Noise Suppression Using Joint Distribution Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise suppressing apparatuses in hands-free telephone systems, such as those used in video conferencing, face challenges in accurately calculating noise suppression coefficients, leading to a deterioration in sound quality and clarity, especially in environments with high background noise like cars, due to insufficient estimation accuracy.
Innovation Solution
A noise suppressing apparatus that calculates a suppression coefficient by converting input signals into frequency spectra, estimating noise levels, calculating weight coefficients, and using a joint distribution model of sound to derive an estimation expression for the sound spectrum, allowing for higher estimation accuracy and effective noise suppression without over-suppressing sound.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If MAP estimation and Bayes' theorem are used to calculate noise suppression coefficient, then the amount of computation is kept low, but the accuracy with which noise suppression coefficient can be calculated is not sufficient
Solution Approach 1:
The sound distribution model is segmented into multiple statistical distribution models (N≥2), each representing different acoustic conditions. By dividing the overall sound model into distinct statistical components, the system can selectively apply appropriate models to different frequency regions or time periods, improving estimation accuracy without requiring complete recalculation of a single complex model.
Solution Approach 2:
The weight coefficients for combining statistical distribution models are dynamically adjusted based on signal characteristics. The system adaptively determines the contribution of each statistical model according to the actual acoustic conditions, allowing the noise suppression coefficient calculation to respond to changing environments while maintaining computational efficiency through selective model application.
2Object-generated harmful factors
If noise suppression coefficient is calculated with insufficient accuracy, then noise suppression effect can be obtained, but sound signal is also suppressed causing deterioration in sound quality
Solution Approach 1:
Different statistical distribution models are applied to different frequency regions or signal conditions, with each model optimized for specific acoustic characteristics. By assigning appropriate statistical models to different local conditions (frequency bands, signal levels), the system achieves accurate noise suppression in each region while preserving sound quality through region-specific optimization.
Solution Approach 2:
The sound distribution model is constructed as a composite of multiple statistical distribution models combined with weight coefficients. This composite modeling approach allows the system to capture diverse acoustic characteristics that cannot be represented by a single statistical model, thereby improving both noise suppression accuracy and sound quality preservation through the synergistic combination of multiple models.
Data Source
AI summary
A noise suppressing apparatus that calculates a suppression coefficient for suppressing noise of an input signal by using a frequency spectrum of the input signal includes a frequency converting section that converts the input signal into a frequency spectrum; a noise level estimating section that calculates an estimated noise level of the input signal; a weight coefficient calculating section that calculates N (N is 2 or more) weight coefficients at predetermined intervals; and a suppression coefficient calculating section that calculates a joint distribution model of sound by weighting N statistical distribution models with the N weight coefficients, derives an estimation expression for a sound spectrum of the input signal on the basis of posteriori probability using the calculated joint distribution model of sound as priori probability, and calculates the suppression coefficient on the basis of the derived estimation expression and level of the input signal.