The cambridge university 2014 BOLT conversational telephone Mandarin Chinese LVCSR system for speech translation

• DOI: 10.21437/interspeech.2015-633
• Publication Date: 2021-08-27T05:58:44Z
• Authors: Xunying Liu; Federico Flego; Linlin Wang; C. Zhang; Mark J. F. Gales; Philip C. Woodland
• Keywords: Mandarin Chinese; Word error rate; Speech translation

This paper presents the development of 2014 Cambridge University conversational telephone Mandarin Chinese LVCSR system for DARPA BOLT speech translation evaluation. A range advanced modelling techniques were employed to both improve recognition performance and provide a suitable integration with system. These include an improved combination technique using frame level acoustic model via joint decoding. Sequence trained deep neural network (DNN) based hybrid tandem systems combined on-the-fly produce consistent decoding output during search. multi-level paraphrastic recurrent LM (RNNLM) alternative paraphrase expressions character sequences while preserving word segmentation was also used. gave overall error rate (CER) 29.1% on dev14 set. Index Terms: transcription, translation, combination, RNNLM,