X-DC: Explainable Deep Clustering Based on Learnable Spectrogram Templates.

Area of Science:

• Speech processing
• Machine learning
• Signal processing

Background:

• Deep neural networks (DNNs) excel in speech processing, with deep clustering (DC) effectively handling monaural speech separation.
• A key limitation of DC is its black-box nature, hindering interpretability and adaptation to varying conditions like reverberation.

Purpose of the Study:

• To introduce explainable deep clustering (X-DC), an interpretable DNN architecture for speech separation.
• To address the interpretability and model adaptation limitations of existing deep clustering methods.

Main Methods:

• X-DC interprets network architecture as fitting learnable spectrogram templates to input spectrograms, followed by Wiener filtering.
• Nonnegativity constraints on spectrogram template elements and activations promote sparsity and interpretability during training.

Main Results:

• X-DC provides visualizations, clarifying the model's decision-making process for embedding vectors.
• The proposed X-DC achieves speech separation performance comparable to traditional DC models.

Conclusions:

• X-DC offers an interpretable alternative to black-box DNNs for speech separation.
• The physically interpretable structure of X-DC naturally facilitates model adaptation, enhancing robustness to different acoustic conditions.