edVirtus Reference

What Is Code-Excited Linear Prediction?

What Is CELP?

Code-Excited Linear Prediction (CELP) is a low-bit-rate speech coding technique that models human speech using Linear Predictive Coding (LPC) while employing a codebook of carefully designed excitation signals to reproduce natural-sounding speech. Introduced during the 1980s, CELP represented a major advance over earlier vocoders by significantly improving speech quality without requiring high transmission bit rates. Today, CELP and its derivatives form the basis of many digital telephone, mobile communication, and Voice over IP (VoIP) speech codecs.

Like other vocoder techniques, CELP is based on the source-filter model of speech production. The vocal tract is represented by a mathematical filter whose characteristics are estimated continuously from the incoming speech. Instead of transmitting the complete speech waveform, the encoder sends only the parameters describing the filter together with information describing the excitation that drives it.

The principal innovation of CELP is the use of a codebook containing many possible excitation sequences. For each short segment of speech, the encoder searches the codebook to find the excitation that, after passing through the LPC filter, most closely reproduces the original speech. Rather than transmitting the excitation itself, only the index of the selected codebook entry is sent. This dramatically reduces the amount of information that must be transmitted while maintaining good speech quality.

A useful analogy is recreating a painting using a large collection of coloured tiles. Instead of painting every detail from scratch, the artist selects the tiles that best match each part of the original image. Similarly, a CELP encoder selects the codebook entry that most closely reproduces the desired speech segment.

Because the encoder evaluates many possible codebook entries before selecting the best one, CELP is known as an analysis-by-synthesis algorithm. The encoder effectively contains a copy of the decoder and repeatedly synthesizes candidate speech waveforms until it finds the one that most closely matches the original. Although computationally intensive, this approach produces much more natural speech than earlier LPC-based vocoders.

CELP has become the foundation of numerous international speech-coding standards. Variants are used in digital cellular systems, satellite communications, Voice over IP (VoIP), digital radio, video conferencing, and Internet telephony. Well-known codecs such as G.729, AMR (Adaptive Multi-Rate), and several military speech coders employ CELP or closely related techniques to achieve excellent speech quality at bit rates ranging from approximately 4 to 13 kbps.

It is important to distinguish CELP from Mixed Excitation Linear Prediction (MELP). Both are model-based speech coders derived from LPC, but MELP uses a mixed excitation model optimized for very low bit rates, whereas CELP employs a codebook search to achieve higher speech quality at somewhat higher bit rates. Consequently, CELP is commonly used where natural speech quality is the primary objective, while MELP is preferred where bandwidth is extremely limited.

Today, CELP remains one of the most successful speech-coding techniques ever developed. Although newer codecs employ increasingly sophisticated algorithms, many continue to build upon the same analysis-by-synthesis principles introduced by CELP. Its combination of efficient compression and high speech quality has made it one of the defining technologies of modern digital voice communications.