Noise Canceling Apparatus Using Deep Learning and Statistical Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise canceling methods, including those using beam-forming and deep learning, face challenges in effectively canceling noise in dynamic environments and fail to minimize distortion, particularly when noise sources resemble voice or have strong low-frequency signals.
Innovation Solution
A noise canceling apparatus and method that employs a machine learning or deep learning algorithm to primarily cancel noise from an input voice signal, followed by secondary cancellation using statistical analysis based on speech presence probability, to generate a clear voice signal while minimizing distortion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning algorithm is used for noise canceling, then ability to cancel noise similar to voice is improved, but voice distortion is increased
Solution Approach 1:
The patent segments the noise canceling process into two distinct stages: primary noise canceling using deep learning algorithm and secondary residual noise canceling using statistical analysis. This segmentation allows each method to be optimized for its specific function, with deep learning handling the complex task of removing voice-like noise and statistical analysis refining the result to minimize distortion.
Solution Approach 2:
The patent applies different processing qualities to different components of the noise canceling task. Deep learning is applied locally to the primary noise cancellation task where high noise removal capability is needed, while statistical analysis is applied locally to the residual noise where precision and distortion control are critical.
2Adaptability or versatility
If beam-forming method is used for noise canceling, then location-based noise canceling is improved, but performance deteriorates when user moves or noise resembles voice
Solution Approach 1:
The patent replaces the mechanical beam-forming approach with a data-driven deep learning model. Instead of relying on spatial filtering and user location information, the system uses a trained neural network that directly processes audio signals to identify and remove noise, achieving superior performance especially when noise resembles voice or when users move dynamically.
3Measurement precision
If noise canceling model is used, then primary noise cancellation is improved, but residual noise remains
Solution Approach 1:
The patent ensures continuous noise canceling action by sequentially applying two different methods. The deep learning model performs primary noise cancellation continuously, and the statistical analysis continuously refines the output to remove residual noise, creating an uninterrupted chain of noise reduction that maintains high quality throughout the processing.
Data Source
AI summary
An embodiment of the present invention provides an apparatus for noise canceling that includes: an input unit configured to receive an input voice signal; and one or more processors configured to perform a first noise cancellation using as input the received input voice signal to generate a first voice signal by cancelling noise from the input voice signal using a noise canceling model which is trained using a plurality of reference voice signals, perform a second noise cancellation using as input the first voice signal generated by the noise canceling model to generate a second voice signal in which residual noise is canceled from the first voice signal using statistical analysis, and generate an output voice signal comprising an encoding of the second voice signal.