Extended weighted linear prediction (XLP) analysis of speech and its application to speaker verification in adverse conditions

Jouni Pohjalainen, Rahim Saeidi, Tomi Kinnunen, Paavo AlkuInterspeech 2010

TL;DR

This paper introduces Extended Weighted Linear Prediction (XLP), a novel method for speech analysis that improves robustness in speaker verification. By weighting individual speech samples differently, XLP and its stabilized version (SXLP) significantly outperform traditional methods like FFT and LP, especially in noisy and channel-distorted conditions.

What Problem Does This Paper Address?

Speaker verification systems, which aim to identify individuals based on their voice, are notoriously susceptible to real-world challenges. These include background noise and variations in transmission channels (like different phone lines or microphones), collectively known as "mismatch." Traditional methods for analyzing speech signals, such as the Fast Fourier Transform (FFT) and standard Linear Prediction (LP), struggle to maintain accuracy when the conditions during recognition differ from those during training. While previous work introduced temporally weighted LP (WLP) to mitigate these issues by emphasizing less corrupted speech segments, this paper addresses the need for an even more refined approach to spectrum analysis that can adapt to varying levels of distortion across different parts of the speech signal.

Methods

The core innovation is the Extended Weighted Linear Prediction (XLP) method. Unlike conventional LP, which treats all speech samples equally, or WLP, which applies a single temporal weight to entire frames, XLP allows for individual weighting of each lagged speech sample within the prediction model. This is achieved by introducing a two-dimensional weighting function, $Z_{n,j}$, where $n$ represents the time frame and $j$ represents the lag.

The paper also introduces a specific weighting scheme called Absolute Value Sum (AVS), which emphasizes prominent signal samples and their corresponding lags. This is motivated by the idea that louder speech segments are likely less corrupted by noise. The XLP method aims to minimize a weighted prediction error energy, leading to a new set of normal equations for calculating the predictor coefficients.

Furthermore, the authors present a stabilized version of XLP (SXLP), which ensures the stability of the resulting synthesis model, a crucial aspect for reliable signal processing. Both XLP and SXLP are then integrated into the Mel-Frequency Cepstral Coefficients (MFCC) feature extraction pipeline of a standard Gaussian Mixture Model-Universal Background Model (GMM-UBM) speaker verification system.

Datasets

The evaluation was performed on the NIST 2002 speaker recognition evaluation (SRE) corpus. This corpus contains speech samples from various cellular networks and handsets, introducing realistic channel distortion and handset variations. To further test robustness against additive noise, speech samples were augmented with noise from the NOISEX-92 database, specifically the "factory2" noise. The experiments involved training speaker models on clean data and testing on data with varying levels of channel distortion and added noise (including an "original" condition representing the unadulterated NIST data).

Key Insights and Ideas

  • Finer-grained weighting is beneficial: The ability of XLP to weight individual lagged samples, rather than just entire frames, provides a more nuanced approach to handling signal corruption.
  • Prominent samples are more reliable: The AVS weighting scheme, which prioritizes louder speech segments, proved effective, suggesting that these segments are indeed less affected by noise.
  • Stabilization matters: The SXLP method, which builds upon XLP by ensuring model stability, consistently performed as well as or better than XLP, highlighting the importance of a stable synthesis model.
  • Beyond additive noise: While previous work focused on additive noise, this paper demonstrates the effectiveness of XLP/SXLP in scenarios with channel distortion, a common challenge in real-world speaker verification.

Key Results

The XLP and SXLP methods showed significant improvements in speaker verification performance compared to baseline FFT and LP methods across various adverse conditions.

  • In scenarios with channel distortion and low to moderate additive noise, SXLP emerged as the overall best-performing method, closely followed by XLP.
  • Both XLP and SXLP consistently outperformed the FFT baseline in terms of Equal Error Rate (EER) across all evaluated mismatch scenarios.
  • For instance, at 0 dB SNR (significant additive noise), the EER for XLP and SXLP was around 16.68% and 16.58% respectively, compared to 18.34% for FFT.
  • Even in heavily corrupted conditions (0 dB SNR), the weighted LP models (WLP, SWLP, XLP, SXLP) significantly outperformed the baseline FFT and LP models by a wide margin.

Limitations and Future Work

  • Weighting scheme exploration: The authors suggest that future work could focus on searching for new two-dimensional weighting schemes for XLP, potentially leading to further performance gains.
  • Application to new tasks: The paper proposes applying these novel methods to other speech processing tasks beyond speaker verification.
  • Parameter tuning: While specific parameters (like the averaging parameter $m=20$ for AVS and window duration $M=20$ for STE) were used, further investigation into optimal parameter selection for different conditions could be beneficial.

Summary updated

BibTeX
@inproceedings{pohjalainen2010extended,
  author = {Jouni Pohjalainen and Rahim Saeidi and Tomi Kinnunen and Paavo Alku},
  title = {{Extended weighted linear prediction (XLP) analysis of speech and its application to speaker verification in adverse conditions}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure