Weighted Linear Prediction for Harmonic Bias Suppression

22 concepts10 papers4 proposals

In speech processing, accurately separating the sound source (the vocal folds) from the acoustic filter (the vocal tract) is a foundational challenge. The standard mathematical tool for estimating this vocal tract filter is Linear Prediction (LP). LP models the vocal tract as an all-pole digital filter, assuming that the input driving the filter has a flat spectral envelope, such as white noise or a single isolated impulse.

While this assumption holds reasonably well for unvoiced speech (like the "sh" sound), it breaks down catastrophically for voiced speech (like vowels). In voiced speech, the vocal folds vibrate periodically, producing a glottal source signal characterized by a fundamental frequency ($F_0$) and a series of sharp, discrete harmonics. As noted by Hideki Kawahara et al. in Beyond bandlimited sampling of speech spectral envelope imposed by the harmonic structure of voiced sounds, this periodic nature imposes a discrete harmonic structure that biases standard LP coefficients.

The Problem: Harmonic Biasing

When standard LP is applied to a voiced speech signal, the optimization algorithm minimizes the squared prediction error uniformly across the entire analysis frame. Because the true speech spectrum contains sharp harmonic peaks separated by deep valleys, the LP model is heavily penalized if it misses a high-energy harmonic peak.

As a result, the estimated spectral envelope suffers from harmonic biasing: the filter's frequency response is forcibly "pulled" toward individual harmonic peaks rather than smoothly tracing the underlying formant resonances of the vocal tract. As Carlo Magi et al. demonstrate in Stabilised weighted linear prediction - a robust all-pole method for speech processing, this bias degrades the representation of the true vocal tract characteristics.

Standard LP:  [Harmonic Peak]  [Harmonic Peak]  [Harmonic Peak]
                     \               /                /
                      \             /                /
Estimated Envelope: ---*-----------*----------------*--- (Pulled to peaks)
True Vocal Tract:   _____________________________________ (Smooth Formant)

This problem exacerbates significantly in high-pitched speech, such as the voices of women or children. In A toolkit for voice inverse filtering and parametrisation, Matti Airas et al. explain that a higher pitch means the harmonics are spaced wider apart in frequency. This sparse sampling makes it incredibly easy for analysis tools to lock onto individual harmonics instead of true resonant frequencies.

Aníbal J. S. Ferreira confirms this vulnerability in New signal features for robust identification of isolated vowels, showing that high pitches cause traditional formant-tracking tools to lose discrimination power. The lower formants (such as $F_1$) are especially prone to this distortion because they sit closest to the high-energy lower harmonics, as detailed by Paavo Alku et al. in Improved formant frequency estimation from high-pitched vowels by downgrading the contribution of the glottal source with weighted linear prediction.

The Intuition: The Free-Ringing Vocal Tract

To eliminate harmonic bias, we must decouple the vocal tract filter estimation from the glottal source excitation. The key to doing this lies in the time domain, specifically within the mechanics of the human glottal cycle.

A single cycle of vocal fold vibration can be divided into three main phases:

  1. The Open Phase: The vocal folds open, and air flows from the lungs into the vocal tract.
  2. The Glottal Closure Instant (GCI): The vocal folds snap shut, creating a sharp acoustic excitation pulse.
  3. The Closed Phase: The vocal folds remain completely closed.

During the closed phase, a remarkable physical phenomenon occurs: the vocal tract is completely isolated from the subglottal system (the lungs). Because the vocal folds are shut, no new acoustic energy enters the cavity. The air column inside the mouth and throat simply rings down freely, vibrating exclusively at its natural resonant frequencies.

Mathematically, during this brief closed window, the speech signal behaves exactly like the ideal, unexcited all-pole system that standard LP assumes. If we can isolate our analysis to this closed phase, the glottal source's harmonic influence vanishes, allowing us to extract the true, unbiased vocal tract envelope.

Mechanism: Weighted Linear Prediction (WLP)

To exploit this time-domain intuition, researchers transitioned from standard LP to Weighted Linear Prediction (WLP). Instead of treating every speech sample equally, WLP introduces a time-varying weighting function, $w_n$, into the error minimization objective:

$$\text{Error} = \sum_{n} w_n \left( x_n - \sum_{k=1}^{p} a_k x_{n-k} \right)^2$$

By strategically designing $w_n$, we can control which parts of the glottal cycle dictate the filter coefficients. To suppress the glottal source, Paavo Alku et al. introduced the Attenuated Main Excitation (AME) weighting function. The AME function assigns a very low weight to samples immediately surrounding the GCI and the open phase, while assigning a high weight to samples located within the glottal closed phase. By down-weighting the heavily driven portions of the speech signal, the model focuses almost entirely on the free-decaying resonances of the closed phase, effectively bypassing the harmonic bias.

Scaling to High Pitches: Quasi-Closed-Phase (QCP) Analysis

While pure closed-phase analysis works beautifully for low-pitched male voices, it encounters a mathematical barrier in high-pitched speech. As the fundamental frequency rises, the duration of the glottal cycle shrinks, leaving a closed phase that is exceptionally brief. A standard closed-phase analysis window becomes too short to contain enough audio samples to reliably solve the linear prediction matrix equations, resulting in severe numerical instability.

To overcome this limitation, Manu Airaksinen et al. developed Quasi-Closed-Phase (QCP) analysis in Quasi closed phase analysis for glottal inverse filtering. Instead of strictly truncating the analysis window to the closed phase, QCP uses a smooth, continuous AME weighting function that spans across multiple glottal cycles.

As explained by Anna Sfakianaki et al. in Investigating Voice Function Characteristics of Greek Speakers with Hearing Loss Using Automatic Glottal Source Feature Extraction, QCP allows the algorithm to incorporate more speech samples within each analysis frame while still heavily discounting the samples affected by the main glottal excitation. This provides a mathematically robust, stable filter estimate even when the glottal closed phase is short or incomplete.

Glottal Signal:   /\        /\        /\        /\
_________________/  \______/  \______/  \______/  \______
                  ^         ^         ^         ^
                 GCI       GCI       GCI       GCI

QCP Weight (w_n):__         __         __         __
                 \ \_______/ / \_______/ / \_______/ / \_
                  \_________/   \________/   \_______/
                 (High weight during closed phase, low at GCI)

To handle continuous, dynamic speech where formants shift rapidly over time, Dhananjaya Gowda et al. introduced Time-Varying Quasi-Closed-Phase (TVQCP) analysis in Time-Varying Quasi-Closed-Phase Weighted Linear Prediction Analysis of Speech for Accurate Formant Detection and Tracking. TVQCP integrates the QCP weighting function with time-continuity constraints, allowing the system to smoothly track formants across changing glottal cycles without succumbing to sudden jumps caused by harmonic locking.

Applications and Alternative Domains

The capacity to cleanly separate the source from the filter via harmonic suppression has proven vital across speech technology applications:

  • Speech Synthesis: In GlottDNN - A Full-Band Glottal Vocoder for Statistical Parametric Speech Synthesis, the authors utilize closed-phase weighted linear prediction to isolate high-fidelity vocal tract filters, preventing the "buzzy" artifacts common to traditional vocoders and enabling highly natural deep-learning-based speech synthesis.
  • Perceptual Modeling: The principle of harmonic suppression is not restricted to the time domain. In An Auditory Model of Speaker Size Perception for Voiced Speech Sounds, Toshio Irino et al. implement a spectral weighting function applied directly to an auditory spectrogram. This function selectively down-weights the energy of lower glottal harmonics, allowing an auditory model to accurately mimic how the human brain isolates vocal tract resonances to perceive speaker size.

By shifting the analytical focus away from the disruptive moments of glottal excitation, harmonic suppression weighting transforms linear prediction from a brittle mathematical curve-fit into a robust reflection of physical speech production.

Key Foundational Work

Stabilised weighted linear prediction - a robust all-pole method for speech processing (2007) — contributed 1 concept elements to this direction.
A toolkit for voice inverse filtering and parametrisation (2005) — contributed 1 concept elements to this direction.
New signal features for robust identification of isolated vowels (2005) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_harmonic_suppression_weighting_14b057fa,
  title = {harmonic suppression weighting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/harmonic-suppression-weighting/rdr_14b057fa0ac24362b5426b3da4c71316},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_harmonic_suppression_weighting_14b057fa_v1,
  title = {harmonic suppression weighting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/harmonic-suppression-weighting/rdr_14b057fa0ac24362b5426b3da4c71316/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake