Traditional Linear Prediction (LP) is a cornerstone of speech processing, widely used to estimate the spectral envelope of the vocal tract. It operates by modeling speech as the output of an all-pole filter, minimizing the sum of squared differences between actual speech samples and their predicted values.
However, standard LP has a major vulnerability: it treats every signal sample equally. If a specific segment of speech is corrupted by additive noise, or if the signal contains sharp, high-energy glottal pulses, the prediction error at those exact moments will spike. Because standard LP minimizes the squared error, it over-indexes on these large outliers, warping the resulting spectral envelope and miscalculating vital features like formant frequencies.
To overcome this, researchers developed Weighted Linear Prediction (WLP). The core intuition is simple: instead of treating every error equally, we can introduce a time-varying weight to the optimization process. By downweighting regions of the signal that are noisy or unrepresentative of the vocal tract, and upweighting clean or highly informative regions, we can achieve a far more robust spectral estimate.
The Mechanics of Temporal Weighting
In standard LP, we minimize the residual energy $E = \sum e_n^2$. In WLP, as described by Airaksinen et al. (2013), the error function is modified by a temporal weighting function $w_n$:
$$E_{\text{weighted}} = \sum w_n e_n^2$$
By altering $w_n$, the model can be dynamically steered to focus on the most reliable parts of the signal. The choice of the weighting function depends entirely on the acoustic challenge at hand:
- Combating Additive Noise: In noisy environments, the signal-to-noise ratio (SNR) fluctuates rapidly. Pohjalainen et al. (2009) demonstrated that using the Short-Time Energy (STE) of the speech signal as a weighting function allows the model to prioritize high-energy, high-SNR regions, significantly reducing automatic speech recognition (ASR) error rates in severe noise.
- High-Pitched Vowels: When analyzing high-pitched speech (such as female or child voices), the fundamental frequency ($F_0$) can interfere with formant estimation. Alku et al. (2012) introduced WLP with Attenuated Main Excitation (WLP-AME), which applies a specialized weighting function to downweight the samples immediately following a glottal opening. This prevents the strong glottal pulse from blinding the model to the underlying vocal tract resonances.
- Glottal Inverse Filtering: To isolate the source air flow from the vocal tract, Airaksinen et al. (2013) utilized WLP to emphasize the glottal closed phase—the brief window where the vocal folds are shut and the vocal tract is free from source excitation—leading to highly accurate glottal flow estimations.
The Stability Crisis and SWLP
While basic WLP provides excellent spectral representations, it introduces a dangerous mathematical flaw: it does not guarantee filter stability.
In standard LP, using the autocorrelation method ensures that all estimated poles lie strictly inside the unit circle, meaning the filter is stable and will not blow up during synthesis. As Magi et al. (2007) pointed out, introducing arbitrary temporal weights destroys this inherent mathematical guarantee. The resulting filters can easily become unstable, making basic WLP unusable for applications like speech coding or synthesis where guaranteed filter convergence is mandatory.
To resolve this, Magi et al. (2007) and Pohjalainen et al. (2009) formulated Stabilized Weighted Linear Prediction (SWLP). SWLP modifies the underlying normal equations of the optimization process, embedding a stabilization mechanism that mathematically forces the estimated poles to remain inside the unit circle. This modification delivers the noise-robust benefits of temporal weighting while restoring the absolute reliability required for speech synthesis and coding.
Expanding to Two Dimensions: Extended WLP
Standard WLP applies a one-dimensional weight based solely on the time index of the speech sample. However, linear prediction inherently operates on pairs of samples: the current sample being predicted and the historical, lagged samples used to predict it.
To exploit this relationship, Pohjalainen et al. (2010) and Keronen et al. (2011) proposed Extended Weighted Linear Prediction (XLP). Instead of a simple weighting vector, XLP introduces a two-dimensional temporal weighting matrix. This allows the optimization framework to independently weight both the current time instant and the specific lagged samples contributing to the prediction error.
By providing finer control over which signal components influence the model, XLP offers even greater flexibility in adapting to non-stationary noise. Just like its predecessor, XLP can be stabilized (SXLP) to ensure the resulting synthesis models remain perfectly stable under adverse conditions (Pohjalainen et al., 2010).
Why It Works: Applications in Feature Extraction
By decoupling the spectral envelope from localized noise spikes and glottal pulse timing, temporally weighted linear prediction yields smoother, more reliable spectral profiles. Jokinen et al. (2014) found that this reduced sensitivity to glottal timing makes SWLP an exceptional tool for speech transformation and intelligibility enhancement algorithms.
Furthermore, this robustness translates directly to improved feature extraction for speaker verification and formant detection:
- Vocal Effort Mismatch: When a speaker's vocal effort changes between training and testing (e.g., whispering vs. shouting), traditional spectrum estimators fail because the glottal source characteristics shift dramatically. Hanilçi et al. (2013) demonstrated that SWLP mitigates this mismatch, outperforming standard methods by focusing on the stable vocal tract characteristics rather than the volatile glottal excitation.
- Phase Information: Beyond magnitude spectra, Rajan et al. (2013) leveraged SWLP to extract robust group delay functions, capturing underutilized phase information that remains stable even under high vocal effort.
- Formant Tracking: By combining the temporal robustness of SWLP with the high spectral resolution of group delay functions, the authors in Gowda et al. (2013) created the SWLP-GD spectrum, which isolates vocal tract resonances with high precision, even when the speech is heavily corrupted by noise.
Key Foundational Work
Weighted linear prediction for speech analysis in noisy conditions (2009) — contributed 7 concept elements to this direction.
Stabilised weighted linear prediction - a robust all-pole method for speech processing (2007) — contributed 1 concept elements to this direction.