Is ACELP the same as CELP?

It is a CELP approach distinguished by an algebraically structured excitation codebook.

Does the decoder need the original speech to choose the pulses?

No. It reconstructs the selected excitation from the transmitted indices and other parameters using the agreed codec rules.

Does a sparse excitation imply mostly silent output?

No. The excitation drives a synthesis filter, and the complete excitation can include another component. Sparse pulses do not directly describe silent stretches in the final waveform.

AI Glossary / MLOps and deployment / Intermediate

Algebraic code-excited linear prediction

Definition

Algebraic code-excited linear prediction (ACELP) is a speech-coding approach that represents speech using a prediction filter and a compact, mathematically structured excitation signal. Its algebraic codebook describes candidate excitation patterns through rules, often using a small number of signed pulses, rather than storing every possible pattern explicitly.

What it means

Speech has a spectral structure and, for voiced sounds, repeating patterns. Predictive speech coders represent these using a model and its driving signal.

In ACELP, the word algebraic refers to the organization of an excitation codebook. It does not mean that an AI system reasons about algebra. The codebook can describe many candidate pulse arrangements compactly, and the encoder selects an arrangement suited to the current speech segment.

Compression is lossy: transmitted parameters reconstruct an approximation, not every original microphone sample.

How it works

The encoder analyses a short speech frame and estimates linear-prediction parameters. These describe a filter that helps model the signal's short-term spectral structure.

In a typical CELP arrangement, an adaptive codebook contributes a component related to previous excitation and pitch. A fixed, or innovation, codebook supplies an additional component. ACELP gives that innovation codebook an algebraic structure, restricting where nonzero pulses may occur and how they are signed.

The encoder searches candidate parameters by evaluating reconstructed speech against a target. A perceptually weighted error measure gives some differences greater importance than others. The selected indices, gains and filter parameters are transmitted. The decoder uses them to reconstruct the excitation and pass it through the synthesis filter.

The exact pulse layout, bit allocation and search strategy depend on the codec. G.729's CS-ACELP is one specific design, not a complete definition of every ACELP implementation. Searching compactly represented candidates still requires computation.

Three practical examples

1. Telephone speech: An ACELP-based codec can represent a speech conversation using parameters and excitation indices, reducing payload compared with sending uncompressed samples at the same sampling rate.

2. Mobile audio integration: An application handling a speech format that uses ACELP must identify the exact codec and mode before decoding. The family name alone does not identify the bitstream.

3. AI transcription testing: A team can test recognition on audio decoded from its real communication codec. Evaluating only clean original recordings may miss the effects of compression and repeated transcoding.

Think of it this way

Think of describing a musical phrase using the instrument's settings and a short set of playing instructions instead of sending a detailed measurement of every pressure fluctuation. The receiver uses the agreed model to recreate an approximation.

The comparison is limited: ACELP's pulses are mathematical excitation components, not notes or words. It does not transcribe speech into language, and the reconstructed sound depends on precise codec rules rather than a human musician's interpretation.

Worked example

Build an invented algebraic codebook for an eight-sample excitation vector. Require exactly one pulse at an even position, chosen from 0, 2, 4 and 6, and one at an odd position, chosen from 1, 3, 5 and 7. Each pulse can be +1 or -1.

There are 4 × 4 = 16 position combinations and 2 × 2 = 4 sign combinations, giving 64 possible vectors. Six bits can identify one vector: two bits for each position choice and one bit for each sign.

For example, choose +1 at position 2 and -1 at position 5. The vector is [0, 0, 1, 0, 0, -1, 0, 0]. Its index describes a whole eight-sample pattern without storing eight separate amplitudes in the bitstream.

This is only the excitation description. A complete coder still needs gains, prediction parameters, state and framing. The toy structure is not a G.729 codebook.

For a separate bandwidth calculation, G.729's basic 8 kbit/s mode represents a 10 ms frame with 80 bits: 8,000 × 0.010 = 80. That is ten payload bytes per frame, before transport overhead, and does not mean that every ACELP codec uses that rate.

Illustrative case study

Illustrative case study - fictional

A fictional support platform adds search to its archived voice calls. Studio tests produce encouraging transcripts. Real calls, however, use an ACELP-based codec and may undergo another conversion before archiving.

It constructs a representative test set using the actual decoding and conversion path. Reviewers examine errors involving names, numbers and speech over background noise. The team also measures latency and preserves original encoded recordings for reproducibility.

Deployment follows those tests; the scenario claims no measured accuracy change.

Common misconceptions and mistakes

Algebraic does not mean symbolic mathematical reasoning or a neural network.
The excitation pulses are not the final speech waveform; synthesis filtering changes their effect.
A codebook index is only part of the information needed to reconstruct speech.
All ACELP-based codecs do not share one pulse count, bitrate or bitstream.
Speech-oriented compression should not automatically be judged suitable for music or exact waveform preservation.

Business use

ACELP is relevant to voice-system integration and to AI services built on recorded or transmitted speech. Specify the actual codec, rate, sampling characteristics and container in technical requirements.

Assess end-to-end quality with realistic speakers and noise, and include transcoding if it occurs in production. Keep listening quality, intelligibility and downstream recognition performance as distinct evaluation measures. A lower payload rate can help storage or transmission, but packet overhead, delay and processing requirements still affect the complete service.

Check your understanding

A toy codebook has four choices for each of two pulse positions and two choices for each pulse's sign. How many vectors can it describe?

Answer: 4 × 4 × 2 × 2 = 64. Six bits identify a vector, but a full speech codec requires additional parameters and framing.

Related terms

Code-excited linear prediction
Linear predictive coding
Codebook
excitation
speech codec
perceptual weighting
G.729

Learn next

Pass two different toy pulse vectors through the same simple filter and compare their outputs. Then change the filter while keeping the excitation fixed to separate the roles of excitation and synthesis parameters.

Sources

Sources: ITU-T G.729 official record: https://www.itu.int/rec/T-REC-G.729/e ; original 1996 specification, hosted by UCLA: https://www.seas.ucla.edu/spapl/codecs/g729.pdf . Relevant model, encoder, decoder and bit-allocation sections were reviewed. The historical specification explains the principles; it is not presented as the latest implementation specification.

Previous: Σ-algebra Next: Algebraic connectivity