Data-driven temporal filters obtained via different optimization criteria evaluated on Aurora2 database

Jeih-Weih Hung, Lin-Shan LeeInterspeech 2002 · DOI

TL;DR

This paper explores how to create better speech recognition features by using "temporal filters" that learn from data. They found that by first cleaning up the speech features (Cepstral Normalization) and then applying filters optimized with Principal Component Analysis (PCA) or Minimum Classification Error (MCE), they achieved significant improvements in recognition accuracy, especially in noisy conditions.

What Problem Does This Paper Address?

Speech recognition systems often struggle when the audio they are trained on sounds different from the audio they encounter in real-world use. This "mismatch" can be caused by background noise, different microphones, or varying acoustic environments. Traditional methods like Cepstral Mean Subtraction (CMS) and Relative Spectral (RASTA) filtering help by smoothing out speech features over time, but they use fixed filters that aren't tailored to specific recognition tasks. The authors aim to develop "data-driven" temporal filters that can be optimized for a particular task or dataset, leading to more robust speech recognition. They also investigate a preprocessing step, Cepstral Normalization, to further enhance the effectiveness of these data-driven filters.

Methods

The core of this paper is the design and evaluation of data-driven temporal filters. The authors consider three optimization criteria for these filters:

  • Principal Component Analysis (PCA): This method aims to find a filter that maximizes the variance of the output signal in a one-dimensional space. It's applied independently to each feature dimension.
  • Linear Discriminant Analysis (LDA): Unlike PCA, LDA considers class information (e.g., different phonemes or digits) to find a filter that maximizes the separability between classes.
  • Minimum Classification Error (MCE): This criterion directly optimizes the filter to minimize the classification error of the speech recognition system. It uses a gradient-descent approach to iteratively refine the filter coefficients.

A key novelty introduced in this work is the Cepstral Normalization (CN) preprocessing step. Before applying any of the temporal filters, the authors first normalize the cepstral features to have zero mean and unit variance. This is hypothesized to remove slowly varying components that might be distorted by channel effects and could be amplified by the low-pass nature of the temporal filters. The temporal filters are then derived and applied to these normalized features. The filters are designed as L-sample Finite Impulse Response (FIR) filters.

Datasets

  • AURORA2 Database: This is the primary dataset used for both training and evaluation. It contains clean speech data and various noisy conditions (subway, babble, car, exhibition) at different Signal-to-Noise Ratios (SNRs). The authors specifically mention using the clean speech data for training the filters and the multi-condition testing sets (Sets A, B, and C) for evaluation.

Key Insights and Ideas

  • Cepstral Normalization is Crucial: The paper strongly suggests that applying Cepstral Normalization before temporal filtering is a critical step. It significantly boosts the performance of the data-driven filters, especially when dealing with mismatched conditions.
  • Data-Driven Filters Outperform Fixed Filters: The results indicate that filters optimized using PCA and MCE criteria, when applied after CN, offer substantial improvements over baseline MFCCs and even CN-only features.
  • PCA and MCE are Strong Contenders: Both PCA and MCE-derived temporal filters showed excellent performance, often outperforming LDA-derived filters, particularly in challenging noisy conditions.
  • Low-Pass Nature and Syllabic Rate: The temporal filters are observed to be low-pass, with cutoff frequencies around 6-12Hz. This range is thought to be effective because it aligns with the syllabic rate of speech (around 4Hz), potentially enhancing this important temporal information.
  • Mismatched Conditions are Key: The most significant gains from these data-driven temporal filters are observed under conditions where the training and testing environments are highly mismatched, especially at low SNRs.

Key Results

  • Significant Accuracy Improvements: The paper reports substantial improvements in word accuracy. For example, when using PCA- and MCE-derived temporal filters on cepstral-normalized features, they achieved improvements of up to 16% over original unnormalized features and 7% over cepstral-normalized features alone.
  • Robustness to Noise: The proposed methods, particularly PCA and MCE with CN, demonstrated significant improvements in recognition accuracy across various noise types and SNRs in the AURORA2 database.
  • PCA and MCE vs. LDA: While all data-driven filters showed some improvement, PCA and MCE-derived filters generally outperformed LDA-derived filters, especially in the more challenging low-SNR conditions.
  • Filter Characteristics: The frequency responses of the filters showed they are low-pass. PCA filters had the widest main lobe, LDA the narrowest, and MCE in between.

Limitations and Future Work

  • Limitations of Low-Pass Filters: The paper acknowledges that the low-pass nature of these temporal filters can amplify slowly varying components (below 1Hz) that correspond to convolutional channel distortion, potentially distorting speech features. Cepstral Normalization is proposed as a mitigation strategy.
  • Filter Length: The length of the temporal filters (L) was set to 10. The paper doesn't extensively explore the impact of different filter lengths.
  • Future Work: The authors suggest that further investigation into the optimal filter design and preprocessing steps could lead to even greater improvements. The effectiveness of these filters on other databases and for different speech tasks (e.g., speaker recognition) could also be explored.

Summary updated

BibTeX
@inproceedings{hung2002data,
  author = {Jeih-Weih Hung and Lin-Shan Lee},
  title = {{Data-driven temporal filters obtained via different optimization criteria evaluated on Aurora2 database}},
  booktitle = {Interspeech},
  year = {2002},
  doi = {10.21437/icslp.2002-14},
}
Figure