Immittance Spectral Frequencies

Description

Immittance Spectral Frequencies (ISF) are a representation of Linear Predictive Coding (LPC) parameters used in speech coding. The term is defined in 3GPP specifications 26.190 (Adaptive Multi-Rate Wideband, AMR-WB) and 26.290 (Extended Adaptive Multi-Rate Wideband, AMR-WB+), and is treated as a synonym for Immittance Spectral Pairs (ISP). The ISF representation is obtained by decomposing the LPC inverse filter transfer function A(z) into two polynomials, one with even symmetry and one with odd symmetry. The roots of these polynomials on the z-unit circle are the Immittance Spectral Frequencies.

The mathematical construction begins with the LPC polynomial A(z) of order p, which models the vocal tract as an all-pole filter. Two auxiliary polynomials P(z) and Q(z) are formed: P(z) = A(z) + z^-(p+1) A(z^-1) and Q(z) = A(z) - z^-(p+1) A(z^-1). P(z) has even symmetry and Q(z) has odd symmetry. Because of these symmetry properties, the roots of P(z) and Q(z) lie on the unit circle in the z-plane and interlace with each other. The angular positions of these roots, expressed as frequencies, are the ISF. For an LPC order p, there are p ISF values, which fully describe the spectral envelope of the speech signal.

In the AMR-WB codec, LPC analysis is performed once per 20 ms frame, and the resulting ISF are quantized and transmitted to the decoder. The ISF are preferred over raw LPC coefficients because they have better quantization properties: they are more evenly distributed, they are less sensitive to quantization errors, and they allow efficient vector quantization with a small number of bits. The ISF also have a natural ordering property (they are sorted in ascending frequency), which simplifies quantization and interpolation. In AMR-WB, the ISF are quantized using split vector quantization, and the quantized values are interpolated between frames to provide smooth spectral transitions.

The ISF representation is also used in the EVS codec (26.445) and other 3GPP speech codecs. In the decoder, the quantized ISF are converted back to LPC coefficients, which are then used to filter the excitation signal to reconstruct the speech. The stability of the synthesis filter is guaranteed if the ISF remain ordered and within the unit circle, which is ensured by the quantization scheme. The ISF are also used in speech enhancement and noise suppression algorithms within the codec.

In summary, Immittance Spectral Frequencies are a mathematically elegant and practically efficient representation of the short-term spectral envelope of speech. They are central to the operation of wideband speech codecs in 3GPP systems, enabling high-quality voice services over mobile networks. The term is used interchangeably with Immittance Spectral Pairs, and both refer to the same underlying concept.

Purpose & Motivation

The Immittance Spectral Frequencies representation exists to provide a robust and efficient way to encode the short-term spectral envelope of speech in LPC-based codecs. Raw LPC coefficients are difficult to quantize because small errors can cause the synthesis filter to become unstable, and their dynamic range varies widely. ISF solve this problem by transforming the LPC coefficients into a set of frequencies that are inherently ordered and bounded, making quantization more reliable and efficient.

The use of ISF also enables efficient interpolation between frames, which is important for smooth transitions and high speech quality at low bit rates. In the context of 3GPP codecs such as AMR-WB and EVS, ISF allow the codec to achieve transparent quality at bit rates as low as 6.6 kbps, which is essential for spectral efficiency in mobile networks. The concept was introduced in Release 8 as part of the AMR-WB codec specification, and it has been carried forward in subsequent releases without fundamental changes to its definition.

Historically, the ISF representation was developed in the speech coding research community as an alternative to other LPC parameterizations such as Line Spectral Frequencies (LSF) and reflection coefficients. ISF are closely related to LSF but are derived from a slightly different polynomial decomposition. The 3GPP adopted ISF for AMR-WB because of its excellent quantization properties and its ability to support wideband speech (50-7000 Hz) with low complexity. The term has remained stable across releases, reflecting its mature and well-understood nature.