AN INVESTIGATION INTO THE INTERACTIONS BETWEEN SPEAKER DIARISATION SYSTEMS AND AUTOMATIC SPEECH TRANSCRIPTION
S. E. Tranter, K. Yu, D. A. Reynolds, G. Evermann, D. Y. Kim and P. C. Woodland
Oct 2003
This report looks at the tasks of speaker diarisation, namely labelling who spoke when, and Speech-To-Text (STT) transcription, namely transcribing what was said, in some audio. Many experiments are performed within the context of the 2003 RT-03s spring Rich Transcription evaluation, investigating the performance and interactions between both diarisation and STT systems in both the broadcast news (BN) and conversational telephone speech (CTS) domains.
Questions addressed include can the same segmentation system be used for both diarisation and STT? Can knowledge from one system help to improve performance in the other? Can information from one help predict how well the other will do? Can combining different aspects of different systems help improve performance? and in short what are the interactions between diarisation and STT systems? Results and conclusions are drawn from many systematic experiments, involving diarisation and STT systems from several different sites.
For CTS it was found that diarisation systems can be used as input to STT and diarisation scores can accurately predict the resulting word error rates (WERs); using information from the recogniser output can improve diarisation score and thus word error rate further; STT systems may be slightly bias towards their own segmentations, but this effect is small; and the best automatic segmentations still trailed the best manually-derived segmentations by approximately 0.6% WER.
For BN it was found that there was no correlation between the diarisation score and the WER; automatic advert removal could be successfully achieved if contemporaneous (untranscribed) audio was available, but only helped performance if commercials were not excluded from the scoring; the CUED segmentation performed as well or better than the diarisation reference for STT, but still trailed the performance of using the STT reference segmentation by around 0.8% WER. Potential for improving both the STT and diarisation segmentations were also identified.
Finally, a hybrid diarisation system was built using CUED and MIT-LL components and shown to perform slightly better than the individual systems for the diarisation task on the development data.
If you have difficulty viewing files that end '.gz',
which are gzip compressed, then you may be able to find
tools to uncompress them at the gzip
web site.
If you have difficulty viewing files that are in PostScript, (ending
'.ps' or '.ps.gz'), then you may be able to
find tools to view them at
the gsview
web site.
We have attempted to provide automatically generated PDF copies of documents for which only PostScript versions have previously been available. These are clearly marked in the database - due to the nature of the automatic conversion process, they are likely to be badly aliased when viewed at default resolution on screen by acroread.