A causal and talker-independent speaker separation/dereverberation deep learning algorithm: Cost associated with...

Area of Science:

• Speech processing
• Computational auditory scene analysis
• Deep learning

Background:

• Real-time speech processing requires causal algorithms that do not use future information.
• Computational auditory scene analysis (CASA) aims to separate target speech from interfering sounds.
• Hearing impairment can significantly impact speech intelligibility in noisy environments.

Purpose of the Study:

• To evaluate the performance of a fully causal deep CASA algorithm for speech enhancement.
• To assess the intelligibility benefit provided to hearing-impaired (HI) and normal-hearing (NH) listeners.
• To quantify the performance cost of causal processing compared to a non-causal version.

Main Methods:

• A fully causal deep CASA algorithm utilizing Dense-UNet and temporal convolutional networks was developed.
• The algorithm estimated both magnitude and phase of target speech in complex acoustic scenes.
• Speech intelligibility was measured for HI and NH listeners in conditions with interfering talkers and reverberation.

Main Results:

• The causal algorithm provided significant benefit across all tested conditions.
• Hearing-impaired listeners experienced a mean benefit of 46.4 percentage points in intelligibility.
• Intelligibility decrements due to causal processing were observed in some conditions but were modest relative to the overall benefit.

Conclusions:

• The fully causal deep CASA algorithm effectively enhances speech intelligibility, particularly for hearing-impaired individuals.
• The algorithm demonstrates a significant benefit in separating target speech from complex interference.
• The performance cost of causal implementation is acceptable given the substantial gains in speech clarity.