Study Of Speaker Adaptation Based On Neural Network Acoustic Model

Although in recent years,speech recognition has become mature,great process has been made in speaker-independent(SD)tasks,and the recognition rate of speech recognition has also been significantly improved.But there is still a gap between the speaker-independent(SI)speech recognition system and the speaker-dependent speech system.Speaker adaptation technologies adapt a SI acoustic model to a SD acoustic model by using a small amount of adaptation datasets of target speaker.In the process of speaker adaptation,a major problem is the spareness of the data.Existing speaker adaptation technologies do not essentially solve the data sparsity problem during speaker adaptation.This thesis proposed a speaker adaptation method based on auxiliary data regularization under the LF-MMI criterion.The main innovations are as follows:1)This paper explores the supervised and unsupervised speaker adaptation under the LF-MMI criterion,and selects the appropriate lattice for subsequent speaker adaptation experiments.2)The problem of data sparsity is analyzed.A speaker adaptation method based on auxiliary data regularization is proposed.The auxiliary data is introduced to alleviate the sparsity problem of data.The adaptation data is expanded and the speaker adaptation loss function is modified to improve the performance of the model.In this paper,experiments are carried out on Librispeech and AISHELL1 datasets.The experimental results show that,compared with the speaker independent model,the relative improvement of supervised and unsupervised speaker adaptation under 1minute speech is 18% and 12% respectively,and the improvement of consistency is also achieved on AISHELL1.