Research on Speech Enhancement Algorithms for Wearable Devices

Jingze Fu, Fulang Sun, Hao Yin, Bingyu Shen, Zihao Wang, Hengfan ZhangAtlantis 2023 · DOI · OpenAlex

Abstract

Speech is the primary information carrier in human communication, yet environmental noise frequently degrades signal quality. Speech enhancement algorithms serve as a critical tool for noise reduction and improving the subjective audio experience. This research analyzes two primary digital signal processing (DSP) paradigms: single-channel and dual-channel methods. Single-channel approaches are categorized into traditional DSP-based algorithms and modern data-driven learning algorithms. While traditional methods struggle with non-stationary noise, deep learning has demonstrated significant performance gains. Dual-channel systems further enhance quality by utilizing a secondary microphone to provide a more accurate noise reference. This paper synthesizes current methodologies and highlights the shift toward hybrid, deep-learning-integrated systems for wearable technology.

Introduction

The proliferation of wearable devices—such as wireless headsets, hearing aids, and head-worn displays—has made clear voice communication in diverse environments a necessity. These devices often operate in "noisy" settings characterized by wind, traffic, or background chatter. Such "non-stationary" noise poses a significant challenge for traditional enhancement techniques, which often assume noise characteristics remain stable over time. This paper explores the evolution of speech enhancement from rigid mathematical models to adaptive, data-driven architectures designed to meet the subjective quality demands of the human ear.

Methodological Framework

Speech enhancement architectures are broadly divided based on their input configuration.

Single-Channel Speech Enhancement (SCSE)

SCSE techniques operate on a single microphone input and are split into two generations:

  1. Traditional DSP Algorithms: This category includes spectral subtraction, statistical models (e.g., Hidden Markov Models and Vector Quantization), and subspace algorithms. Subspace methods partition signals into "speech" and "noise" Euclidean spaces to isolate the target signal. However, these methods often produce "musical noise" artifacts when dealing with dynamic environments.
  2. Learning-Based Algorithms: Modern approaches utilize Deep Neural Networks (DNNs) and Sparse Representation. These are "data-driven" models that learn to identify and mask noise patterns from massive datasets, offering superior performance in handling non-stationary noise compared to traditional assumptions.

Dual-Channel Speech Enhancement (DSE)

DSE systems utilize a second microphone to capture ambient noise, providing a "ground truth" reference:

  • Adaptive Noise Cancelling (ANC): Employs algorithms like Least Mean Square (LMS) and Recursive Least Squares (RLS) to subtract the secondary microphone's noise signal from the primary speech signal.
  • First-Order Difference Microphone (FDM): A spatial filtering technique that uses the physical distance between microphones to calculate a "difference signal," effectively nullifying noise based on its direction of arrival.
  • Phase Time-Frequency Masking: A multi-step process involving beamforming and Wiener filtering to isolate the target voice within a three-dimensional acoustic space.

Key Insights and Discussion

The research identifies the "stationary trap" as a primary limitation of legacy systems. Traditional spectral subtraction assumes noise is relatively constant; when it is not, the resulting audio often contains distracting artifacts. Deep learning addresses this by acting as a "soft mask," predicting which signal components are likely speech.

A critical insight for wearable design is that hardware configuration is paramount. Dual-channel systems are inherently more robust because they do not have to "guess" the noise profile; they measure it. The most effective modern systems often combine these paradigms, using spatial beamforming (DSE) alongside DNN-based post-filtering (SCSE) to achieve maximum clarity.

Conclusion and Future Outlook

Deep learning has established itself as the gold standard for processing non-stationary noise in single-channel systems. For high-end wearables, dual-channel configurations provide the most realistic noise reference. However, the reliance of DNNs on large "parallel" datasets (clean and noisy pairs) remains a bottleneck. Future research into weak supervision—training on non-parallel data—and the optimization of these models for low-power wearable processors will be essential for the next generation of communication devices.

Summary updated

Abstract

Speech is the primary information carrier in human communication and the interference could be caused by environmental noise.The speech enhancement algorithm is an effective method to reduce noise and improve the subjective feeling of the human ear.The existing speech enhancement algorithms can be divided into two categories: single-channel methods and dual-channel methods.This research analyzed the two main digital signal processing topics for speech enhancement algorithms.This included the traditional single channel speech enhancement algorithm and the dual-channel speech enhancement algorithm.The traditional single channel speech enhancement algorithm was further divided into two categories: traditional enhancement algorithms based on digital signal processing and learning enhancement algorithms based on data-driven.The traditional algorithms have limitations in processing non-stationary noise, whereas deep learning has significantly improved handling non-stationary noise.The dual-channel speech enhancement algorithm adds a second microphone to collect noise, providing more real noise for enhancement.This research briefly described the main dual-channel speech enhancement algorithms, including adaptive noise cancelling and firstorder difference microphone.In conclusion, the research found that deep learning has shown significant improvement in handling non-stationary noise in singlechannel speech enhancement algorithms, and dual-channel speech enhancement algorithms have the potential to provide more real noise for enhancement.
BibTeX
@inproceedings{fu2023research,
  author = {Jingze Fu and Fulang Sun and Hao Yin and Bingyu Shen and Zihao Wang and Hengfan Zhang},
  title = {{Research on Speech Enhancement Algorithms for Wearable Devices}},
  booktitle = {Atlantis},
  year = {2023},
  doi = {10.2991/978-94-6463-200-2\_101},
}
Figure