Conference Audio Voice Separation for Multi-Speaker Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing systems face challenges in accurately recognizing semantics from voice data due to low accuracy in voice recognition, particularly in environments with multiple speakers and potential voice overlapping.
Innovation Solution
An audio processing method that includes determining voices of multiple targets in a segment of audio, performing voice separation and recognition, and utilizing multi-channel filtering and concurrent voice recognition to improve accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If voice recognition is performed on audio captured by a system, then voice recognition can be conducted, but accuracy of determined semantics is low
Solution Approach 1:
The patent segments the audio stream into multiple voice paragraphs based on voice activity detection, and further segments each voice paragraph into multiple voice sections based on pitch contour analysis. This multi-level segmentation allows the system to process and recognize semantics of each target's voice separately, significantly improving the accuracy of determined semantics compared to processing the entire audio stream as a single unit.
2Measurement precision
If voice separation is performed on all to-be-processed voices, then voice recognition accuracy can be improved, but resource occupation increases
Solution Approach 1:
The patent applies voice separation only to specific voice sections where multiple targets are detected to speak alternately, rather than performing voice separation on all to-be-processed voices. By identifying segments with alternating speech patterns and applying separation only to those segments, the system improves recognition accuracy for multi-speaker portions while minimizing resource consumption on single-speaker segments.
3Reliability
If overlap detection is performed on each to-be-processed voice, then voice overlapping can be identified, but resource occupation increases
Solution Approach 1:
The patent performs pitch contour analysis and target identification on voice sections before performing overlap detection. By preliminarily analyzing the pitch contours to identify alternating speech patterns and detecting which sections contain multiple targets, the system can then perform overlap detection only on those specific sections rather than on every to-be-processed voice, thereby reducing overall resource occupation while maintaining reliable voice overlapping detection.
Data Source
AI summary
An audio processing method and an electronic apparatus are provided. The audio processing method is applied to a conference system, and the conference system includes at least one audio capturing device. The audio processing method includes: receiving at least one segment of audio captured by the at least one audio capturing device; determining voices of a plurality of targets in the at least one segment of audio; and performing voice recognition on a voice of each of the plurality of targets, to obtain semantics corresponding to the voice of each target. Voice recognition is separately performed on voices of different targets, thereby improving accuracy of voice recognition.