Multi-channel Speech Separation via Beamforming and Neural Networks

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current audio signal processing systems fail to effectively separate and recognize individual speakers in highly overlapped speech environments, commonly known as the 'cocktail party problem', due to large variations in mixing sources, limiting their performance in applications like speech recognition and transcription.

Innovation Solution

The system employs a microphone array coupled with a beamformer for spatial filtering, followed by a source separator using a machine learning algorithm that learns a ratio mask for each source, and a post-selection process involving spectral clustering and speech quality detection to isolate and separate audio sources, utilizing techniques like permutation invariant training to address the label permutation problem.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional audio signal processing methods are used, then the system structure remains simple, but the speech separation accuracy deteriorates in highly overlapped environments

Engineering Contradiction:
Improvespeech separation accuracyVSAvoidsystem structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the mixed speech signal processing into multiple independent single-channel separation tasks by using multiple beamformers to create separate audio channels, each processed by its own neural network separator, thereby improving separation accuracy while managing complexity through modular architecture

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system transitions from single-channel to multi-channel processing by introducing spatial dimension through microphone arrays and beamforming, creating multiple parallel processing paths that improve speech separation accuracy in overlapping environments

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If multiple channel separation architecture is used, then speech separation accuracy improves, but the computational complexity increases

Engineering Contradiction:
Improvespeech separation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The computational load is segmented and distributed across multiple independent single-channel separation tasks, each handled by a dedicated neural network processor, allowing parallel computation that improves accuracy while managing overall computational complexity

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system applies partial separation by focusing each neural network on extracting specific speech sources from its assigned channel, performing sufficient separation for each task rather than attempting complete simultaneous separation of all sources, thereby reducing computational complexity

Inventive Principle:
Principle #16Partial or excessive action

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

This approach effectively separates and recognizes multiple speakers, improving speech recognition and transcription accuracy by isolating individual audio sources, even in complex environments, and can be applied in various settings such as hearing aids, meeting transcription, and music information retrieval.

Implementation Method 1

a beamformer to output multiple channels of audio such that each channel isolates audio in a particular direction

Methodology Applied
Scientific EffectBeamforming:

Implementation Method 2

utilizing techniques like permutation invariant training to address the label permutation problem

Methodology Applied
Scientific EffectPermutation invariant training:

Data Source

PatentEP3707716B1Multi-channel speech separation
Publication Date: 2021.12.01 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP3707716B1 patent drawingFigure 1
  • EP3707716B1 patent drawingFigure 2
  • EP3707716B1 patent drawingFigure 3

AI summary

Representative embodiments disclose mechanisms to separate and recognize multiple audio sources (e.g., picking out individual speakers) in an environment where they overlap and interfere with each other. The architecture uses a microphone array to spatially separate out the audio signals. The spatially filtered signals are then input into a plurality of separators, so each signal is input into a corresponding signal. The separators use neural networks to separate out audio sources. The separators typically produce multiple output signals for the single input signals. A post selection processor then assesses the separator outputs to pick the signals with the highest quality output. These signals can be used in a variety of systems such as speech recognition, meeting transcription and enhancement, hearing aids, music information retrieval, speech enhancement and so forth.