A Tibetan speech corpus labeling method and system based on collaborative batch active learning

Abstract

The invention discloses a Tibetan speech corpus tagging method and system based on cooperative batch active learning, wherein the system includes: a sample selection module, a manual tagging module, a tagging decision module, a tagger evaluation module, and a training set generation module. The invention solves the construction of the sample evaluation function and the proof of the nature of the submodular function through the approach to the optimal batch sample selection method, and solves the construction of the labeling decision function, the construction of the labeler evaluation model and the labeler-assisted learning model through the labeling committee collaborative labeling method mold. In addition, the system of the present invention can realize functions such as optimal selection of samples, annotation evaluation of users, sharing of annotation information and Tibetan phonetic knowledge, assisted learning by annotators, and is intended to improve the tagging quality of Tibetan speech data and speed up the construction of audio corpus.

Description

technical field [0001] The invention relates to the fields of speech recognition and corpus training, in particular to a method and system for labeling Tibetan speech corpus based on cooperative batch active learning. Background technique [0002] In the field of speech recognition, traditional speech recognition algorithms (such as HMM, DBNs, ANN, and DTW, etc.) use a supervised learning method to establish a speech recognition model. In order to establish a speech recognition model with high accuracy, this learning method requires a large number of Annotating speech corpus, and annotating speech corpus is an extremely time-consuming and labor-intensive task. Usually, the time spent on tagging words as a speech recognition unit is 10 times the actual audio sentence time (for example, a one-minute speech sentence takes nearly 10 minutes to tag), and the speech tagging work using phonemes as a recognition unit It will reach 400 times the length of the speech sentence (assumi...

AI Technical Summary

Problems solved by technology

Solve the construction of sample evaluation function and the proof of its submodularity property by approaching the optimal batch sample selection method

Through the collaborative labeling method of the labeling committee, the construction of the labeling decision function, the modeling of the labeler evaluation model and the labeler-assisted learning model are solved

Images