Search

2 Mar 2000: The 1997 system used N-gram language models trained on 132 million words of broadcast news texts, the LDC-distributed 1995 newswire texts, and the transcriptions from BNtrain97 (LMtrain97). ... The word N-grams were trained by interpolating (and merging)

1 Mar 2000: The system uses a language model trained on 132 million words of broadcast news texts, the LDC-distributed 1995 newswire texts, and the transcriptions from BNtrain96. ... using the broadcast news training texts, the acoustic training data and 1995

15 Feb 2005: Recent Developments at Cambridgein Broadcast News Transcription. D.Y. Kim, H.Y. Chan, G. ... Presentation Overview. • RT03 Broadcast News System Review• Training & Test Data• Improved Acoustic Model Building. –

1 Mar 2000: Experiments in Broadcast News Transcription. P.C. Woodland, T. Hain, S.E. Johnson, T. ... Cross-word context dependent decision tree state clustered mixture Gaussian HMMs are used with a 65k word vocabulary and a language model trained on 132 million

16 Feb 2008: 744.1.3 WFST Computations. 77. 4.2 Source Language Texts, Bitexts, and Phrase-Pair Inventories. ... tion of texts from one natural language (such as French) to another (e.g.

8 Mar 2000: The1997 system used N-gram language models trained on 132million words of broadcast news texts, the LDC-distributed1995 newswire texts, and the transcriptions from BNtrain97(LMtrain97). ... 5.1. Language Models. The word N-grams were trained by

27 May 2004: A word-based 4-gram language model was trained on the acous-tic transcriptions, additional Broadcast News data (427M words oftext) plus 62M words of “conversational texts” collected from theWorld Wide ... Niesler, A.Tuerk, E.W.D. Whittaker and S.J.

27 May 2004: IMPROVING BROADCAST NEWS TRANSCRIPTION BY LIGHTLY SUPERVISEDDISCRIMINATIVE TRAINING. H.Y. Chan & P.C. ... These include the TDT2,TDT3 and TDT4 closed captions, broadcast news acoustic train-ing data transcriptions, commercial broadcast news transcripts

10 Apr 2000: Cross-word context dependent deci-sion tree state clustered mixture Gaussian HMMs are used with a65k word vocabulary and a language model trained on 132 millionwords of broadcast news texts, along ... Young S.J. (1997) TheDevelopment of the 1996

8 Mar 2000: 132 million words of broadcast news texts, the LDC-distributed 1995 newswire texts, and the transcriptions of the acoustic training data. ... Furthermore new language models were estimated from an extended set of broadcast news transcripts and newspaper

8 Mar 2000: 4.6. Static Language Models. The evaluation system language model (LM) had a word listcontaining 65423 words chosen from the most frequent wordsin the broadcast news training texts, with the most ... 5. CONCLUSIONThis paper has described our initial

8 Mar 2000: The 1997 HTK Broadcast News Transcription SystemP.C. Woodland, T. Hain, S.E. ... These were trained on the 132million words of the LDC broadcast news training texts,the transcriptions of the BNtrain97 data (added twice), the1995 newswire texts (both

2 Mar 2000: Our retrieval system was run on 7 different sets of automatically transcribed broadcast news texts with a WER varying from 24.8% to 66.2%. ... J. Young. Segment Generation and Clustering in the HTK Broadcast News Transcription System [in Proc.

5 Oct 2000: The system uses word-based N-gram LMs estimated from an interpolation of Hub5 acoustic training transcriptions and Broadcast News texts.

30 Mar 2000: Three fixed backoff word-based language models were trained, from broadcast news text, newspaper texts and acoustic transcriptions, which were all generated using data from before 31st January 1998. ... 1996 to Jan. 1998 (from the Primary Source Media

8 Mar 2000: The paper is organised as follows. Firstly an overview ofthe HTK broadcast news transcription system used to generatehigh quality automatic transcriptions is given. ... of broadcast news texts, the LDC-distributed 1995newswire texts, and the

23 Sep 2003: BN LMswere trained on the acoustic transcripts (2 million words), anumber of other broadcast news transcript sources (343Mwords) plus a variety of newspaper texts (674M words).More details on the task ... More details on the effectiveness of these

10 Apr 2000: different sets ofautomatically transcribed broadcast news texts with aWER varying from24.8% to 66.2%. ... The input data is presented to the system as com-plete episodes of broadcast news shows and these are.

10 Apr 2000: Three fixed backoff word-based language models were trained,from broadcast news text, newspaper texts and acoustic tran-scriptions, which were all generated using data from before 31stJanuary 1998. ... The first model was built using 190 million wordsof

6 Nov 2000: It was reported in [32] that FD didn’t improveerror rates over MLE trained models for a broadcast newsrecognition task. ... The system uses word-based N-gramLMs estimated from an interpolation of Hub5 acoustic train-ing transcriptions and Broadcast