Deep neural network-based generalized sidelobe canceller for dual-channel far-field speech recognition

Area of Science:

• Speech processing
• Machine learning
• Acoustic signal processing

Background:

• Traditional generalized sidelobe cancellers (GSC) optimize signal levels, not automatic speech recognition (ASR) performance.
• Far-field ASR systems require robust speech enhancement to mitigate noise.
• Existing GSC methods do not guarantee optimal ASR outcomes.

Purpose of the Study:

• To propose a novel deep neural network (DNN)-based GSC (nnGSC) optimized for ASR performance.
• To develop a GSC structure that integrates learnable modules and joint optimization with acoustic models.
• To enhance noise robustness and performance of far-field ASR systems.

Main Methods:

• Developed a dual-channel DNN-based GSC (nnGSC) structure.
• Initialized nnGSC with traditional GSC coefficients for guided network learning.
• Employed joint optimization between the GSC and the acoustic model.
• Enabled frame-by-frame target direction-of-arrival (DOA) tracking without external algorithms.
• Trained nnGSC with multi-geometry data to improve robustness against array mismatches.

Main Results:

• nnGSC achieved significant relative character error rate (CER) improvements: 23.7% over microphone observation, 13.5% over oracle direction-based super-directive beamformer, 12.2% over oracle direction-based traditional GSC, and 5.9% over oracle mask-based MVDR beamformer.
• The proposed nnGSC automatically tracks target DOA without additional localization.
• nnGSC demonstrated improved robustness against array geometry mismatches.

Conclusions:

• The novel nnGSC structure optimizes ASR performance by integrating DNNs and joint acoustic model training.
• nnGSC offers superior speech enhancement and noise robustness for far-field ASR compared to traditional methods.
• The approach leverages both signal processing knowledge and data-driven learning for effective speech recognition enhancement.