Registered Voice Extraction Using Similarity-Based Noise Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional voice extraction methods in complex scenarios result in low accuracy due to the inclusion of noise from other objects, leading to incomplete separation of the desired voice signal.
Innovation Solution
A voice processing method utilizing artificial intelligence to extract a registered voice feature, perform initial recognition, determine voice similarity, and filter out noise based on a preset similarity threshold, employing techniques like convolutional neural networks and Fourier transforms to enhance accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If voice extraction is performed directly on complex voice without preprocessing, then processing speed is maintained, but voice extraction accuracy deteriorates due to noise inclusion
Solution Approach 1:
The voice extraction process is divided into multiple stages: initial voice extraction to obtain candidate voice segments, voice similarity determination to identify target voice segments, and noise filtering to remove non-target segments. This segmentation allows systematic improvement of extraction accuracy through structured processing.
Solution Approach 2:
Voice segments are extracted in advance from the complex voice signal, and voice similarity is determined beforehand for each segment. This preliminary processing enables the system to identify and retain only high-similarity segments, improving final extraction accuracy while maintaining efficient processing.
2Measurement precision
If multiple processing steps are added to improve voice extraction accuracy, then extraction quality improves, but processing time increases
Solution Approach 1:
Traditional mechanical voice separation methods are replaced with AI-based voice similarity determination. The system uses trained models to automatically identify target voice segments through similarity comparison, significantly reducing processing time while maintaining high extraction accuracy.
Solution Approach 2:
The system changes the processing parameter from direct signal filtering to voice similarity scoring. By extracting voice segments and evaluating their similarity to target voice characteristics, the system efficiently identifies target segments without requiring complex real-time signal processing, thus reducing processing time.
3Ease of manufacture
If conventional voice extraction methods are used, then processing simplicity is maintained, but noise from other objects remains in the extracted voice
Solution Approach 1:
The system extracts individual voice segments from the complex voice signal and separately evaluates each segment's similarity to the target voice. This extraction approach allows selective retention of only those segments that match target voice characteristics, effectively removing noise from other objects while maintaining processing simplicity.
Solution Approach 2:
Voice similarity serves as an intermediary criterion between the extracted voice segments and the final output. By using similarity determination as a mediating step, the system can identify and retain only target voice segments while filtering out noise, achieving effective noise removal without complex processing.
Data Source
Figure 1
Figure 2~3
Figure 4~6
AI summary
A voice processing method includes: obtaining a registered voice of a speaker, and obtaining a mixed voice, the mixed voice comprising voice data of a plurality of sounding objects, to which the speaker belongs (202); determining a registered voice feature of the registered voice (204); extracting an initially recognized voice of the speaker from the mixed voice based on the registered voice feature (206); determining, for each voice data portion comprised in the initially recognized voice, based on the registered voice feature, a voice similarity between the registered voice and the voice data portion (208); and filtering out, from the initially recognized voice, voice information of which voice similarity is lower than a preset similarity, to obtain a clean voice of the speaker (210).