AP17:OLR-special session

Title

Minor- and Multilingual speech and language processing

Organizers

Dong Wang: Tsinghua University (wangdong99@mails.tsinghua.edu.cn)

Dong Wang:

Dr. Dong Wang got his PhD degree at the University of Edinburgh, and worked in Oracle, IBM, and Nuance. He is now an assistant professor at the certer for speech and language technologies (CSLT) at Tsinghua University. Dr. Wang’s research interest covers speech processing, language processing and financial processing. He has published more than 80 academic papers in the related area, including three best paper awards. Dr. Wang plays active roles in the speech research community: he serves as the secretary in national conference of machine-man speech communication (NCMMSC) and a country representative of the mainland China in Oriental COCOSDA. He was the local chair of ChinaSIP 2013, special session co-chair of ISCSLP 14 and plenary talk co-chair of ISCSLP 16. Dr. Wang is now serving as the vice Chair of the SLA track of APSIPA.

Guanyu Li: Northwest National University (guanyu-li@163.com)

Mijit Ablimit: Xinjiang University (mijit@xju.edu.cn)

Target track

Speech and Language processing

Introduction

Minor- and multilingual phenomenon is a important for modern international societies. This special session focuses on minor- and multilingual speech and language processing, including but not limited to the following topics:

• Minor- and Multilingual phonetic and phonological analysis
• Minor- and Multilingual speech recognition
• Minor- and Multilingual speaker recognition
• Minor- and Multilingual speech synthesis
• Minor- and Multilingual language understanding
• Resource construction for minor- and multilingual langauges

Potential Papers

Title: Prior-constrained multilingual speech recognition

• Author: Ying Shi, Zhiyuan Tang, Dong Wang

• Abstract: Conventional multilingual speech recognition follows ether a tandem approach (language identification)

or parallel architecture (parallel decoding). This paper presented a novel prior-constrained approach that conduct the decoding in a multilingual linguistic space, where a prior of the language is used to constrain the decoding frame by frame. Our experiments found that this approach can realize true simultaneous multilingual speech recognition.

Title: Memory-based Uyghur-Chinese Translation

• Author: Shiyue Zhang, Guli, Mijit Ablimit, Askar Hamdulla

• Abstract: Neural machine translation (NMT) has achieved significant performance. However, this NMT approach

has not yet effectively applied to minor languages such as Uyghur to Chinese translation. The main problem here is that the limited training data does not support an end-to-end neural learning. In this paper, we propose to use a memory structure to assist the NMT inference under the condition of limited resource languages. Our experiments demonstrated that the this approach is highly efficient compared to the vanilla NMT, and outperforms the conventional statistical machine translation (SMT) approach.

Title: Resource construction for Mongolia

• Author: Shipeng Xu, Guanyu Li, Hongzhi Yu

• Abstract: Mongolia is a typical low-resource language. The resource limitation is in various aspects, from acoustic

analysis, phonetic rules, lexicon, speech and text data. This paper describes our recent progression on Mongolia resource construction supported by the NSFC project.

Title: Resource construction for tibetan

• Author: Guanyu Li, Hongzhi Yu

• Abstract: Tibetan is an important low-resource language in China. The syllable structure of Tibetan is similar

as Chinese, but the composition rules in orthographic forms is highly complex. Additionally, the lexicon resource is far from standard and rich. This paper describes our recent progression on Tibetan resource construction supported by the NSFC M2ASR project.

Title: A large Kazak speech database and a speech recognition baseline

• Author: Askar Hamdulla, Ying Shi

• Abstract: We describe the construction process of a large scale Kazak speech database. The database involves

150 hours of speech signals, recorded by more than 200 speakers. A speech recognition baseline system based on the Kaldi toolkit was also constructed. We hope this database will be a standard dataset for a multiple Kazak speech processing tasks, including ASR, speaker recognition and language understanding.