Neural Network ANC Predicts Speech Signals for Better Noise Suppression

Conventional active noise control (ANC) systems excel at canceling stationary noise like engine hum or fan drone, but they struggle with highly non-stationary signals such as human speech. In a new paper accepted to APSIPA Annual Summit and Conference 2026, researchers Manami Nishikata and Shoichi Koyama introduce a feedforward ANC approach that leverages neural networks for time-series prediction of future speech signals. The core idea is to predict the upcoming speech waveform and use that prediction to update a linear control filter more effectively. Unlike traditional adaptive filters that rely only on current and past reference signals, their method incorporates the predicted future signal into the update rule, enabling proactive suppression of rapidly changing speech.

The proposed algorithm was evaluated in numerical experiments using both an oracle (true) predicted signal and a signal predicted by neural networks. In both cases, the noise reduction performance improved compared to baseline ANC methods, demonstrating that even imperfect neural predictions can boost suppression of non-stationary speech. This work could have significant implications for hearing aids, noise-cancelling headphones, and voice communication systems in noisy environments. By predicting speech rather than just reacting to it, the method offers a smarter path to handling real-world acoustic scenes. The paper is available on arXiv (2608.16092) with full code and experimental details.