Group sparse hidden Markov models for speech recognition

Jen-Tzung Chien, Cheng-Chun ChiangInterspeech 2012

TL;DR

This paper introduces Group Sparse Hidden Markov Models (GS-HMMs) that leverage a novel Bayesian learning approach to represent speech features more efficiently. By using common and individual basis vectors with a sparse prior, GS-HMMs achieve improved robustness and accuracy in noisy speech recognition tasks.

What Problem Does This Paper Address?

Traditional Hidden Markov Models (HMMs) used in speech recognition can struggle with noisy environments and varying acoustic conditions. While sparse representation techniques have shown promise in pattern recognition by selecting relevant features, applying them effectively to sequential data like speech has been challenging. Existing methods, like Bayesian sensing HMMs, haven't fully exploited the potential of sparse priors for basis representation. This paper aims to bridge this gap by developing a group sparse representation for sequential data that is more robust to noise and better captures the underlying speech characteristics.

Methods

The core innovation is the Group Sparse Hidden Markov Model (GS-HMM). This model extends standard HMMs by representing each acoustic feature vector using two sets of basis vectors:

  • Common bases: Shared across different states within an HMM, capturing general speech characteristics.
  • Individual bases: Specific to each state, accounting for state-specific residual information.

The weights associated with these bases are learned using a Bayesian learning framework. A key component is the use of a Laplacian scale mixture (LSM) distribution as a sparse prior. This distribution, which is a product of a Laplace variable and a Gamma variable, is designed to induce even greater sparsity than a standard Laplace prior. This sparsity helps in selecting the most relevant bases, a process known as Automatic Relevance Determination (ARD). The estimation of both the sensing weights and the basis vectors is performed using an Expectation-Maximization (EM) algorithm, with specific adaptations for group sparsity.

Datasets

The experiments were conducted using the Aurora2 database, which consists of English digits spoken in various noise environments. The database includes clean speech and noisy speech with different noise types (subway, babble, car, exhibition) and signal-to-noise ratios (SNRs) ranging from -5 dB to 20 dB.

Key Insights and Ideas

  • Group Sparsity for Sequential Data: The paper successfully extends the concept of group sparsity, previously applied to static data, to sequential data like speech, demonstrating its effectiveness.
  • Laplacian Scale Mixture (LSM) for Enhanced Sparsity: Using LSM as a prior distribution leads to even sparser representations compared to standard Laplace priors, which is crucial for effective feature selection in noisy conditions.
  • Common vs. Individual Bases: The distinction between common and individual bases allows the model to capture both general acoustic patterns and state-specific nuances, leading to a more comprehensive representation.
  • Robustness to Noise: The proposed GS-HMMs show significant improvements in recognition accuracy under noisy conditions, highlighting their ability to handle acoustic variability.

Key Results

The GS-HMMs demonstrated significant improvements in noisy speech recognition accuracy compared to baseline HMMs.

  • Under the clean training (CT) condition, the average accuracy for noisy speech increased from 62.9% to 71.0%.
  • Under the multi-conditional training (MT) condition, the average accuracy improved from 81.3% to 86.2%.
  • The improvements were more pronounced at lower SNRs, indicating the model's strength in challenging acoustic environments.
  • Histograms of sensing weight sparsity showed that weights in transition states and under multi-conditional training were generally sparser, supporting the model's ability to adapt to varied conditions.

Limitations and Future Work

  • Further Noise Robustness: The authors suggest that incorporating other noise-robust schemes into GS-HMMs could lead to further improvements.
  • Hyperparameter Tuning: While hyperparameters were selected using validation data, the paper implies that optimal tuning is crucial for performance.
  • Computational Complexity: While not explicitly detailed as a limitation, Bayesian learning and EM algorithms can be computationally intensive, which might be a consideration for real-time applications. Future work could explore more efficient estimation methods.

Summary updated

BibTeX
@inproceedings{chien2012group,
  author = {Jen-Tzung Chien and Cheng-Chun Chiang},
  title = {{Group sparse hidden Markov models for speech recognition}},
  booktitle = {Interspeech},
  year = {2012},
}
Figure