Linear Predictive Coding

Description

Linear Predictive Coding (LPC) is a foundational speech coding and compression technique employed within 3GPP voice codec standards, most notably the Adaptive Multi-Rate (AMR) and Enhanced Voice Services (EVS) codecs. It operates on the principle that a speech sample can be approximated as a linear combination of its past samples, effectively modeling the spectral envelope of the human vocal tract. The core process involves analyzing a short segment of the input speech signal to compute a set of LPC coefficients that define a time-varying digital filter. This filter, known as the synthesis filter, is designed to have a frequency response that matches the formants (resonant frequencies) of the original speech. The excitation signal—representing the glottal pulse for voiced sounds or noise for unvoiced sounds—is then determined. The encoder transmits these LPC coefficients and a quantized version of the excitation signal, achieving significant data rate reduction compared to the original waveform.

Architecturally, LPC is integrated into the speech encoder and decoder components of a voice codec. Key components include the LPC analysis module, which performs autocorrelation or covariance methods to derive the coefficients, and the perceptual weighting filter, which shapes quantization noise to be less audible. The synthesis filter at the decoder reconstructs the speech waveform using the received coefficients and excitation. In codecs like AMR, LPC parameters are often transformed into more robust representations for transmission, such as Line Spectral Pairs (LSPs) or Immittance Spectral Frequencies (ISFs), which are less sensitive to quantization errors and ensure better interpolation between frames.

LPC's role in the network is critical for the physical layer processing of voice traffic. It directly impacts the air interface efficiency by determining the bit rate allocated to voice channels. In GSM and UMTS, AMR codecs utilizing LPC allow dynamic adaptation of the speech bit rate based on channel conditions, trading off speech quality for increased error protection when needed. In LTE and 5G NR, the EVS codec employs advanced LPC-based techniques like Bandwidth Extension (BWE) and Time-Domain Aliasing Cancellation (TDAC) to support super-wideband and full-band audio, providing superior voice quality while maintaining backward compatibility. The efficiency of LPC is thus a cornerstone of voice service quality and network capacity across all generations of 3GPP systems.

Purpose & Motivation

LPC was developed to address the fundamental challenge of transmitting intelligible, natural-sounding human speech over bandwidth-constrained and error-prone digital radio channels in mobile networks. Prior to its adoption in digital cellular standards like GSM, speech coding techniques such as Pulse Code Modulation (PCM) used in wired networks required high bit rates (64 kbps) that were impractical for scarce wireless spectrum. LPC-based codecs like the Full-Rate codec in early GSM aimed to compress speech to around 13 kbps or lower while preserving acceptable quality, enabling efficient use of radio resources and making digital cellular voice services commercially viable.

The motivation for LPC stems from its ability to provide a high compression ratio by exploiting the predictable nature of speech production, rather than simply digitizing the waveform. It solves the problem of maintaining voice service capacity and coverage without requiring excessive bandwidth. By separating the slowly varying vocal tract filter (LPC coefficients) from the rapidly varying excitation signal, the codec can prioritize bits effectively, allocating more to the perceptually critical spectral envelope. This approach also facilitates robust error concealment; if some parameters are corrupted during transmission, the decoder can use previous good frames to mask errors, preventing complete voice dropouts.

Historically, LPC evolved from government and academic research into low-bit-rate speech coding for secure communications and satellite links. Its integration into 3GPP standards, starting with GSM, established a scalable framework for voice compression that has been enhanced over decades. Each new codec generation (AMR, AMR-WB, EVS) has refined LPC techniques to deliver higher quality at equivalent or lower bit rates, supporting the evolution from narrowband telephony to high-definition voice and multimedia services in 5G.

Key Features

• Models the human vocal tract as a time-varying linear filter defined by LPC coefficients.
• Separates speech signal into filter parameters and an excitation signal for efficient compression.
• Enables variable bit rate operation in codecs like AMR to adapt to channel conditions.
• Uses robust parameter representations like Line Spectral Pairs (LSPs) for quantization and transmission.
• Forms the core spectral modeling component in 3GPP codecs including AMR, AMR-WB, and EVS.
• Supports bandwidth extension techniques to enhance audio bandwidth beyond the coded core band.

Evolution Across Releases

Defining Specifications