An environment model-based robust speech recognition

Lei Zhang, Jiqing Han, Chengguo Lv, Chengfa WangInterspeech 2000 · DOI

TL;DR

This paper introduces a novel "Environmental Discrimination Learning" (EDL) approach to make speech recognition systems more robust to real-world noise and channel distortions. By modeling and learning the effects of these environmental factors, EDL significantly improves recognition accuracy, especially in challenging acoustic conditions.

What Problem Does This Paper Address?

Speech recognition systems often struggle in real-world environments due to the presence of various noises, such as additive background sounds and distortions introduced by communication channels. Traditional methods often assume that noisy speech follows a Gaussian distribution, which doesn't always hold true when environmental effects are significant. This paper tackles the challenge of improving speech recognition robustness by explicitly modeling and learning these environmental distortions, aiming to bridge the gap between clean speech recognition performance and performance in noisy, real-world scenarios.

Methods

The core of the proposed method is Environmental Discrimination Learning (EDL). This approach first models the environmental distortions (additive noise and linear channel filtering) in the cepstral domain using a simplified model inspired by Acero. The key innovation lies in how the environmental parameters are learned. Instead of traditional methods, EDL employs a Minimum Classification Error (MCE) criterion combined with a Generalized Probabilistic Descent (GPD) algorithm. This allows the system to discriminatively train the environmental parameters by considering information across all classes, aiming to minimize the overall error rate. The process involves:

  1. Environment Modeling: Representing noisy speech features as a function of clean speech features, environmental parameters (noise statistics, channel transfer function), and additive noise. A vector Taylor series (VTS) approach is used to simplify this non-linear relationship in the cepstral domain.
  2. Discriminant Function: Defining a discriminant function based on Gaussian distribution assumptions for clean speech segments.
  3. Misclassification Measure: Developing a measure that quantifies how incorrectly a speech segment is classified.
  4. Cost Function: Embedding the environmental parameters within a cost function that is optimized using MCE.
  5. GPD Algorithm: Iteratively updating the environmental parameters using the GPD algorithm to minimize the cost function.

Once optimized, the learned environmental parameters are used to transform noisy speech features into the clean speech feature space, which are then fed into a standard HMM-based recognizer.

Datasets

The paper does not explicitly name specific public datasets. However, it describes the experimental setup:

  • Vocabulary: A vocabulary of 208 isolated words was used.
  • Confusable Words: For training the environmental parameters, a subset of 18 "confusable words" was selected from the larger vocabulary. This is a common practice to focus the discriminative training on the most challenging cases.
  • Features: Speech features consisted of 39 elements: 12 cepstral coefficients, 12 delta-cepstral coefficients, 12 delta-delta-cepstral coefficients, log energy, delta log energy, and delta-delta log energy.

Key Insights and Ideas

  • Learning Environmental Effects: The most significant insight is that explicitly modeling and learning the parameters of environmental noise and channel distortions can lead to substantial improvements in speech recognition robustness.
  • Discriminative Training for Environment: Applying MCE/GPD for training environmental parameters, rather than just acoustic model parameters, allows for a more direct optimization towards reducing recognition errors.
  • Cross-Class Optimization: The EDL approach leverages information from all classes to optimize parameters for a specific class, which is a powerful way to improve discrimination.
  • Cepstral Domain Simplification: Extending VTS to the cepstral domain effectively simplifies the complex environmental model, making it computationally tractable for training.

Key Results

The experimental results demonstrate the effectiveness of the EDL approach:

  • Iterative Improvement: The recognition rate improved significantly with each iteration of the EDL training process. Starting at 74% on a test set (85% on training), it reached 92.6% on the test set (96.8% on training) after 10 iterations, indicating convergence.
  • Significant Error Reduction: Compared to a "bench mark" (likely a system without EDL), the EDL approach achieved substantial reductions in word error rate across different Signal-to-Noise Ratios (SNRs):
  • 30dB SNR: 58.9% error rate reduction.
  • 20dB SNR: 37.5% error rate reduction.
  • 10dB SNR: 33.9% error rate reduction.
  • Overall Effectiveness: On average, the EDL method reduced the error rate by approximately 43.4%, confirming its effectiveness in improving robust speech recognition.

Limitations and Future Work

  • Limitations:
    • The paper uses a simplified model of environmental degradation (additive stationary noise and linear filtering). Real-world environments can be much more complex.
  • The method is evaluated on isolated word recognition. Its performance on continuous speech is not discussed.
  • The paper focuses on a specific set of features (cepstral coefficients and their derivatives).
  • Future Work: The authors suggest that the EDL approach can be extended to continuous speech recognition and potentially applied to other types of environmental distortions. The paper implies that further refinement of the environment model could yield even better results.

Summary updated

BibTeX
@inproceedings{zhang2000environment,
  author = {Lei Zhang and Jiqing Han and Chengguo Lv and Chengfa Wang},
  title = {{An environment model-based robust speech recognition}},
  booktitle = {Interspeech},
  year = {2000},
  doi = {10.21437/icslp.2000-604},
}
Figure