Line Spectral Pair

Description

Line Spectral Pairs (LSPs), also known as Line Spectral Frequencies (LSFs), are a fundamental concept in speech coding, particularly within the Adaptive Multi-Rate (AMR) and Enhanced Variable Rate Codec (EVRC) families standardized by 3GPP. They provide a mathematically equivalent but more robust representation of the Linear Predictive Coding (LPC) filter coefficients that model the human vocal tract. The LPC filter, typically an all-pole filter of order 8 or 10, is used to capture the spectral envelope of speech. Direct quantization of LPC coefficients is problematic because small quantization errors can lead to unstable synthesis filters. LSPs solve this by transforming the LPC coefficients into a domain where stability is easily preserved and quantization is more efficient.

The transformation from LPC to LSP involves decomposing the inverse filter transfer function A(z) into two auxiliary polynomials: P(z) and Q(z). These polynomials are constructed such that P(z) has even symmetry and Q(z) has odd symmetry. The roots of P(z) and Q(z) on the unit circle in the z-plane are the Line Spectral Pairs. For an LPC filter of order M, there are M LSPs, which are interleaved and sorted in ascending order. The key property is that the LPC filter is stable if and only if the LSPs are strictly ordered and lie on the unit circle. This property allows for simple stabilization procedures by adjusting the LSPs to maintain ordering.

In practical speech codecs, the LSPs are quantized using predictive or multi-stage vector quantization. Due to their ordering property, quantization can be performed on the differences between adjacent LSPs, ensuring that the reconstructed LSPs remain ordered and the synthesis filter remains stable. Furthermore, LSPs exhibit good interpolation properties, allowing for smooth transitions between frames, which is crucial for maintaining speech quality during frame erasures or when using variable-rate codecs.

The 3GPP specifications 26.090 (AMR speech codec; Transcoding functions) and 46.060 (Enhanced Full Rate (EFR) speech codec; Transcoding functions) detail the use of LSPs in the respective codecs. In AMR, the LSPs are quantized using a split vector quantizer with 3 or 4 sub-vectors, depending on the mode. The quantized LSPs are then interpolated and converted back to LPC coefficients for each subframe to synthesize the speech signal. The use of LSPs is one of the key factors enabling the high quality and robustness of 3GPP speech codecs at low bit rates.

In summary, Line Spectral Pairs are a critical component of modern speech coding, providing a stable, efficient, and quantizable representation of the vocal tract filter. Their mathematical properties and practical benefits have made them the de facto standard in 3GPP codecs from Rel-8 through Rel-19 and beyond.

Purpose & Motivation

The primary motivation for developing Line Spectral Pairs was to overcome the limitations of directly quantizing Linear Predictive Coding (LPC) coefficients in speech codecs. Direct quantization of LPC coefficients is highly sensitive to errors; even small quantization noise can shift the roots of the LPC polynomial outside the unit circle, causing the synthesis filter to become unstable and producing severe artifacts in the reconstructed speech. Additionally, LPC coefficients do not lend themselves to efficient interpolation, which is necessary for smooth transitions between frames and for handling packet loss.

LSPs address these issues by transforming the LPC coefficients into a domain where stability is guaranteed by the ordering property of the LSPs. This allows for simple and effective stabilization techniques, such as sorting and enforcing a minimum distance between adjacent LSPs. Moreover, LSPs have been shown to be more robust to quantization errors, meaning that fewer bits are required to achieve a given level of speech quality. This is particularly important in low-bit-rate codecs where bandwidth is at a premium.

Historically, LSPs were introduced in the 1980s and quickly became a standard tool in speech coding, including in early 3GPP codecs like EFR and AMR. Their inclusion in 3GPP specifications from Rel-8 onwards reflects their proven effectiveness and continued relevance in modern communication systems. The evolution of LSP quantization techniques, such as split vector quantization and predictive quantization, has further improved efficiency and quality, ensuring that LSPs remain a cornerstone of speech coding in 3GPP networks.