Hearing Aid Speech Classification via EEG Cepstral Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current hearing devices struggle to accurately identify and enhance the attended speech source in noisy environments, with existing methods achieving low reconstruction accuracy and classification accuracy due to the complexity of mapping speech signals to EEG data.
Innovation Solution
A speech classification apparatus that uses cepstral processing to extract features from both sound and EEG signals, calculating sound and EEG cepstra, and correlating selected coefficients to identify the attended speech source through linear or non-linear regression methods.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If linear regression methods are used to map speech envelope to EEG data, then the system can track neural activity, but the reconstruction accuracy remains low (correlation coefficients r = 0.1 - 0.2)
Solution Approach 1:
The patent transforms the mapping approach by changing from direct speech envelope regression to cepstral coefficient correlation. This parameter transformation extracts features from both speech and EEG signals in the cepstral domain, fundamentally altering the representation space to achieve higher correlation coefficients (r > 0.5) compared to traditional linear regression methods.
Solution Approach 2:
The patent introduces cepstral coefficients as an intermediary representation between raw speech signals and EEG data. By computing cepstra from both modalities and correlating selected coefficients, the system creates a bridge that captures neural entrainment more effectively than direct speech-EEG mapping, resolving the low accuracy issue.
2Reliability
If directional hearing aids are used to reduce sounds not directly in front of the listener, then signal-to-noise ratio improves, but the listener must face the signal source and be within a certain distance
Solution Approach 1:
The patent employs EEG-based feedback to dynamically identify which speech source the listener is attending to. By continuously monitoring neural entrainment through cepstral correlation, the system adapts to the listener's cognitive attention rather than relying on fixed directional constraints, enabling effective processing regardless of the listener's orientation or distance from sound sources.
Solution Approach 2:
The system transitions from static directional beamforming to dynamic attention-based processing. The cepstral correlation method continuously identifies the attended speaker based on real-time EEG feedback, allowing the hearing device to adapt its processing to the listener's changing attention focus without requiring physical repositioning or facing constraints.
3Quantity of substance
If standard hearing aids amplify both speech and noise, then hearing capability is enhanced, but comprehension in noisy environments remains difficult for hearing-impaired persons
Solution Approach 1:
The patent extracts the attended speech signal from the mixed audio input by using EEG-based cepstral correlation to identify which speaker the listener is focusing on. This extraction approach separates the target speech from background noise and other speakers, delivering only the relevant audio content to the hearing-impaired listener rather than amplifying all sounds equally.
Data Source
Figure 1a~1b
Figure 2
Figure 3
AI summary
There is provided a speech classification apparatus 100 for hearing devices with electroencephalography, EEG, dependent sound processing, comprising a sound processing unit 10 configured to capturing sound input signals 11 from at least one external microphone 15 and segmenting said captured sound input signals into segmented sound signals 17, a speech classification unit 20 comprising a speech cepstrum calculation unit 21 configured to calculate a speech cepstrum 14 for each segmented sound signal 17, an EEG cepstrum calculation unit 40 configured to calculate an EEG cepstrum 13 for an EEG signal 12 of a user's brain 220, a mapping unit 22 configured to select a predetermined number of coefficients from each calculated sound cepstrum 14 and from the calculated EEG cepstrum 13, and a correlation unit 23 configured to calculate a correlation value for each captured sound input signal 11 based on a correlation of the predetermined number of selected coefficients 13, 14 from the respective calculated sound cepstrum with the predetermined number of selected coefficients from the calculated EEG cepstrum, wherein an attended speech source is classified based on the obtained correlation values.