A dual-stream deep attractor network with multi-domain learning for speech dereverberation and separation

Area of Science:

• Speech processing
• Machine learning
• Audio signal analysis

Background:

• Deep attractor networks (DANs) excel at speech separation using discriminative embeddings but struggle in reverberant environments.
• Existing methods like permutation invariant training (PIT) have limitations in complex acoustic conditions.

Purpose of the Study:

• To develop an improved speech separation and dereverberation method for reverberant environments.
• To enhance the performance of deep attractor networks (DANs) in challenging acoustic conditions with multiple speakers.

Main Methods:

• A novel dual-stream DAN with multi-domain learning was proposed, integrating a speaker encoding stream (SES) and a speech decoding stream (SDS).
• The SES models speaker information in a Fourier transform-based embedding space, while the SDS estimates the early sound component in the time domain.
• Clustering losses were employed to refine attractor estimation.

Main Results:

• The dual-stream DAN achieved significant scale-invariant source-to-distortion ratio (SI-SDR) improvements: 9.8 dB for 2-speaker and 7.5 dB for 3-speaker evaluations.
• Performance gains exceeded the baseline DAN by 2.0/0.7 dB and convolutional time-domain audio separation network (Conv-TasNet) by 1.0/0.5 dB on reverberant datasets.
• The method demonstrated effectiveness in handling variable numbers of speakers in reverberant settings.

Conclusions:

• The proposed dual-stream DAN effectively addresses speech separation and dereverberation challenges in reverberant environments.
• This multi-domain learning approach offers a substantial improvement over existing speech separation techniques.
• The method shows promise for real-world applications requiring robust speech processing in adverse acoustic conditions.