Audio Processing

Audio processing encompasses the analysis, transformation and synthesis of sound signals to extract information, enhance quality or facilitate new user experiences. At its core lie spectral methods—short-time Fourier transforms, wavelet and empirical mode decompositions—that decompose complex waveforms into time-frequency representations. Statistical measures such as permutation entropy quantify signal complexity, while adaptive filtering and beamforming address noise, reverberation and source localisation. Machine learning has transformed the field: deep neural networks now learn end-to-end mappings for denoising, separation and classification tasks. Applications range from hearing aids and teleconferencing to underwater acoustics, bioacoustics and voice biometrics. Emerging trends include cooperative sensor arrays, self-supervised pre-training and privacy-preserving embeddings, reflecting the field’s interdisciplinary convergence of signal processing, acoustics and artificial intelligence.

Research from Nature Portfolio

Speech intelligibility for hearing-impaired users has been restored to normal-hearing levels through a deep learning algorithm optimised by neural architecture search and a data-driven intelligibility metric. Operating on single-microphone inputs, the network suppresses noise while preserving speech cues in real time, demonstrating state-of-the-art denoising performance across diverse noise categories on human-graded assessments.

A wireless acoustic swarm self-organises into an adaptive microphone array that captures spatial information in reverberant rooms. Coupled with an attention-based neural framework, it localises and separates three to five concurrent speakers with centimetre-level accuracy, enabling configurable speech zones for selective activation or muting.

Recurrence plot embeddings have been introduced as stand-alone nonlinear features for speaker identification across air, bone and throat conduction modes. By mapping short signal segments to dynamical system representations, these features power unimodal and multimodal classifiers that achieve over 99 percent accuracy, underscoring the value of nonlinear dynamical cues in speaker modeling.

Audio Processing publication trend

The graph below shows the total number of articles in audio processing across all publications each year (not limited to Nature Index journals).

Technical terms

Empirical Mode Decomposition (EMD): A data-driven algorithm that decomposes non-stationary signals into intrinsic mode functions without predefined bases.

Permutation Entropy: A measure of signal complexity that quantifies the randomness of value orderings in time series segments.

Mask Inference: A neural network–based technique that estimates time-frequency masks to separate or enhance sources in a mixture.

Recurrence Plot Embedding: A representation of nonlinear dynamical systems by mapping time series segments into recurrence matrices of state distances.

Speech Zone: A spatial region defined by adaptive microphone arrays and neural separation frameworks to isolate or suppress audio sources.

References

1. Restoring speech intelligibility for hearing aid users with deep learning. Scientific Reports (2023).
2. Creating speech zones with self-distributing acoustic swarms. Nature Communications (2023).
3. Recurrence plot embeddings as short segment nonlinear features for multimodal speaker identification using air, bone and throat microphones. Scientific Reports (2024).
4. Deep neural network techniques for monaural speech enhancement and separation: state of the art analysis. Artificial Intelligence Review (2023).
5. A New Underwater Acoustic Signal Denoising Technique Based on CEEMDAN, Mutual Information, Permutation Entropy, and Wavelet Threshold Denoising. Entropy (2018).
6. HilbertHuang Transform With Intelligent Noise Reduction for Passive SONAR Signal Processing. IEEE Journal of Oceanic Engineering (2025).