Wind Noise Attenuation Using Microphone Array Coherence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise suppression methods fail to adequately remove wind noise from microphone signals due to the difficulty in differentiating wind noise and speech through energy or SNR analysis in the time or frequency domains.
Innovation Solution
The system employs space selectivity and signal correlation properties at two or more microphones to determine wind noise, using three properties: wind noise being uncorrelated with speech signals, wind noise at different locations being largely uncorrelated, and speech signals being correlated across microphones. This approach quickly constructs a reliable wind noise detector and applies an effective wind noise attenuator, using coherence, phase of the cross power spectrum, and probabilities of speech and wind noise to derive an attenuation gain factor.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If energy or SNR analysis is used to differentiate wind noise and speech, then noise suppression can be attempted, but wind noise and speech cannot be adequately differentiated
Solution Approach 1:
The patent transitions from traditional time-domain and frequency-domain analysis to the spatial domain by utilizing multiple microphones arranged in a specific geometry. This dimensional change enables the system to differentiate wind noise from speech by analyzing spatial correlation patterns, achieving reliable wind noise detection where traditional energy or SNR analysis failed.
Solution Approach 2:
The patent changes the analysis parameters from energy-based metrics to coherence-based metrics in the frequency domain. By computing coherence between signals from multiple microphones across different frequency bins, the system can reliably distinguish wind noise (which appears incoherently across microphones) from speech (which maintains coherent spatial patterns), resolving the differentiation problem.
2Object-affected harmful factors
If traditional noise suppression methods are applied, then some noise reduction may occur, but wind noise is not adequately removed
Solution Approach 1:
The patent extracts wind noise components from the mixed signal by utilizing the incoherence property of wind noise across multiple microphone channels. Through coherence-based spectral analysis, the system identifies and extracts frequency components characteristic of wind noise, then suppresses them while preserving speech components that maintain coherent spatial relationships.
Solution Approach 2:
The patent implements a feedback mechanism where coherence estimates from multiple microphones continuously inform the wind noise suppression process. The system computes coherence metrics, uses them to estimate wind noise power spectral density, applies suppression gains, and can iteratively refine the estimation, creating a closed-loop system that reliably reduces wind noise while preserving speech.
3Reliability
If multiple microphones are used to detect wind noise, then detection reliability improves, but device complexity increases
Solution Approach 1:
The patent makes the multi-microphone array serve multiple functions: it simultaneously captures speech signals, detects wind noise through coherence analysis, and provides spatial information for noise suppression. This multi-functionality justifies the increased device complexity by extracting maximum utility from the microphone array, enabling reliable wind noise detection and suppression that single-microphone systems cannot achieve.
Data Source
Figure 1
Figure 2
Figure 3A
AI summary
Approaches for detecting and reducing wind noise from audio signals captured at multimicrophone array are described. In aspects, the wind noise detector is constructed from probabilities of speech presence and wind noise presence, which are derives from statistics of the phase differences among the time aligned signals of multi-microphones in separate frequency regions. Wind noise, if detected, is reduced by a gain in frequency domain, which is also a function of the phase difference and its statistics.