Unsupervised modulation filter learning for noise-robust speech recognition.

Area of Science:

• Signal Processing
• Machine Learning
• Speech Recognition

Background:

• Robust automatic speech recognition (ASR) is challenged by noise and reverberation.
• Traditional modulation filtering enhances relevant spectral regions but requires manual design.
• Data-driven approaches offer potential for optimizing modulation filters.

Purpose of the Study:

• To propose a data-driven, unsupervised method for learning modulation filters for robust ASR.
• To investigate the effectiveness of these learned filters on noisy and reverberant speech.
• To explore the application of this method in semi-supervised learning scenarios.

Main Methods:

• A convolutional restricted Boltzmann machine (CRBM) was employed for unsupervised filter learning.
• Initial filters were learned from speech spectrograms.
• Subsequent filters were learned from residual spectrograms to refine the process.

Main Results:

• Modulation filtered spectrograms significantly improved ASR performance on noisy and reverberant speech.
• The proposed features outperformed other robust features in experimental evaluations.
• The method demonstrated potential for semi-supervised learning applications.

Conclusions:

• The proposed unsupervised modulation filter learning scheme effectively enhances speech features for robust ASR.
• This data-driven approach offers a significant improvement over traditional methods in adverse acoustic conditions.
• The technique shows promise for advancing semi-supervised learning in speech recognition.