Blind Source Separation Using Signal Absence Probability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional blind source separation methods, such as DUET, fail to effectively separate mixture signals in indoor environments with reverberations and non-stationary noise, particularly when white Gaussian noise or fan noise is present, leading to deteriorated signal separation performance.
Innovation Solution
A method and apparatus that calculate global and local signal absence probabilities for each frequency band, estimate noise-free spectrum vectors, and generate source label vectors using attenuation and delay parameters to separate source signals from mixture signals received through two microphones, enabling real-time noise elimination and source signal separation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional blind source separation methods (DUET) are used, then source signals can be separated in ideal conditions, but signal separation performance deteriorates in indoor environments with reverberations and non-stationary noise
Solution Approach 1:
The patent changes the parameter representation from time-domain to time-frequency domain using Short-Time Fourier Transform (STFT). This transformation allows the algorithm to operate on spectral components where noise and reverberation effects can be more effectively distinguished from source signals through statistical analysis of signal absence probabilities across frequency bins and time frames.
Solution Approach 2:
The patent segments the mixture signal into multiple frequency bins using STFT, and then applies independent statistical analysis to each frequency bin. This segmentation allows the algorithm to identify signal absence probabilities locally in the time-frequency plane, making the separation process robust to non-stationary noise and reverberation that affect different frequency components differently.
2Object-affected harmful factors
If white Gaussian noise or fan noise is present, then noise elimination becomes more difficult, but the patent achieves effective noise elimination through probability-based filtering
Solution Approach 1:
The patent uses feedback by iteratively updating the estimate of the source signal spectrum based on the calculated signal absence probabilities. The algorithm refines its noise estimation and source signal separation through multiple passes, using the separated signal information to improve subsequent noise elimination decisions, thereby achieving effective removal of white Gaussian noise and fan noise.
Solution Approach 2:
The patent transforms the noise elimination problem from the time domain to the frequency domain using STFT. In the frequency domain, the algorithm can apply magnitude-based filtering using signal absence probabilities, which is particularly effective for removing stationary and non-stationary noise components including white Gaussian noise and fan noise while preserving source signal characteristics.
3Device complexity
If the w-disjoint orthogonality assumption is made, then separation is simplified, but performance deteriorates when noise occupies entire frequency bands
Solution Approach 1:
The patent applies partial action by using the w-disjoint orthogonality assumption only as a starting point for clustering frequency bins, rather than enforcing it strictly throughout the entire algorithm. The method initially groups frequency bins based on this assumption to reduce complexity, then refines the separation using statistical signal absence probability analysis that does not rely on the orthogonality assumption, thereby maintaining both computational efficiency and robustness to noise.
Solution Approach 2:
The patent segments the frequency spectrum into multiple bins and applies independent statistical analysis to each bin rather than treating the entire spectrum as a single unit. This segmentation allows the algorithm to identify which frequency bins contain source signals and which contain only noise, even when noise occupies entire frequency bands, by calculating signal absence probabilities locally in each bin and using clustering to group bins with similar characteristics.
Data Source
AI summary
A method and apparatus to separate first and second mixture signals received from two sensors and transformed into the frequency domain in two or more source signals. The signal separation method includes: calculating a global signal absence probability for each frame and a local signal absence probability for each frequency band of a corresponding frame for at least one of the first and second mixture signals; estimating a spectrum vector for each frequency band in which a noise signal is eliminated using the global signal absence probability; determining a plurality of frequency bands including at least one of a noise signal and a source signal using the local signal absence probability, and generating a source label vector which consists of a plurality of frequency bands assigned to each source, using an attenuation parameter and a delay parameter generated for each of the determined frequency bands; and multiplying the spectrum vector estimated for each frequency band by the source label vector, and obtaining signals separated according to the source signals.