Speaker Recognition System Noise Filtering Spectral Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speaker recognition methods are inefficient when dealing with noisy speech signals and require large amounts of speech samples, leading to reduced accuracy and increased computational burden.

Innovation Solution

A speaker recognition system that pre-filters speech signals to remove noise and non-speech portions, using spectral analysis and energy estimation to discard frames with high noise or no speech, followed by feature extraction and classification using Support Vector Machines.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If speech signals are pre-filtered to remove noise and non-speech portions, then speaker recognition accuracy is improved, but device complexity increases

Engineering Contradiction:
Improvespeaker recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by performing spectral analysis and energy estimation on speech signals before the main speaker recognition process. The system pre-filters frames to remove noise and non-speech portions ahead of time, so that only clean speech portions are processed in the subsequent speaker recognition stages, thereby improving accuracy without overwhelming the main processing system

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the speech signal into discrete frames and applies filtering operations to individual frames based on their spectral and energy characteristics. This segmentation allows the system to process only relevant portions of the signal, reducing overall computational complexity while maintaining high accuracy for clean speech segments

Inventive Principle:
Principle #1Segmentation

2Reliability

If speech signals are pre-filtered using spectral analysis and energy estimation, then noise impact is reduced, but computational burden increases

Engineering Contradiction:
Improvenoise resistanceVSAvoidcomputational burden
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The patent extracts only the essential spectral and energy features from each speech frame that are necessary for noise detection and removal. By taking out only the critical parameters needed for filtering decisions, the system achieves effective noise resistance without computing unnecessary features, thereby reducing overall computational burden

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies partial action by performing filtering operations selectively on frames that require it, based on spectral flatness and energy criteria. Not all frames undergo the complete filtering process - only those identified as containing noise or non-speech portions, thus avoiding excessive computation on already clean segments

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If large amounts of speech samples are used to generate Universal Background Model, then model accuracy is improved, but processing time increases

Engineering Contradiction:
Improvemodel accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary filtering and selection of speech samples before they are used to train the Universal Background Model. By pre-processing the training data to remove noisy and non-speech portions in advance, the system achieves high model accuracy using fewer, higher-quality samples, thereby reducing the processing time required compared to using large amounts of unfiltered data

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3613040B1Speaker recognition method and system
Publication Date: 2024.03.13 TELECOM ITALIA SPA
  • EP3613040B1 patent drawingFigure 1
  • EP3613040B1 patent drawingFigure 2
  • EP3613040B1 patent drawingFigure 3

AI summary

A speaker recognition system (100) for assessing the identity of a speaker through a speech signal based on speech uttered by said speaker is provided. The system comprises: - a framing module (110) configured to subdivide said speech signal over time into a set of frames; - a filtering module (115) configured to analyze the frames of the set to discard frames affected by noise and frames which do not comprise a speech, based on a spectral analysis of the frames; - a feature extraction module (120) configured to extract audio features from frames which have not been discarded; - a classification module (140) configured to process the audio features extracted from the frames which have not been discarded for assessing the identity of the speaker.