Does CELP reproduce the original waveform exactly?

Usually not. It is a lossy model-based speech-coding approach that aims for an appropriate reconstruction quality.

What is analysis by synthesis?

The encoder generates candidate reconstructions and compares them with the input to choose coding parameters.

Is every CELP codec identical?

No. Codebooks, predictive models, frame structures, bit allocation and other details vary across standards and implementations.

AI Glossary / MLOps and deployment / Intermediate

Code-excited linear prediction

Definition

Code-excited linear prediction, or CELP, is a speech-coding approach that represents speech using a predictive synthesis filter driven by a selected excitation signal. The encoder searches for an excitation and parameters whose synthesized speech closely matches the input under a chosen error criterion.

What it means

Speech contains structure that can be represented more compactly than a sequence of independent waveform samples. Linear prediction models part of that structure using relationships among nearby samples, often interpreted as an approximation to the vocal tract's filtering effect. The remaining excitation drives the model. CELP selects that excitation from a structured or conceptual codebook and transmits parameters needed for reconstruction. The goal is useful speech quality at a limited bitrate, not an exact copy of every input sample. CELP is a family of techniques with codec-specific designs.

How it works

The encoder analyzes short speech segments to estimate predictive-filter parameters. It considers candidate excitation signals, synthesizes their outputs through the model and evaluates the difference from the original, commonly using perceptual weighting. It then encodes the chosen excitation information, gains and filter-related parameters.
Many practical designs combine an adaptive codebook related to past excitation with an innovation or fixed-codebook component. The decoder uses the transmitted information to generate excitation and run the synthesis filter. Stability, parameter quantization, frame structure and error handling are crucial implementation details. The analysis-by-synthesis search can be computationally demanding, while the decoder often follows a more direct reconstruction path.

Three practical examples

A speech communication codec uses predictive modeling to transmit intelligible audio with limited bandwidth.
A signal-processing class reconstructs a speech segment from filter parameters and a selected excitation.
A codec evaluation compares speech quality, delay and complexity under different operating conditions.

Think of it this way

Think of describing a musical instrument's resonance and how it is excited, rather than sending a separate instruction for every tiny pressure change. The decoder recreates sound from that compact description. The analogy is limited because CELP uses a mathematical speech model, not a complete physical simulation of a person's vocal tract or a real instrument.

Worked example

Consider a simplified one-tap synthesis rule y[n] = 0.5y[n−1] + e[n], starting with y[−1] = 0. Candidate excitation A is [1,0,0], producing [1,0.5,0.25]. Candidate B is [0.8,0.2,0], producing [0.8,0.6,0.3].
For a target segment [1,0.5,0.25], candidate A has zero squared error under this toy model. Candidate B has squared error 0.2² + (−0.1)² + (−0.05)² = 0.0525. A real CELP encoder uses richer filters, codebooks, gain choices and perceptual weighting. The example demonstrates analysis-by-synthesis selection, not a standards-compliant speech encoder.

Illustrative case study

In a fictional voice-assistance project, a team evaluates a low-bitrate speech link before sending audio to a recognizer. The decoded speech sounds intelligible to listeners, but some names become harder for the recognizer to distinguish. The team evaluates both listening quality and recognition performance, including background noise and packet loss. It also accounts for coding delay. This case illustrates why a speech codec's usefulness depends on the complete application rather than a bitrate or subjective impression alone.

Common misconceptions and mistakes

CELP does not transmit a text transcript, and it is not simply a lookup table of recorded words. The codebook concerns excitation patterns used by a synthesis model. Linear prediction does not mean that the entire speech-coding system is linear; search, quantization and adaptation introduce other operations. Another mistake is assuming that speech-oriented coding will preserve music or arbitrary sounds equally well. Performance depends on the codec's design and intended signal class.

Business use

For communication and AI speech products, evaluate bitrate, latency, computational load and robustness alongside listening and task quality. Test representative speakers, languages, noise conditions and channel impairments. Use the exact format specification and compatible implementations when interoperability matters. If audio will feed a model, assess the model on decoded audio from the intended pipeline. A compact signal representation is valuable only if it preserves the information the service actually needs.

Check your understanding

A CELP encoder selects the excitation with the smallest perceptually weighted reconstruction error. Does that mean it sends the original speech samples unchanged?

Answer: No. It sends a compact description of the selected model and excitation so the decoder can synthesize an approximation.

Related terms

Linear prediction
Speech coding
Codebook
Codec
excitation
analysis by synthesis
perceptual weighting
synthesis filter
quantization
bitrate

Learn next

Learn linear predictive coding and analysis-by-synthesis next. Work through a tiny predictive filter with two candidate excitations, then examine how a real codec adds adaptive excitation, quantization and perceptual error weighting.

Sources

ITU-T, G.729 speech-coding recommendation: https://www.itu.int/rec/T-REC-G.729 IETF, Opus specification discussing predictive speech coding components: https://www.rfc-editor.org/rfc/rfc6716.html Primary standards search excerpts consulted; full codec-specific and editorial review remain pending.

Previous: Code-break procedure Next: Code point