Voice Model Training Using Segmented Data Subsets

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The continuous expansion of voice data sets leads to significantly increased training times for voice models, particularly in methods involving hidden Markov models (HMM) and deep neural networks (DNN), making existing training processes inefficient.

Innovation Solution

A method that involves selecting a first voice data set to generate training and test sets, training HMM+GMM models, and then using the optimal parameters to perform second voice model training on a second data set only if the similarity condition between the two sets is met, thereby reducing the overall training time by omitting unnecessary training steps.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the entire voice data set is used for training HMM+GMM and HMM+DNN models, then the model accuracy is improved, but the total training time is significantly increased

Engineering Contradiction:
Improvemodel accuracyVSAvoidtotal training time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent divides the voice data set into multiple subsets and processes them in stages. First, HMM+GMM models are trained on the entire data set to obtain initial parameters, then HMM+DNN models are trained on subsets of the data using these parameters as initialization. This segmentation allows the system to maintain accuracy while reducing total training time by avoiding redundant full-data training iterations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary training with HMM+GMM on the entire voice data set before proceeding to HMM+DNN training. The parameters obtained from this preliminary stage are used to initialize the HMM+DNN models, providing a good starting point that reduces the number of iterations needed in subsequent training stages, thereby reducing total training time while maintaining accuracy.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If HMM+GMM and HMM+DNN training are performed sequentially on the entire data set, then comprehensive model performance is achieved, but training efficiency is reduced

Engineering Contradiction:
Improvemodel performanceVSAvoidtraining efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The training process is segmented into distinct stages: first training HMM+GMM on the entire data set, then dividing the data into multiple subsets for HMM+DNN training. This segmentation maintains comprehensive model performance by ensuring all data is represented while improving efficiency by processing subsets in parallel or sequential batches rather than requiring multiple full-data passes.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes training parameters between stages - using full data set for HMM+GMM initialization, then switching to subset-based training for HMM+DNN with modified hyperparameters. This parameter adaptation allows the system to achieve comprehensive performance characteristics from diverse data sampling while reducing the computational burden of processing the entire data set repeatedly.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3594940B1Training method for voice data set, computer device and computer readable storage medium
Publication Date: 2023.07.26 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • EP3594940B1 patent drawingFigure 1~2
  • EP3594940B1 patent drawingFigure 3
  • EP3594940B1 patent drawingFigure 4

AI summary

A method for training a voice data set includes: reading a first test set generated by selecting data from a first voice data set, and obtaining a first voice model parameter by training the first voice data set; obtaining a second voice data set, and randomly selecting data from the second voice data set to generate a second test set; and performing second voice model training for the second voice data set by using the first voice model parameter obtained by training when detecting that the second test set and the first test set satisfy a similarity condition.