DNN Noise Echo Removal via Integrated Gain Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice signal noise and echo elimination technologies face performance degradation when trained in normal environments but fail in abnormal noise conditions, and are inefficient when using far-end speech signals and microphone input information alone.
Innovation Solution
A deep neural network (DNN) method that integrates and eliminates noise and echo by using noise and echo information as additional inputs, estimating integrated and eliminated gains, and employing signal-to-echo and signal-to-noise ratios to improve feature vectors for effective noise and echo removal.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a noise eliminator and echo eliminator are independently designed and connected in series, then each component can be optimized separately, but the overall performance degrades due to sequential processing and non-linear computation interference
Solution Approach 1:
The patent merges the noise eliminator and echo eliminator into a single integrated apparatus that processes voice signals simultaneously. The voice signal processor combines both noise elimination and echo elimination functions in one system, allowing the noise eliminator and echo eliminator to operate cooperatively rather than sequentially, thereby maintaining overall performance while enabling independent optimization of each component.
2Productivity
If the noise eliminator is positioned at the front end of the echo eliminator, then noise can be eliminated first, but the echo eliminator performance degrades due to non-linear computation of the noise eliminator
Solution Approach 1:
The patent implements dynamic processing where the noise eliminator and echo eliminator operate simultaneously with adaptive processing. The voice signal processor dynamically adjusts the processing of noise and echo components based on their interactions, allowing both eliminators to function optimally without the performance degradation that occurs in fixed sequential arrangements.
3Productivity
If the echo eliminator is positioned at the front end of the noise eliminator, then echo can be eliminated first, but the noise estimation performance degrades because the spectrum is distorted in the echo elimination process
Solution Approach 1:
The patent combines noise elimination and echo elimination into a unified processing system where both functions operate simultaneously on the voice signal. This integrated approach allows the system to maintain accurate noise estimation while eliminating echo, avoiding the spectrum distortion problem that occurs when echo elimination is performed separately before noise estimation.
4Reliability
If a statistical model is used for noise and echo elimination, then the system works well in normal noise environments, but performance is greatly degraded in abnormal noise environments
Solution Approach 1:
The patent employs deep neural networks that can dynamically adapt processing parameters based on the noise characteristics of the input signal. Unlike fixed statistical models, the DNN-based system changes its processing parameters according to the specific noise environment, enabling it to maintain high performance in both normal and abnormal noise conditions by learning optimal processing strategies for different scenarios.
5Ease of manufacture
If only far-end speech signal and microphone input information are used for DNN training, then the training process is simple, but the DNN is rarely trained effectively and noise and echo elimination performance is insufficient
Solution Approach 1:
The patent performs preliminary processing to generate additional training features from the available far-end speech signal and microphone input information. By pre-computing features such as signal-to-noise ratios, signal-to-echo ratios, and other statistical characteristics, the system creates more informative training data without requiring additional input signals, thereby enabling effective DNN training with the existing data sources.
Data Source
AI summary
Disclosed is a deep neural network-based method and apparatus for combining noise and echo removal. The deep neural network-based method for combining noise and echo removal according to one embodiment of the present invention may comprise the steps of extracting a feature vector from an audio signal that includes noise and echo; and acquiring a final audio signal from which both noise and echo have been removed, by using a combined nose and echo removal gain estimated by means of the feature vector and deep neural network DNN.