Reproduce “Speech Boosting: Low-Latency Live Speech Enhancement for TWS Earbuds” on OpenResearch

About this paper

Hanbin Bae, Pavel Andreev, Azat Saginbaev, Nicholas Babaev, Won-Jun Lee, Hosang Sung, Hoon-Young Cho

Researchers at Samsung Research developed an ultra-low-latency, highly efficient speech enhancement system for True Wireless Stereo (TWS) earbuds, enabling clear conversations in noisy environments with Active Noise Cancellation (ANC) active. The optimized system achieves 3 ms algorithmic latency and 0.21 GMAC computational complexity, providing a subjective MOS of 3.71, comparable to much larger non-real-time models.

Key insights

• Time-domain architectures like Wave-U-Net + LSTM are more robust and perform better than constrained state-of-the-art time-frequency models (TF-GridNet) in ultra-low-latency, low-complexity scenarios.
• A two-stage loss function, combining PFPL with adversarial and perceptual metric losses, significantly improves subjective speech quality compared to single-stage or regression-only approaches.
• Advanced pruning techniques like SPDY + OBC are crucial for achieving drastic complexity reduction (e.g., ~90%) in deep learning models without substantial quality degradation, outperforming simple magnitude pruning.

Summary by alphaXiv.