Distance-Based Sound Separation Using Neural Source Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sound separation technologies face challenges in efficiently separating sounds based on distance, particularly in noisy environments, with limitations in computational power, memory, and user-specific sound preferences, and existing methods often fail to include non-speech sounds or require cumbersome user interaction.
Innovation Solution
A neural network-based sound separation model that estimates source distances using acoustic cues, allowing separation of near and far sounds without physical distance limitations, leveraging intensity and reverberation ratios to group sources based on distance thresholds, and can be trained on simulated audio environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If neural network model size is increased to improve sound separation quality, then separation quality improves, but computational power and memory requirements increase
Solution Approach 1:
The sound separation task is segmented into two independent stages: first estimating distance parameters (DRR, distance) from the mixed audio signal, then using these parameters to separate near and far sources. This segmentation allows each neural network to be smaller and more efficient, avoiding the need for a single large end-to-end model while achieving comparable or better separation quality.
2Measurement precision
If beamforming techniques are used to separate sounds based on direction, then sound separation is achieved, but maximum distance is limited by the size of the microphone array
Solution Approach 1:
The patent introduces distance estimation parameters (DRR and distance) as intermediary variables that bridge the gap between microphone array measurements and source separation. These intermediaries encode spatial information in a compact form that can be used to separate sources beyond the physical limits of the array, effectively extending the separation range without increasing array size.
3Adaptability or versatility
If selective listening methods are used to enhance specific sounds, then user preferences can be accommodated, but user effort and interaction are required
Solution Approach 1:
The system performs automatic sound separation based on estimated distance parameters without requiring explicit user input or interaction. The neural network autonomously identifies and separates near and far sources according to the spatial information encoded in the audio signal, making the system both adaptive and easy to use simultaneously.
4Adaptability or versatility
If class-based methods are used to include non-speech sounds, then sound diversity is improved, but interference from sounds of the same class reduces effectiveness
Solution Approach 1:
Instead of classifying sounds by type (speech, music, etc.) and then separating them, the patent inverts the approach by first estimating spatial parameters and separating sources based on distance, regardless of sound class. This inversion avoids the fundamental problem of class-based methods where interfering sounds of the same class cannot be distinguished, achieving reliable separation for all sound types equally.
Data Source
AI summary
A computer-implemented method of applying a trained neural network for sound separation based on distance estimation is provided. The method includes receiving, by an audio input component of a computing device, an audio mixture from one or more sources. The method includes predicting, by a trained distance estimation neural network and based on the audio mixture, respective distances of the one or more sources from the audio input component. The method includes determining one or more near sounds and one or more far sounds based on the respective distances. The near sounds correspond to sources that are located within a threshold distance of the audio input component, and the far sounds correspond to sources that are not located within the threshold distance of the audio input component. The method includes providing the predicted one or more near sounds.