Speech Dereverberation Filter Coefficient Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional dereverberation methods for automatic speech recognition systems require high computation costs, previous knowledge of reverberation times, or large CPU resources, making them impractical for embedded systems that need real-time processing and cannot utilize room-specific information after deployment.
Innovation Solution
A method that divides speech signal data into frames and computes filter coefficients to minimize the weighted sum of residual speech power in a reverberation segment and subtracted speech power, allowing for effective dereverberation without requiring extensive computational resources or prior knowledge of reverberation characteristics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Hidden Markov Model and Expectation Maximization algorithm are used to determine the coefficient, then the dereverberation performance is improved, but the computation cost increases and supervised training is required
Solution Approach 1:
The patent extracts only the essential statistical properties of reverberation (exponential attenuation of speech power spectrum) and uses them to derive a simple coefficient determination method. By taking out the core characteristic and discarding complex HMM/EM training mechanisms, the solution achieves good dereverberation performance with much lower computation cost.
Solution Approach 2:
The method determines the coefficient automatically by analyzing the statistical properties of the observed speech signal itself, without requiring external supervised training data or acoustic models. The system serves itself by using the input signal's own characteristics to adapt the coefficient.
2Reliability
If inverse filter is used for dereverberation, then the dereverberation effect is improved, but the device complexity increases and multiple microphones are required
Solution Approach 1:
The patent extracts only the power spectrum information and statistical attenuation properties of reverberation, ignoring the complex phase information and impulse response characteristics that make inverse filtering difficult. This extraction of essential features enables a simpler single-microphone solution.
Solution Approach 2:
Instead of using complex inverse filters that require precise room impulse response measurement and multiple microphones, the patent employs a simple spectral subtraction method with a lightweight coefficient that can be quickly adapted online without expensive preparatory measurements.
3Reliability
If comb filter is used to form transfer function, then the dereverberation is achieved, but the computation cost increases and fundamental problems remain
Solution Approach 1:
The patent extracts only the essential statistical property that reverberation power spectrum exponentially attenuates over time, discarding the complex comb filter structure and its assumptions about harmonic structure. This simplified extraction achieves dereverberation with much lower computation cost.
4Productivity
If spectral subtraction with fixed coefficient is used, then the computation cost is reduced, but the adaptability to different reverberation environments deteriorates
Solution Approach 1:
The patent makes the coefficient dynamic by deriving it from the statistical properties of the observed speech signal itself. The coefficient adapts automatically to different reverberation environments through online analysis of the signal's power spectrum attenuation characteristics, without requiring fixed pre-determined values.
Solution Approach 2:
The system determines its own coefficient by analyzing the input signal's statistical properties, making the adaptation process self-service and automatically adjustable to different reverberation conditions without external intervention or training data.
Data Source
AI summary
Method and computing apparatus for processing speech signal data. A speech signal is divided into frames. Each frame is characterized by a frame number T representing a unique interval of time. Each speech signal is characterized by a power spectrum with respect to frame T and frequency band ω. A speech segment and a reverberation segment of the speech signal is determined. L filter coefficients W(k) (k=1, 2, . . . , L) respectively corresponding to L frames immediately preceding frame T are computed such that the L filter coefficients minimize a function Φ that is a linear combination of sum of squares of a residual speech power in the reverberation segment and a sum of squares of a subtracted speech power in the speech segment. The computed L filter coefficients are stored within storage media of the computing apparatus.