Corpus and transcription system of Chinese Lecture Room

Author

Sheng Li ; Akita, Yuya ; Kawahara, Toshio

Author_Institution

Sch. of Inf., Kyoto Univ., Kyoto, Japan

fYear

2014

fDate

12-14 Sept. 2014

Firstpage

442

Lastpage

445

Abstract

The paper introduces our project on automatic speech recognition (ASR) of Chinese lectures. For a comprehensive study on spontaneous Chinese, we compile a corpus of Chinese Lecture Room (CCLR), which has faithful transcripts and caption texts. Based on the annotated alignment of these texts, we conduct analysis on linguistic phenomena of spontaneous Chinese speech. We also develop a baseline ASR system with this corpus, and refine it with the DNN-HMM framework. By exploiting the lecture data without faithful transcripts and conducting unsupervised speaker adaptation, significant improvement of ASR accuracy is achieved.

Keywords

computer aided instruction; hidden Markov models; linguistics; natural language processing; neural nets; speech recognition; text analysis; ASR accuracy; CCLR; DNN-HMM framework; annotated alignment; automatic speech recognition; baseline ASR system; caption texts; corpus of Chinese Lecture Room; deep neural network; faithful transcripts; hidden Markov models; linguistic phenomena; spontaneous Chinese speech; transcription system; unsupervised speaker adaptation; Acoustics; Adaptation models; Data models; Neural networks; Speech; Speech recognition; Training; acoustic model; lecture; speech recognition;

fLanguage

English

Publisher

ieee

Conference_Titel

Chinese Spoken Language Processing (ISCSLP), 2014 9th International Symposium on

Conference_Location

Singapore

Type

conf

DOI

10.1109/ISCSLP.2014.6936595

Filename

6936595