Voice Identity Feature Extractor Training for Short Durations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice identity recognition systems face reliability issues when voice duration is short, as traditional I-vectors may not accurately represent speaker information in such cases, leading to reduced recognition accuracy.
Innovation Solution
A novel voice identity feature extractor is developed, which trains a neural network model using I-vectors as target outputs to adjust weights, then fine-tunes using posterior means of I-vector latent variables, resulting in a more reliable F-vector extractor that can accurately represent speaker information even with short voice durations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional I-vector methods are used for voice identity recognition, then the system is simple to implement, but reliability deteriorates when voice duration is short
Solution Approach 1:
The patent segments the feature extraction process into two distinct stages: first training a neural network model to map voice features to I-vectors, then fine-tuning it to extract F-vectors. This segmentation allows each stage to optimize for its specific purpose, improving overall reliability while managing complexity through modular design.
Solution Approach 2:
The patent performs preliminary training of the neural network model using I-vectors as target outputs before the actual F-vector extraction task. This preliminary action pre-adjusts the model weights to be more sensitive to voice identity features, enabling reliable performance even with short voice durations.
2Ease of operation
If short voice durations are accepted for recognition, then user convenience improves, but recognition accuracy deteriorates
Solution Approach 1:
The patent changes the output parameter from traditional I-vectors to F-vectors through neural network transformation. This parameter change enables the system to extract more discriminative voice identity features from short voice durations, maintaining accuracy while improving convenience.
Solution Approach 2:
The patent replaces the traditional mechanical I-vector extraction method with a neural network-based F-vector extraction system. This substitution allows the model to learn complex non-linear mappings from voice features to identity representations, achieving high accuracy with minimal voice input.
3Reliability
If neural network fine-tuning is performed with posterior mean, then feature reliability improves, but training time increases
Solution Approach 1:
The patent performs preliminary training with I-vectors as targets before fine-tuning with posterior means. This preliminary action initializes the model weights in a favorable state, reducing the number of fine-tuning iterations needed and thus minimizing additional training time while achieving improved reliability.
Data Source
AI summary
A voice identity feature extractor training method includes extracting a voice feature vector of training voice. The method may include determining a corresponding I-vector according to the voice feature vector of the training voice. The method may include adjusting a weight of a neural network model by using the I-vector as a first target output of the neural network model, to obtain a first neural network model. The method may include obtaining a voice feature vector of target detecting voice and determining an output result of the first neural network model for the voice feature vector of the target detecting voice. The method may include determining an I-vector latent variable. The method may include estimating a posterior mean of the I-vector latent variable, and adjusting a weight of the first neural network model using the posterior mean as a second target output, to obtain a voice identity feature extractor.