Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Multilingual model training method based on cross-language self-training

A model training and multilingual technology, applied in speech analysis, speech recognition, instruments, etc., can solve the problem that the teacher's model cannot be decoded, and achieve the effect of enhancing the ability of semantic representation

Pending Publication Date: 2021-09-03
UNIV OF SCI & TECH OF CHINA
View PDF0 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

Since the teacher model cannot decode the speech of other languages, the self-training method can only use the data of this language

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Multilingual model training method based on cross-language self-training
  • Multilingual model training method based on cross-language self-training
  • Multilingual model training method based on cross-language self-training

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0026] The current unsupervised pre-training method has the limitation of insufficient semantic representation, while the supervised pre-training method has the limitation of insufficient data. Therefore, the present invention proposes a multilingual training method based on cross-language self-training, with the purpose of using a small amount of labeled data to enhance the semantic representation ability of the model on multilingual unlabeled data. The model pre-trained by this method can be used as a multilingual general initialization model and migrated to any low-resource language to improve the accuracy of the low-resource language ASR model.

[0027] The present invention proposes a model training method based on Cross-lingual Self-training (XLST). In the framework of this training model, we assume that frame-level speech representations have shared properties across languages. The method first trains an acoustic phoneme classifier on labeled data in a high-resource la...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention provides a multilingual model training method based on cross-language self-training. The method comprises the steps: firstly training an acoustic phoneme classifier on labeled data of a certain high-resource language as a target network, and then training a main network to approach the characterization of the acoustic phoneme classifier in multiple languages. The method specifically comprises: obtaining a target network; training the main network; and migrating the trained main network to the automatic speech recognition model of the target language.

Description

technical field [0001] The invention relates to the field of low-resource speech recognition and speech representation learning, in particular to a multilingual model training method based on cross-language self-training. Background technique [0002] The current advanced Automatic Speech Recognition (ASR) models usually need to be trained on hundreds or thousands of labeled data, and this scale of labeled data is usually difficult to obtain in low-resource languages. The pre-training method can effectively solve the data problem of low-resource ASR. It first pre-trains a model through other resources (data in other languages ​​or unlabeled data in this language), and then migrates the model to the low-resource ASR model. [0003] Existing pre-training methods can be divided into supervised and unsupervised methods. Early research focused on supervised pre-training, in which the pre-trained model is usually trained on labeled data in one or more other languages, and then us...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G10L15/00G10L15/02G10L15/06G10L15/16
CPCG10L15/005G10L15/02G10L15/063G10L15/16G10L2015/025G10L2015/0631
Inventor 张自强戴礼荣
Owner UNIV OF SCI & TECH OF CHINA
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products