Spatial Audio Cropping Using Blind Source Separation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing audio source separation technologies face challenges in accurately isolating desired audio sources from multiple sources in a scene without location information or microphone geometry, leading to noisy spatial responses and difficulties in determining the likelihood of demixing separation filters corresponding to a selected field of view.
Innovation Solution
The method employs Blind Source Separation techniques, specifically Maximum Likelihood Independent Component Analysis (ML-ICA) in the time-frequency domain, using Short Time Fourier Transforms and importance weighting, along with stochastic gradient-based improvements and importance subsampled stochastic gradient ICA, to separate audio sources and crop out unwanted components based on spatial positioning, leveraging Bayesian Rank Statistics for robust source selection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Blind Source Separation is performed without location information or microphone geometry, then audio sources can be separated into components, but the spatial response becomes noisy and it becomes difficult to determine the likelihood of demixing separation filters corresponding to a selected field of view
Solution Approach 1:
The patent introduces an intermediary statistical model (Bayesian Rank Statistics) that mediates between the noisy spatial responses from blind source separation and the reliable determination of source likelihood. This intermediary layer processes the uncertain spatial information through probabilistic ranking, converting unreliable phase correlation data into reliable source selection decisions based on rank probabilities rather than raw spatial coordinates
Solution Approach 2:
The patent changes the parameter representation from deterministic spatial coordinates to probabilistic rank distributions. Instead of relying on precise spatial responses that are noisy without microphone geometry, the system transforms the problem into parameter space of rank probabilities, where the likelihood of a source being in the field of view is determined by statistical ranking rather than direct spatial measurement
2Measurement precision
If phase correlations are calculated for all possible directions to determine source alignment, then the probability of source alignment can be determined, but the computational complexity increases
Solution Approach 1:
The patent applies partial action by calculating phase correlations only for a limited set of discrete directions rather than continuously for all possible directions. The spatial domain is sampled at specific angular intervals, and phase correlations are computed only at these discrete points, reducing computational load while maintaining sufficient accuracy for source alignment probability determination
Solution Approach 2:
The patent segments the continuous spatial domain into discrete directional bins or sectors. Instead of processing all possible directions as a continuous range, the spatial field is divided into discrete segments, and phase correlations are calculated independently for each segment, enabling efficient computation through parallel processing and reducing the overall computational complexity
Data Source
AI summary
A method of cropping a portion of an audio signal captured from a plurality of spatially separated audio sources in a scene, the method comprising: capturing the audio signal with one or more recording devices; separating the audio signal into a plurality of components each associated with one or more of the plurality of audio sources; selecting a spatial region in the scene; determining which of the plurality of components are associated with an audio source positioned outside of the selected spatial region; and cropping the plurality of components associated with an audio source positioned outside of the selected spatial region out of the audio signal.