Maximum Likelihood Beamforming for Speech Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing beamforming methods, such as MVDR, struggle to accurately estimate noise signals in environments where a target sound is present, leading to deteriorated speech recognition performance due to noise distortion.

Innovation Solution

A beamforming method using maximum likelihood estimation is employed, where a filter is estimated to maximize the log likelihood of a probability density function under the assumption that the beamforming output signal satisfies a complex generalized Gaussian or complex gamma distribution, with the goal of improving sound quality and speech recognition accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If MVDR beamforming algorithm is used to minimize beamforming output signal, then noise removal capability is improved, but speech recognition performance deteriorates when target sound is present

Engineering Contradiction:
Improvenoise distortionVSAvoidspeech recognition performance
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent changes the fundamental parameter of filter estimation from minimizing output power (MVDR) to maximizing log-likelihood under complex generalized Gaussian distribution assumption. This parameter change allows accurate noise estimation even when target sound is present, resolving the contradiction between noise removal and speech recognition performance

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the traditional MVDR mechanical optimization approach with a statistical maximum likelihood estimation approach based on complex generalized Gaussian distribution. This substitution enables more accurate noise signal estimation in the presence of target sound, improving speech recognition reliability

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Object-affected harmful factors

If filter estimation minimizes beamforming output signal power, then noise components are removed, but log likelihood of probability distribution is not maximized

Engineering Contradiction:
Improvenoise signalVSAvoidlog likelihood estimation accuracy
Core Design Contradiction:
Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The patent fundamentally changes the estimation parameter from output power minimization to log-likelihood maximization under complex generalized Gaussian distribution. This parameter change ensures that the filter estimation accurately reflects the statistical properties of the noise signal, improving measurement precision

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent employs an iterative algorithm where the filter estimation and variance estimation mutually refine each other. The algorithm uses the estimated filter to compute variance, then uses the variance to improve filter estimation, creating a self-improving system that maximizes log-likelihood

Inventive Principle:
Principle #25Self-service

Data Source

PatentUS10916239B2Method for beamforming by using maximum likelihood estimation for a speech recognition apparatus
Publication Date: 2021.02.09 MPWAV INC
  • US10916239B2 patent drawing
  • US10916239B2 patent drawing

AI summary

Provided is a method for beamforming by using maximum likelihood estimation in a speech recognition apparatus, including: (a) receiving an input signal (Xn,k) at a time frame n and a frequency k where noise is mixed: (b) determining a probability density function for a target signal (Yn,k) obtained by removing the noise from the input signal satisfies a complex generalized Guassian distribution or a complex gamma distribution where an average value is zero in a time-frequency domain; (c) estimating a variance (λn,k) of the target signal so as to maximize log likelihood for the probability density function; (d) estimating a filter (wk) maximizing a cost function so as to maximize the log likelihood for the probability density function; and (e) repeatedly performing the estimation of the steps (c) and (d) until the filter (wk) coverages, and finally acquiring a final filter (wk).