Multi-Step Linear Prediction for Late Reverberation Removal

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing dereverberation techniques struggle with accurately removing reverberation in environments where the room transfer function contains many maximum-phase components, leading to decreased accuracy in speech recognition and other applications.

Innovation Solution

A multi-channel multi-step linear prediction model is employed, using discrete acoustic signals from multiple sensors to accurately estimate and eliminate late reverberation, even in environments with complex room transfer functions, by pre-whitening the signals and calculating linear prediction coefficients that account for both direct and reverberant components.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If exponential function is used to estimate late reverberation energy, then the processing is simple and fast, but the accuracy deteriorates in environments with complex room transfer functions containing many maximum-phase components

Engineering Contradiction:
Improveprocessing simplicityVSAvoiddereverberation accuracy
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent changes the fundamental parameter used for reverberation estimation from a simple exponential decay model to a multi-step linear prediction model. This model uses multiple prediction steps (e.g., 3-10 steps) with different time constants to capture the complex decay characteristics of late reverberation in environments with many maximum-phase components, thereby improving accuracy while maintaining computational efficiency

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent introduces a dynamic multi-step linear prediction model that adapts to different reverberation conditions. The model uses multiple prediction steps with varying time constants that can be adjusted based on the observed signal characteristics, allowing it to dynamically adapt to complex room transfer functions while maintaining processing efficiency

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If multi-channel multi-step linear prediction model is used to accurately estimate late reverberation, then the dereverberation accuracy is improved, but the device complexity increases

Engineering Contradiction:
Improvedereverberation accuracyVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the late reverberation estimation into multiple discrete prediction steps, where each step uses a specific time constant. This segmentation allows the complex estimation problem to be broken down into manageable computational tasks, improving accuracy while keeping each individual prediction step computationally efficient

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The multi-step linear prediction model serves multiple functions: it estimates late reverberation energy, captures complex decay characteristics, and adapts to different room conditions all within a single unified framework. This multi-functionality reduces the need for multiple separate processing modules, thereby managing complexity while achieving high accuracy

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP1993320B1Reverberation removal device, reverberation removal method, reverberation removal program, and recording medium
Publication Date: 2015.01.07 NIPPON TELEGRAPH & TELEPHONE CORP
  • EP1993320B1 patent drawingFigure 1
  • EP1993320B1 patent drawingFigure 2A~2B
  • EP1993320B1 patent drawingFigure 3

AI summary

A model application unit 10b calculates linear prediction coefficients of a multi-step linear prediction model by using discrete acoustic signals. Then, a late reverberation predictor 10c calculates linear prediction values obtained by substituting the linear prediction coefficients and the discrete acoustic signals into linear prediction term of the multi-step linear prediction model, as predicted late reverberations. Next, a frequency domain converter 10d converts the discrete acoustic signals to discrete acoustic signals in the frequency domain and also converts the predicted late reverberations to predicted late reverberations in the frequency domain. A late reverberation eliminator 10e calculates relative values between the amplitude spectra of the discrete acoustic signals expressed in the frequency domain and the amplitude spectra of the predicted late reverberations expressed in the frequency domain, and provides the relative values as predicted amplitude spectra of a dereverberation signal.