8th European Conference on Speech Communication and Technology (Eurospeech 2003) 2003
DOI: 10.21437/eurospeech.2003-102
|View full text |Cite
|
Sign up to set email alerts
|

The 300k LIMSI German broadcast news transcription system

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1

Citation Types

0
3
0

Year Published

2004
2004
2012
2012

Publication Types

Select...
3
3
2

Relationship

0
8

Authors

Journals

citations
Cited by 18 publications
(4 citation statements)
references
References 6 publications
0
3
0
Order By: Relevance
“…The performance of the ASR is displayed in table 2. Taking the diversity of the test corpus with its high amount of spontaneous speech into account, the word error rates (WER) is comparable to the performance of other systems [10]. The better performance of the sub-word based recognition can be explained by outof-vocabulary (OOV) effects in the word recognition, For evaluating the NER, precision, recall and the fmeasure are defined in the usual way [8].…”
Section: Resultsmentioning
confidence: 99%
See 1 more Smart Citation
“…The performance of the ASR is displayed in table 2. Taking the diversity of the test corpus with its high amount of spontaneous speech into account, the word error rates (WER) is comparable to the performance of other systems [10]. The better performance of the sub-word based recognition can be explained by outof-vocabulary (OOV) effects in the word recognition, For evaluating the NER, precision, recall and the fmeasure are defined in the usual way [8].…”
Section: Resultsmentioning
confidence: 99%
“…The large word count is necessary because of the compounding strategy inherent in the German language [10]. Grapheme-to-phoneme conversion for the word pronunciation lexicon was carried out using the transcription module of the Bonn Open Source Synthesis System (BOSSII) developed by the Institut für Kommunikationsforschung und Phonetik of Bonn University [4].…”
Section: Speech Recognition Systemmentioning
confidence: 99%
“…A more detailed description of the baseline system can be found in [1]. The baseline ASR system yields a performance comparable to other state of the art systems for German such as [2] (see section 4), yet the word error rate was still to high for displaying the corresponding transcripts to the end users of the ARD Mediathek. We apply a twofold adaptation strategy in order to reduce the mismatch between our baseline ASR model and the heterogeneous ARD data: acoustic and language model adaptation.…”
Section: Automatic Speech Recognitionmentioning
confidence: 99%
“…The German BN transcription system (DE e ) in [McTait & Adda-Decker, 2003] e enhances the system in [Lamel & Gauvain, 2002] d essentially by incorporating new language data for the estimation of AM and LM, and by reducing the effects of intense compounding of German words on the lexical coverage by scaling up its size from 65k to 300k words. Noteworthily, decomposing compounds at morpheme boundaries would allow the constitution of a virtually infinite vocabulary and thereby would maintain a relatively small effective vocabulary.…”
Section: Recognition Task Complexitymentioning
confidence: 99%