Sound Source Localization Using Blind Separation and Binary Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current audio signal processing technologies, such as TDOA and DOA algorithms, struggle with localizing a specific sound source in the presence of multiple active sources, often failing to accurately determine the location of the target sound source amidst other noise sources, and are inefficient in underdetermined source localization scenarios where the number of sources exceeds the number of microphones.
Innovation Solution
The proposed solution involves converting microphone channel signals into time-frequency domain spectrograms, applying blind source separation to identify a target sound source, estimating a binary mask to highlight the target source's activity, and using this mask to reduce the underdetermined localization problem to a determined one, allowing for accurate localization of the target sound source using techniques like ESPRIT and non-negative matrix factorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional TDOA or DOA algorithms are used to localize multiple sound sources, then the locations of all sources can be estimated, but it is impossible to determine which location belongs to the target speaker without additional information
Solution Approach 1:
The patent segments the mixed audio signal from multiple speakers into individual speaker signals using blind source separation. By dividing the complex multi-source localization problem into separate single-source problems, the system can identify and localize the target speaker more accurately without the ambiguity of traditional multi-source algorithms.
Solution Approach 2:
The patent performs blind source separation and speaker identification as preliminary steps before localization. By pre-separating the mixed signals and identifying which separated signal corresponds to the target speaker, the system prepares the data in advance, making the subsequent localization step straightforward and accurate.
2Measurement precision
If blind source separation is applied to separate channel spectrograms, then the target source can be identified, but the computational complexity increases
Solution Approach 1:
The patent transforms the time-domain signals into the frequency domain using spectrograms before applying blind source separation. This dimensionality change to time-frequency representation provides better separation characteristics and makes the blind source separation more effective, improving target identification accuracy while managing computational complexity through efficient frequency-domain processing.
3Adaptability or versatility
If the number of microphones is less than the number of active sources, then the system cannot be overdetermined, but traditional algorithms fail in such underdetermined scenarios
Solution Approach 1:
The patent performs blind source separation as a preliminary step that effectively increases the usable signal information before localization. By separating the mixed signals and identifying the target speaker's contribution, the system creates a determined or overdetermined problem from an underdetermined one, enabling accurate localization even when microphones are fewer than sources.
Solution Approach 2:
The patent changes the problem parameters by transforming from raw microphone signals to separated source signals with identified target components. This parameter transformation allows the system to achieve accurate localization in underdetermined scenarios by working with the separated and identified target source rather than the mixed original signals.
Data Source
AI summary
The proposed method for localizing a target sound source from a plurality of sound sources, wherein a multi-channel recording signal of the plurality of sound sources comprises a plurality of microphone channel signals, comprises converting each microphone channel signal into a respective channel spectrogram in a time-frequency domain, blindly separating the channel spectrograms to obtain a plurality of separated source signals, identifying, among the plurality of separated source signals, the separated source signal that best matches a target source model, estimating, based on the identified separated source signal, a binary mask reflecting where the target sound source is active in the channel spectrograms in terms of time and frequency, applying the binary mask on the channel spectrograms to obtain masked channel spectrograms, and localizing the target sound source from the plurality of sound sources based on the masked channel spectrograms.