Wind Noise Reduction in Multi-Microphone Speech Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Automatic speech recognition systems in mobile devices face challenges in varying recording conditions due to microphone placement issues, user behavior, acoustic environments, and environmental noise, particularly wind noise, which degrade signal-to-noise ratio (SNR) and increase word error rates.

Innovation Solution

A mobile device with multiple microphones is configured to estimate acoustic characteristics and detect wind noise, selecting the weakest microphone signal at low frequencies and the strongest at high frequencies to generate a wind noise-reduced mono audio signal, optimizing frequency balance and SNR for improved automatic speech recognition performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single microphone is used for speech capture, then the device complexity is reduced, but the signal-to-noise ratio deteriorates in windy conditions

Engineering Contradiction:
Improvemicrophone configurationVSAvoidsignal-to-noise ratio
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The audio signal processing is segmented into different frequency bands (low frequencies below wind noise threshold and high frequencies above threshold). For low frequencies, the weakest microphone signal is selected to minimize wind noise, while for high frequencies, the strongest signal is selected to maximize speech quality. This segmentation allows the system to handle different frequency components with appropriate strategies.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically adjusts microphone selection based on estimated wind noise characteristics. The processor continuously estimates the wind noise threshold frequency and adapts the signal processing strategy in real-time, switching between selecting the weakest or strongest microphone signals depending on the frequency band and current wind conditions.

Inventive Principle:
Principle #15Dynamics

2Reliability

If multiple microphones are used to improve signal-to-noise ratio, then the reliability of speech recognition improves, but the device complexity increases

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidmicrophone configuration
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

Different processing strategies are applied to different frequency bands. Instead of uniformly processing all frequencies the same way, the system applies local quality processing by selecting the weakest microphone signals for low frequency bands (where wind noise dominates) and strongest signals for high frequency bands (where speech content dominates).

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The system changes the selection criterion parameter based on frequency. The processor dynamically adjusts which microphone signal to select (weakest or strongest) based on the frequency band being processed, thereby optimizing the balance between noise reduction and speech quality for each frequency range.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If wind noise reduction processing is applied, then the signal-to-noise ratio improves, but the word error rate may increase due to speech distortion

Engineering Contradiction:
Improvesignal-to-noise ratioVSAvoidword error rate
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The system applies wind noise reduction selectively only to frequency bands below the estimated wind noise threshold, rather than applying processing to the entire frequency spectrum. This partial action approach reduces wind noise in affected bands while leaving high frequency speech bands unprocessed, thereby avoiding speech distortion that would increase word error rate.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

Different processing intensities are applied to different frequency regions. Low frequency bands suffer from wind noise and receive aggressive noise reduction processing, while high frequency bands contain important speech information and receive minimal or no processing to preserve speech quality and minimize word error rate.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP3757993B1Pre-processing for automatic speech recognition
Publication Date: 2023.12.06 NOKIA TECHNOLOGIES OY
  • EP3757993B1 patent drawingFigure 1
  • EP3757993B1 patent drawingFigure 2
  • EP3757993B1 patent drawingFigure 3

AI summary

An apparatus comprising means configured to: obtain two or more microphone audio signals; analyse the two or more microphone audio signals for a defined noise type; and process the two or more microphone audio signals based on the analysis to generate at least one audio signal suitable for automatic speech recognition.