Method and Apparatus for Determining a Deep Filter

Simple SummaryContent extracted from patent full text and abstract with AI.

This invention describes a method and apparatus for determining and applying a 'deep filter'—a type of multi-dimensional filter estimated by a deep neural network—for extracting, separating, or reconstructing desired signals (such as speech) from complex mixtures (such as noisy audio). Unlike traditional methods that use time-frequency masks or statistical models, the deep filter enables improved extraction or reconstruction, even in cases where interfering signals cause destructive interference or when data is missing, by leveraging learned representations over time, frequency, or sensor domains.

Use CasesContent extracted from patent full text and abstract with AI.

• Noise reduction in audio and speech recordings (e.g., for teleconferencing, hearing aids, mobile devices).
• Speech enhancement in communication systems such as phones, assistive listening devices, and teleconferencing setups.
• Audio source separation (e.g., separating voices or sound sources in music, security, or surveillance recordings).
• Reconstruction of missing or damaged signal parts (e.g., packet loss concealment in streaming, audio repair in recordings).
• Biomedical signal enhancement (e.g., removing noise from ECG, EEG signals for better diagnosis).
• Bandwidth extension or enhancement of low-quality transmissions (e.g., improving narrowband audio).
• Declipping and artifact removal in digital audio restoration.

BenefitsContent extracted from patent full text and abstract with AI.

• Improved separation and extraction performance, especially in adverse environments with strong interference or missing data.
• Ability to reconstruct missing signals or parts of a signal, not just suppress noise.
• Does not rely on statistical assumptions about the signals or interferers, allowing for robust application in highly variable real-world situations.
• Can be applied to single-channel recordings, not requiring complex multi-microphone setups.
• Flexible multi-dimensional filtering allows adaptation to a wide range of signal processing problems (time, frequency, and sensor domains).
• Neural network-based estimation allows end-to-end learning and optimization based on real-world data and application-specific goals.
• Can be integrated as software (e.g., firmware, applications) or embedded in hardware.

Technical Classifications (CPCs)

Inventors & Applicants

Patent Abstract

A method for determining a deep filter comprises the following steps:• receiving a mixture;• estimating using a deep neural network the deep filter, wherein the estimating is performed, such that the deep filter, when applying to elements of the mixture, obtains estimates of respective elements of the desired representation;wherein the deep filter of at least one dimension comprises a tensor with elements.