This paper describes the ITC-irst systems used in the TC-STAR'06 evaluation campaign for transcribing parliamentary speeches delivered in both English and Spanish languages. Systems use a three pass decoding strategy with cluster-based unsupervised acoustic models adaptation. Both rst and second decoding passes use a trigram language model, while the third decoding pass employs a fourgram language model. Acoustic and language models of both English and Spanish transcription systems were trained exploiting the language resources released for the TC-STAR evaluation campaign of year 2006. An additional language resource, i.e. a 200M word text corpus distributed by the Linguistic Data Consortium (LDC), was also utilized to train language models for English. The Word Error Rates (WERs) of the primary English transcription system were 13.0% and 11.0% on the EPPS English development and evaluation data sets, respectively. On both the Spanish development and evaluation data sets, which include both EPPS and Spanish Parliament speech data, the transcription system provided a WER of 13.3%.
The ITC-irst Transcription Systems for the TC-STAR-06 Evaluation Campaign
Brugnara, Fabio;Falavigna, Giuseppe Daniele;Giuliani, Diego;Gretter, Roberto;Seppi, Dino;Stemmer, Georg
2006-01-01
Abstract
This paper describes the ITC-irst systems used in the TC-STAR'06 evaluation campaign for transcribing parliamentary speeches delivered in both English and Spanish languages. Systems use a three pass decoding strategy with cluster-based unsupervised acoustic models adaptation. Both rst and second decoding passes use a trigram language model, while the third decoding pass employs a fourgram language model. Acoustic and language models of both English and Spanish transcription systems were trained exploiting the language resources released for the TC-STAR evaluation campaign of year 2006. An additional language resource, i.e. a 200M word text corpus distributed by the Linguistic Data Consortium (LDC), was also utilized to train language models for English. The Word Error Rates (WERs) of the primary English transcription system were 13.0% and 11.0% on the EPPS English development and evaluation data sets, respectively. On both the Spanish development and evaluation data sets, which include both EPPS and Spanish Parliament speech data, the transcription system provided a WER of 13.3%.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
