Voiceprint Registration via Frame Alignment and GMM Stitching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voiceprint registration and authentication systems are affected by noise and pauses, leading to reduced accuracy and performance in voiceprint models.
Innovation Solution
The method involves performing frame alignment on voice input character strings to extract acoustic features, calculating posterior probabilities using a global Gaussian Mixture Model (GMM) with Baum-Welch statistics, and stitching vector features to create and match voiceprint models for registration and authentication.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional voiceprint authentication is used, then the system is simple to operate, but the accuracy is reduced due to noise and pauses
Solution Approach 1:
The patent segments the voiceprint authentication process into multiple independent stages: acoustic feature extraction, posterior probability calculation, Baum-Welch statistics computation, and vector feature stitching. Each stage processes specific aspects of the voice data separately, allowing noise and pause handling to be addressed at appropriate stages without complicating the entire system.
Solution Approach 2:
The patent performs preliminary noise reduction and pause handling during the acoustic feature extraction stage before subsequent processing. By addressing potential issues early in the pipeline, the system prevents noise and pauses from degrading accuracy in later stages without requiring complex real-time intervention mechanisms.
2Measurement precision
If frame alignment and GMM model processing are applied, then the voiceprint model accuracy is improved, but the processing time increases
Solution Approach 1:
The patent divides the processing workflow into distinct sequential stages: frame alignment operation, acoustic feature extraction, posterior probability calculation, Baum-Welch statistics, and vector feature stitching. This segmentation allows each stage to be optimized independently and enables parallel processing where applicable, reducing overall processing time while maintaining accuracy.
Solution Approach 2:
The patent applies frame alignment and GMM model processing selectively to critical portions of the voice data that most impact authentication accuracy. By focusing computational resources on the most informative segments rather than uniformly processing all data, the system achieves high accuracy without unnecessary processing time expenditure.
3Reliability
If acoustic features are extracted for each character, then the authentication reliability is improved, but the computational complexity increases
Solution Approach 1:
The patent extracts acoustic features for each character in the input voice data and processes them through separate computational stages. This character-level segmentation enables precise authentication by analyzing individual character characteristics while allowing each stage to handle computation efficiently through specialized algorithms.
Solution Approach 2:
The patent combines multiple character-level acoustic features into a unified voiceprint model through the Baum-Welch statistics and vector feature stitching stages. By merging processed character features systematically, the system achieves high authentication reliability without requiring each individual feature extraction to be computationally intensive.
Data Source
AI summary
The present disclosure provides methods and devices for registering a voiceprint and authenticating a voiceprint. The method for registering a voiceprint includes performing a frame alignment operation on a registration character string inputted by a user in voice to extract first acoustic features of each first character constituting the registration character string; calculating a first posterior probability of the first acoustic features of each first character in a global Gaussian Mixture Model (GMM) model to perform a Baum-Welch (BW) statistic; extracting first vector features of each first character through a preset vector feature extractor configured for multi-character; and stitching the first vector features of each first character sequentially, to obtain a registration voiceprint model of the user.