Two-Stage NMF Speech Modeling for Residual Noise Suppression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech enhancement systems often suffer from residual noise due to under-estimation of noise references and corrupted speech envelopes in sub-bands, leading to suboptimal quality of noise-suppressed speech outputs.

Innovation Solution

The method employs a two-stage dictionary-based approach using Non-negative Matrix Factorization (NMF) for speech modeling, where a first stage reduces residual noise using a Wiener filter and a second stage restores the speech envelope, utilizing speaker-specific dictionaries and real-time noise dictionaries to differentiate noise and speech harmonics effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single-stage noise suppression method is used, then the processing complexity is low, but residual noise remains due to under-estimation of noise references

Engineering Contradiction:
Improveprocessing complexityVSAvoidnoise suppression accuracy
Core Design Contradiction:
Device complexityVSReliability

Solution Approach 1:

The noise suppression process is divided into two distinct stages: a first stage that performs initial noise suppression to reduce computational complexity, and a second stage that performs residual noise suppression to improve accuracy. This segmentation allows the system to balance between processing complexity and noise suppression accuracy by handling different noise components in separate stages with appropriate algorithms for each.

Inventive Principle:
Principle #1Segmentation

2Reliability

If aggressive noise suppression is applied, then residual noise is reduced, but speech envelope corruption occurs in sub-bands

Engineering Contradiction:
Improvenoise suppression accuracyVSAvoidspeech envelope integrity
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The system applies different processing strategies to different frequency sub-bands based on their local characteristics. In sub-bands dominated by speech, the system preserves the speech envelope structure, while in noise-dominated sub-bands, more aggressive suppression is applied. This local quality approach ensures that speech information is preserved where present while effectively suppressing noise where dominant.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The noise suppression algorithm dynamically adjusts its behavior based on the local signal-to-noise ratio and speech activity detection in each sub-band. The system adapts the suppression strength and envelope preservation parameters according to the instantaneous characteristics of each frequency band, allowing optimal performance across varying noise conditions without corrupting speech envelopes.

Inventive Principle:
Principle #15Dynamics

3Measurement precision

If real-time noise dictionary is obtained based on noisy speech spectrum, then noise modeling accuracy is improved, but computational load increases

Engineering Contradiction:
Improvenoise modeling accuracyVSAvoidcomputational load
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system performs preliminary noise dictionary construction during periods when noise is dominant or speech activity is low, preparing noise models in advance before they are needed for suppression. This preliminary action allows the computationally intensive dictionary construction to be performed when it has the greatest impact, reducing the real-time computational load during active speech periods while maintaining high noise modeling accuracy.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3111445B1Systems and methods for speaker dictionary based speech modeling
Publication Date: 2018.12.19 QUALCOMM INC
  • EP3111445B1 patent drawingFigure 1
  • EP3111445B1 patent drawingFigure 2
  • EP3111445B1 patent drawingFigure 3

AI summary

A method for speech modeling by an electronic device is described. The method includes obtaining a real-time noise reference based on a noisy speech signal. The method also includes obtaining a real-time noise dictionary based on the real-time noise reference. The method further includes obtaining a first speech dictionary and a second speech dictionary. The method additionally includes reducing residual noise based on the real-time noise dictionary and the first speech dictionary to produce a residual noise-suppressed speech signal at a first modeling stage. The method also includes generating a reconstructed speech signal based on the residual noise-suppressed speech signal and the second speech dictionary at a second modeling stage.