A correlational discriminant approach to feature extraction for robust speech recognition

Vikrant Singh Tomar, Richard C. RoseInterspeech 2012

TL;DR

This paper introduces a new feature extraction method called Correlational Discriminant Analysis (CPDA) for robust speech recognition. CPDA uses a correlation-based distance measure instead of the traditional Euclidean distance, aiming to improve performance in noisy environments by better preserving the relationships between speech features. Experiments show CPDA offers significant improvements, especially in high-noise conditions.

What Problem Does This Paper Address?

Automatic Speech Recognition (ASR) systems often struggle with noisy environments. A key challenge is feature extraction, where raw speech signals are transformed into a more manageable representation. Traditional methods often concatenate short-term speech features into "super-vectors," leading to high dimensionality and strong correlations between feature components. To address this, dimensionality reduction techniques are employed. However, many existing methods, like Linear Discriminant Analysis (LDA) and Locality Preserving Projections (LPP), rely on Euclidean distance. This is problematic because additive noise can distort the magnitude of speech features (like cepstrums) but has less impact on their correlation. This means Euclidean-based methods can fail to capture the underlying structure of noisy speech, leading to recognition errors. This paper aims to develop a feature extraction method that is inherently more robust to noise by using a correlation-based measure.

Methods

The paper proposes Correlational Discriminant Analysis (CPDA), a nonlinear discriminant approach for feature space dimensionality reduction. CPDA builds upon the framework of graph embedding, which represents data relationships using weighted graphs. Specifically, it defines two graphs:

  • Intrinsic Graph ($G_i$): Connects data points within the same class that are close in terms of cosine-correlation. This aims to preserve within-class local manifold relationships.
  • Penalty Graph ($G_p$): Connects data points that are not in the same class but have high cosine-correlation. This aims to maximize separability between classes.

The core novelty lies in using a cosine-correlation based distance measure instead of Euclidean distance to define the "closeness" or "similarity" between feature vectors. This measure is motivated by the observation that feature correlations are more robust to noise than their magnitudes.

The CPDA algorithm aims to optimize a function that maximizes the scatter of the penalty graph while minimizing the scatter of the intrinsic graph. This optimization is performed using a gradient ascent rule. For initialization, a closed-form solution derived from an eigenvalue problem is used, similar to previous manifold learning techniques.

Datasets

The experiments were conducted on the ETSI Aurora-2 speech corpus. This corpus simulates speech in noise by adding noise from various environments to clean speech.

  • Training Data: 8440 multi-noise mixed utterances from 110 speakers.
  • Test Data: 4004 utterances corrupted by four different noise types (subway, car, exhibition hall, airport) at Signal-to-Noise Ratios (SNRs) ranging from 5 to 20 dB.

The paper does not introduce new benchmarks but uses a well-established noisy speech dataset for evaluation.

Key Insights and Ideas

  • Correlation is King in Noise: The central idea is that the correlation between speech features is more resilient to additive noise than their absolute magnitudes. This makes correlation-based measures a promising avenue for robust ASR.
  • Beyond Euclidean Distance: Traditional ASR feature extraction heavily relies on Euclidean distance. CPDA demonstrates that moving to correlation-based measures can unlock significant performance gains in noisy conditions.
  • Combining Discrimination and Manifold Learning: CPDA effectively merges the strengths of discriminant analysis (maximizing class separation) with manifold learning (preserving local data structure), but with a noise-robust distance metric.
  • Graph Embedding as a Flexible Framework: The use of graph embedding provides a unified framework to incorporate both within-class similarity and between-class dissimilarity, allowing for flexible objective function design.

Key Results

The paper presents results in terms of Word Error Rate (WER) on the Aurora-2 task.

  • CPDA consistently outperforms traditional methods like LDA and LPP across various noise types and SNRs.
  • CPDA shows a notable advantage over its predecessor, LPDA (which uses Euclidean distance), particularly in high-noise conditions (lower SNRs).
  • For example, in the "Subway" noise at 5 dB SNR, CPDA achieved a WER of 11.24%, significantly lower than LDA (12.32%), LPP (13.26%), and LPDA (11.73%).
  • CPDA exhibits the smallest drop in ASR performance as the noise level increases, indicating superior robustness.
  • The car noise type was an exception where all techniques performed inefficiently, suggesting specific challenges for that noise environment.

Limitations and Future Work

  • Computational Complexity: The authors acknowledge that CPDA, like other manifold learning approaches (LPDA, LPP), suffers from high computational complexity. The gradient ascent optimization for CPDA adds to this burden compared to LPDA.
  • Kernel Parameter Optimization: The performance of CPDA can be sensitive to the kernel scale parameter ($\rho$). While optimal values were estimated for LPP and LPDA, this was not fully explored for CPDA in this work, potentially explaining some instances where CPDA was slightly outperformed by LPDA. Future work could involve more thorough optimization of this parameter.
  • Correlation in Transformed Space: The paper notes that correlation among projected feature space vectors is a common issue for all feature transformations. Techniques like Maximum Likelihood Linear Transformation (MLLT) or Semi-Tied Covariances (STC) are still needed to handle this for CDHMM-based ASR systems.
  • Further Exploration of Noise Types: While CPDA shows promise, its performance across a wider range of noise types and real-world acoustic conditions could be further investigated.

Summary updated

BibTeX
@inproceedings{tomar2012correlational,
  author = {Vikrant Singh Tomar and Richard C. Rose},
  title = {{A correlational discriminant approach to feature extraction for robust speech recognition}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure