Code Excited Linear Prediction

Description

Code Excited Linear Prediction (CELP) is a sophisticated speech coding algorithm that forms the basis for several 3GPP voice codecs, most notably the Adaptive Multi-Rate (AMR) codec family. At its core, CELP is a hybrid codec combining source coding (linear predictive coding, or LPC) and waveform coding principles. The algorithm operates by analyzing short frames of speech (typically 20 ms) and extracting parameters that model the spectral envelope of the speech signal using linear prediction. This LPC filter represents the short-term correlations in speech, effectively modeling the resonant properties of the vocal tract. The residual signal—the difference between the original speech and the output of the LPC filter—is then approximated using an excitation signal selected from a fixed codebook.

The CELP encoder works through an analysis-by-synthesis loop. For each speech frame, the encoder searches through a codebook of possible excitation vectors (stochastic or algebraic) to find the one that, when passed through the LPC synthesis filter, produces synthesized speech that best matches the original input according to a perceptual error criterion. This perceptual weighting filter emphasizes errors in frequency regions where the human ear is more sensitive, improving subjective speech quality. The selected codebook index, along with the LPC filter coefficients (often converted to Line Spectral Pairs or LSPs for quantization efficiency), gain parameters, and potentially other side information, are transmitted to the decoder.

At the decoder, the received parameters are used to reconstruct the speech signal. The excitation vector is retrieved from the identical codebook, scaled by the gain, and fed into the LPC synthesis filter to produce the synthesized speech frame. Post-processing techniques, such as adaptive post-filtering, are often applied to enhance the perceptual quality by reducing quantization noise. The CELP structure is highly adaptable, allowing for multi-rate operation where the bitrate can be dynamically adjusted based on network conditions and voice activity, trading off between speech quality and bandwidth usage.

In the 3GPP ecosystem, CELP is implemented in various codec specifications. The AMR codec, defined in 3GPP TS 26.090, uses an Algebraic CELP (ACELP) structure for its core modes. ACELP employs an algebraic codebook where the excitation pulses have fixed, predefined positions and amplitudes, simplifying the codebook search. The AMR-WB (Wideband) codec, specified in TS 26.190, extends CELP to cover a wider audio bandwidth (50–7000 Hz) for improved naturalness and intelligibility. These codecs are integral to the Voice over LTE (VoLTE) service, ensuring backward compatibility and high-quality voice across generations of mobile networks.

Purpose & Motivation

CELP was developed to address the critical need for efficient digital speech compression in bandwidth-constrained mobile communication systems. Prior speech codecs, such as Pulse Code Modulation (PCM) used in traditional telephony, consumed 64 kbit/s, which was impractical for the limited radio spectrum of cellular networks. Earlier vocoders like Regular Pulse Excitation-LPC (RPE-LPC) used in GSM Full Rate provided compression but with compromised quality. CELP emerged as a solution to achieve toll-quality speech at significantly lower bitrates (e.g., 4.75 to 12.2 kbit/s in AMR), enabling more voice channels per cell and improving network capacity.

The creation of CELP was motivated by advancements in digital signal processing and computational power in the 1980s. The algorithm's analysis-by-synthesis approach and perceptual optimization allowed it to overcome limitations of simpler waveform coders and parametric vocoders. By accurately modeling the speech production mechanism and using a codebook to represent the excitation, CELP achieved a better balance between bitrate efficiency and speech quality. This made it ideal for the evolving digital cellular standards, where spectrum efficiency and voice quality were paramount for user acceptance and commercial success.

Within 3GPP, standardizing CELP-based codecs like AMR ensured interoperability across equipment from different vendors and consistent voice service quality as networks evolved from GSM/EDGE to UMTS and LTE. The multi-rate capability of CELP codecs allowed networks to adapt speech bitrates dynamically based on radio conditions (e.g., using lower rates in poor coverage to maintain call continuity) and traffic load, optimizing both quality and capacity. This adaptability was crucial for the transition to all-IP networks and VoLTE, where voice becomes just another data service but with stringent latency and quality requirements.

Key Features

• Hybrid speech coding combining LPC source modeling with codebook excitation
• Analysis-by-synthesis optimization with perceptual weighting for high quality
• Multi-rate operation enabling dynamic bitrate adaptation from 4.75 to 12.2 kbit/s (AMR)
• Algebraic Codebook (ACELP) structure for efficient excitation representation
• Wideband extension (AMR-WB) supporting 50-7000 Hz audio bandwidth
• Robustness to transmission errors through parameter redundancy and error concealment

Evolution Across Releases

Defining Specifications