Single-Channel Speech Role Separation via Iterative Re-Clustering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The accuracy of unsupervised role separation based on single-channel speech in conference systems is relatively low, particularly in determining the number of roles and their speaking times without prior role information.

Innovation Solution

A method involving segmenting audio signals into segments, performing clustering based on feature information, determining cluster centers, and iteratively refining these centers to improve role separation accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If unsupervised role separation is performed on single-channel speech, then the system can identify roles without prior information, but the accuracy of role separation is relatively low

Engineering Contradiction:
Improveability to perform role separation without prior role informationVSAvoidaccuracy of role separation
Core Design Contradiction:
Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The audio signal is divided into multiple audio segments, which are then processed through clustering algorithms. This segmentation allows the system to handle complex single-channel speech by breaking it down into manageable parts, improving both adaptability and accuracy simultaneously

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transforms the problem from traditional single-channel processing by introducing multiple processing dimensions through clustering algorithms that analyze audio segments from different feature perspectives, thereby improving accuracy while maintaining the single-channel input constraint

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

2Measurement precision

If clustering processing is performed on audio segments, then role identification can be achieved, but errors in cluster assignments occur

Engineering Contradiction:
Improverole identification accuracyVSAvoidcluster assignment accuracy
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements iterative optimization where clustering results are evaluated and fed back to refine cluster centers. This feedback mechanism continuously improves cluster assignment accuracy, reducing errors while maintaining high role identification precision

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs preliminary clustering to identify initial cluster centers before final role assignment. This preliminary action allows for error correction and refinement, ensuring both high precision and reliability in the final cluster assignments

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4375988B1Audio signal processing method and apparatus, and device and storage medium
Publication Date: 2026.01.14 ZHEJIANG ALIBABA ROBOT CO LTD
  • EP4375988B1 patent drawingFigure 1~2
  • EP4375988B1 patent drawingFigure 3~4
  • EP4375988B1 patent drawingFigure 5

AI summary

The present disclosure relates to an audio signal processing method and apparatus, a device and a storage medium. The present disclosure performs a segmenting processing on an audio signal to obtain multiple audio segments, performs a clustering processing on the multiple audio segments according to feature information of each audio segment in the multiple audio segments to obtain one or more first sets, determines a first cluster center of each first set according to the feature information of the audio segment included in each first set, and performs a clustering processing on the multiple audio segments according to the first cluster center of each first set to obtain one or more second sets, where audio segments in a same second set corresponding to a same role label. That is to say, after an initial clustering processing is performed on the multiple audio segments, a re-clustering processing can also be performed on the multiple audio segments according to the first cluster center of each first set, thereby improving an accuracy of an unsupervised role separation based on a single channel speech.