Vector-Sum Excited Linear Prediction

Description

Vector-Sum Excited Linear Prediction (VSELP) is a speech coding algorithm belonging to the family of Code-Excited Linear Prediction (CELP) vocoders. As a speech codec, its function is to digitally compress the audio signal of human speech for efficient transmission over digital radio channels. The VSELP algorithm works by modeling the human vocal tract using a linear predictive coding (LPC) filter. The excitation signal for this filter—which represents the residual sound after the LPC model is applied—is not generated by a simple pulse or noise source but is selected from a predefined codebook of candidate excitation vectors.

The key architectural innovation in VSELP is the structure of its excitation codebook. Instead of a single, large stochastic codebook, VSELP typically uses a structured codebook built from a basis of overlapping vectors. The final excitation is formed as a sum of a small number of these basis vectors, each scaled by a gain factor. This structure allows for a very efficient search process. The encoder analyzes a short frame of speech (e.g., 20 ms), determines the parameters for the LPC filter (reflecting the vocal tract shape), and then searches the structured codebook to find the excitation vector that, when passed through the LPC filter, produces synthesized speech closest to the original. These parameters (LPC coefficients, codebook indices, and gains) are then transmitted to the decoder.

The decoder reconstructs the speech by using the received indices to look up the excitation vectors from its identical codebook, applying the gains, and passing the resulting excitation signal through the LPC synthesis filter. Key components of the VSELP codec are the LPC analysis/synthesis filter, the structured algebraic codebook, the adaptive codebook (for modeling pitch/pitch), and the perceptual weighting filter used in the error minimization process. Its role in the network, as referenced in 3GPP specs 46.002 and 46.020, was as a defined speech codec option, ensuring interoperability for speech services in relevant system implementations.

Purpose & Motivation

VSELP was created to provide high-quality digital speech compression suitable for the constrained bandwidth of early digital cellular radio systems. The primary problem it solved was the need to transmit intelligible, toll-quality voice within a limited channel bit rate (e.g., around 8 kbps or less) while also being feasible to implement with the digital signal processing (DSP) technology of the late 1980s and early 1990s. It addressed the limitations of earlier speech codecs like Regular Pulse Excitation (RPE) used in full-rate GSM, which had lower quality, and more complex CELP variants that were computationally prohibitive.

The motivation for developing VSELP, particularly by Motorola who pioneered it, was to achieve a favorable balance between speech quality, bit rate, computational complexity, and robustness to channel errors. Its structured codebook search was significantly less computationally intensive than a full search of an unstructured stochastic codebook, making real-time implementation on available DSP chips practical. This was a critical enabler for its adoption in systems like the North American TIA IS-54 digital cellular (D-AMPS) standard.

In the 3GPP context, its inclusion (referenced in TS 46.002 for vocabulary and TS 46.020 for performance) served to maintain a record of this historically significant codec and ensure specifications for systems that incorporated it remained complete. While VSELP itself was largely superseded in mainstream 3GPP systems by more advanced codecs like AMR and AMR-WB, which offer better quality and adaptability, its development was an important step in the evolution of CELP-based speech coding, demonstrating techniques for reducing complexity without severely compromising quality.

Key Features

• Structured algebraic codebook for efficient excitation search
• Code-Excited Linear Prediction (CELP) architecture
• Operates at medium bit rates (e.g., 8 kbps)
• Moderate computational complexity suitable for 1990s DSP
• Provides toll-quality speech intelligibility
• Includes adaptive codebook for long-term (pitch) prediction

Evolution Across Releases

Defining Specifications