Speaker-Aware Audio Processing for Reference-Free Echo Suppression

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional acoustic echo cancellation systems are limited by the need for a reference signal, and struggle to effectively suppress interfering voice signals, particularly when both the desired and interfering signals are voice-based, especially in environments where a time domain reference signal is not available.

Innovation Solution

A machine learning system, such as a neural-network, is used to suppress acoustic interference by determining and adapting to the characteristics of interfering sound sources, utilizing embeddings or other sound characteristics to identify and cancel unwanted sound components without requiring a reference signal.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If conventional echo cancellation is used, then acoustic interference can be suppressed, but the system requires a reference signal which is not always available

Engineering Contradiction:

Improveacoustic interference suppressionVSAvoidapplicability when reference signal is unavailable

Core Design Contradiction:

Object-affected harmful factorsVSAdaptability or versatility

Solution Approach 1:

The system uses self-service by extracting speaker characteristics directly from the microphone signal itself rather than requiring an external reference signal. The embedding model processes the microphone signal to identify and cancel acoustic interference components, enabling the system to serve itself without external reference input.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent introduces an intermediary approach by using embedding models and machine learning systems as mediators between the microphone signal and the cancellation process. These intermediaries extract speaker characteristics and generate cancellation signals without requiring direct reference signals, bridging the gap when reference signals are unavailable.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Object-affected harmful factors

If conventional echo cancellation is used, then acoustic interference can be suppressed, but it struggles to effectively suppress interfering voice signals when both desired and interfering signals are voice-based

Engineering Contradiction:

Improveinterfering voice signal suppressionVSAvoidvoice signal differentiation accuracy

Core Design Contradiction:

Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The system applies local quality by processing different aspects of the audio signal locally - extracting speaker-specific characteristics through embedding models and applying targeted cancellation only to the interfering voice components. This allows precise differentiation between desired and interfering voice signals based on their unique acoustic fingerprints.

Inventive Principle:

Principle #3Local quality

Solution Approach 2:

The patent changes parameters by transforming the approach from time-domain signal processing to embedding-based characteristic extraction. The embedding model converts voice signals into compact representations of speaker characteristics, enabling more precise identification and cancellation of interfering voices based on their unique acoustic parameters.

Inventive Principle:

Principle #35Parameter changes

3Adaptability or versatility

If speaker characteristics are extracted from microphone signal, then reference signal is not needed, but the system complexity increases

Engineering Contradiction:

Improveoperation without reference signalVSAvoidmachine learning system and embedding model

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system uses preliminary action by pre-training embedding models to extract speaker characteristics from microphone signals. These pre-computed embeddings are then used during operation to identify and cancel acoustic interference, reducing the computational burden during real-time operation while maintaining the ability to operate without reference signals.

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS12477291B2Acoustic interference suppression through speaker-aware processing

Publication Date: 2025.11.18 CERENCE OPERATING CO

AI summary

Disclosed are systems, methods, and other implementations for acoustic interference suppression, including a method that includes obtaining a multi-source sound signal sample combining multiple sound components from a plurality of sound sources in a sound environment, with the plurality of sounds sources including one or more interfering sound sources produced by one or more loudspeakers in the sound environment, determining interfering sound characteristics for one or more sound signals that correspond to the one or more interfering sound sources, and suppressing at least one of the multiple sound components associated with the determined interfering sound characteristics for at least one of the one or more sound signals.