that, when filtered by the LPC model, best matches the target speech. The search also estimates a gain for the chosen codeword. The resulting codebook index and gain are quantized and transmitted as the bitstream. The decoder reproduces the excitation from the received index and gain and passes it through the corresponding LPC synthesis filter to synthesize the speech signal.

Variants of CELP include Algebraic CELP (ACELP), which uses specialized algebraic codebooks to reduce computational complexity, and low-delay versions (LD-CELP) that shorten overall latency. Wideband versions extend the approach to broader audio bandwidths, improving naturalness. Collectively, CELP-based codecs have been standardized and deployed in telecommunication and VoIP systems, and they have been widely used in mobile and landline networks, particularly where low bitrates and robustness to transmission errors are important.

Advantages of CELP-based codecs include relatively good speech quality at moderate to low bitrates, moderate computational requirements, and resilience to some packet loss. Limitations include potential artifacts at very low bitrates, limited naturalness compared with newer neural vocoders, and licensing or patent considerations for certain implementations. See also Code-Excited Linear Prediction, ACELP, and LD-CELP.