Fractional Lags

Description

Fractional lags are a concept used in the GSM half-rate speech codec (also known as the Adaptive Multi-Rate or AMR half-rate mode) to represent pitch period delays with sub-sample resolution. In speech coding, the pitch period (or lag) is the fundamental period of the glottal excitation signal, and it is typically estimated in the range of 20 to 160 samples for 8 kHz sampled speech. Integer lags only allow pitch periods that are multiples of the sampling interval, which can lead to quantization errors and degraded speech quality, especially for female or child speakers with short pitch periods. Fractional lags extend the lag resolution to fractions of a sample, typically 1/3 or 1/6 of a sample, by interpolating the correlation function or using an interpolation filter.

The GSM half-rate speech encoder, specified in 3GPP TS 26.090 (AMR speech codec; Transcoding functions) and related specifications such as 26.190 (AMR wideband) and 26.290 (AMR-WB+), employs fractional lags in its pitch analysis and synthesis modules. The encoder computes the open-loop pitch estimate, then refines it with a closed-loop search around the integer lag using fractional resolution. The fractional lag values are quantized and transmitted to the decoder, which uses them to generate the adaptive codebook contribution. The decoder interpolates the past excitation signal at the fractional lag position to produce the pitch-shifted excitation. This process is critical for maintaining high speech quality at low bit rates, as the adaptive codebook (pitch predictor) is a major contributor to the overall speech signal.

The set of allowable fractional lag values is not continuous; it is a discrete set defined by the codec's quantization tables. For example, in the AMR half-rate mode, the fractional part of the lag may be quantized to 1/3 sample resolution, resulting in three possible fractional offsets (0, 1/3, 2/3) for each integer lag. However, not every fractional lag value considered in the encoder's search is an allowable lag value for transmission, because the quantization process restricts the set to those that can be efficiently encoded with the available bits. This distinction is important: the encoder may evaluate many fractional lags during its search, but only a subset can be signaled to the decoder.

Fractional lags are also used in other speech codecs, such as the Enhanced Variable Rate Codec (EVRC) and the Selectable Mode Vocoder (SMV), but the 3GPP specifications focus on GSM/AMR family. The concept is closely related to the adaptive codebook in CELP (Code-Excited Linear Prediction) coders, where the pitch lag determines the delay of the adaptive codebook vector. The use of fractional lags improves the accuracy of the pitch prediction, reduces the prediction error, and ultimately enhances the perceptual quality of the reconstructed speech. In the context of 3GPP, fractional lags are part of the speech codec specifications that ensure interoperability and high-quality voice services over GSM and UMTS networks.

The term has been stable across 3GPP releases from Rel-8 to Rel-19, with only minor editorial changes in the definition text. The specifications 26.090, 26.190, 26.290, 46.020, and 46.060 provide the normative details for the codecs that use fractional lags. Engineers implementing these codecs must carefully handle the interpolation and quantization of fractional lags to meet the conformance test requirements.

Purpose & Motivation

Fractional lags exist to overcome the limitations of integer-sample pitch period resolution in speech coding. In early speech codecs, pitch lags were restricted to integer multiples of the sampling period, which introduced quantization errors in the pitch prediction. These errors are particularly noticeable for speakers with high-pitched voices (e.g., women and children), where the pitch period is short and the relative error of integer lags is larger. The resulting speech may sound rough or have a metallic quality. Fractional lags provide a finer grid for pitch representation, reducing the quantization error and improving the accuracy of the adaptive codebook contribution.

The motivation for fractional lags is rooted in the need for high-quality speech at low bit rates. In GSM half-rate channels, the available bandwidth is limited, so efficient speech compression is essential. The adaptive codebook (pitch predictor) is one of the most important components of a CELP codec, and its performance directly affects the overall speech quality. By allowing sub-sample resolution, the encoder can better match the actual pitch period of the input signal, leading to a smaller prediction residual and more efficient coding of the remaining parameters. This translates to better speech quality at the same bit rate or the same quality at a lower bit rate.

Historically, fractional lags were introduced in the GSM half-rate speech codec as part of the AMR codec family, which was standardized in 3GPP Release 8 and later. The concept was not new—it had been used in other speech coding standards like the US DoD MELP and the ITU-T G.729—but its adoption in 3GPP ensured global interoperability for GSM and UMTS networks. The specifications 26.090 and 26.190 define the exact fractional lag resolution and quantization tables, which are mandatory for compliant implementations. Over successive releases, the definition of fractional lags has remained essentially unchanged, reflecting the maturity and stability of the speech codec technology.