Closed-Loop Pitch Analysis (Adaptive Codebook Search)

Description

Closed-loop pitch analysis is a core component of the adaptive codebook search used in 3GPP speech codecs such as AMR-NB (TS 26.090), AMR-WB (TS 26.190), and the extended AMR-WB codec (TS 26.290). The process estimates the pitch (lag) value from the weighted input speech signal and the long-term filter state. Unlike open-loop pitch estimation, which provides a coarse estimate based on signal correlation, closed-loop analysis refines this estimate using an error minimization loop in an analysis-by-synthesis framework. This means the encoder synthesizes candidate signals and compares them to the original to find the lag that minimizes perceptual error.

In the adaptive multi-rate wideband codec, closed-loop pitch search is performed for every subframe, allowing fine-grained tracking of pitch variations within a frame. The search typically evaluates a limited range of lag candidates around an open-loop estimate, reducing computational complexity while maintaining high accuracy. The chosen lag drives the adaptive codebook, which models the periodic (voiced) component of speech. The error minimization loop computes the weighted synthesis error for each candidate lag and selects the one yielding the lowest error, often using a fractional pitch resolution to improve quality.

The adaptive codebook search is integral to the codec's ability to reproduce voiced speech efficiently. By accurately estimating the pitch period, the codec can represent the long-term correlation of speech with a single lag parameter and a gain, rather than encoding the waveform directly. This significantly reduces bitrate while preserving intelligibility and naturalness. The closed-loop approach ensures that the selected lag aligns with the perceptual weighting filter, which emphasizes formant regions and de-emphasizes spectral valleys, leading to better subjective quality.

Closed-loop pitch analysis is specified in detail in 3GPP TS 26.090 (AMR speech codec; transcoding functions), TS 26.190 (AMR-WB speech codec; transcoding functions), TS 26.290 (Extended AMR-WB codec; transcoding functions), and TS 46.060 (AMR speech codec; comfort noise aspects). These specifications define the exact algorithms, including the computation of the impulse response, the target signal, and the search procedure. The technique has remained stable across releases from Rel-8 to Rel-19, reflecting its maturity and effectiveness.

In the broader context of 3GPP services, closed-loop pitch analysis enables high-quality voice communication at low bitrates, which is essential for mobile networks where bandwidth is limited. It is a key enabler for AMR and AMR-WB codecs, which are widely deployed in GSM, UMTS, and LTE networks. The analysis-by-synthesis paradigm also underpins other codec components, such as the algebraic codebook search for the fixed codebook excitation. Understanding closed-loop pitch analysis is therefore fundamental for engineers working on speech coding and voice over IP (VoIP) in 3GPP systems.

Purpose & Motivation

Closed-loop pitch analysis exists to improve the accuracy of pitch estimation in speech codecs, which directly impacts voice quality and compression efficiency. Early speech codecs often used open-loop pitch estimation, which is computationally simple but less accurate because it does not account for the perceptual weighting and synthesis process. Open-loop methods estimate pitch from the raw signal, leading to errors that degrade reconstructed speech, especially in noisy conditions or for speakers with irregular pitch patterns. Closed-loop analysis solves this by embedding the pitch search within the analysis-by-synthesis loop, where candidate lags are evaluated based on the actual synthesized output and perceptually weighted error.

The motivation for closed-loop pitch analysis stems from the need to achieve toll-quality speech at low bitrates. In mobile communications, bandwidth is a scarce resource, so codecs must compress speech aggressively while maintaining intelligibility and naturalness. The adaptive codebook, which relies on accurate pitch information, models the periodic component of voiced speech. If the pitch lag is inaccurate, the adaptive codebook contribution is misaligned, causing artifacts such as roughness or tonal noise. Closed-loop analysis minimizes these artifacts by selecting the lag that best matches the original signal in the perceptually weighted domain.

Historically, the introduction of closed-loop pitch analysis in codecs like AMR (Rel-8 and earlier) marked a significant advancement over earlier standards such as GSM FR and EFR, which used simpler pitch estimation. The technique has been refined over releases, but its fundamental principles have remained unchanged, demonstrating its robustness. It continues to be a cornerstone of 3GPP speech codecs, ensuring high-quality voice services in evolving networks from 2G to 5G and beyond.