Meeting Recording Transcription

LDC supports the Rich Transcription meeting recording evaluations by producing training and evaluation transcripts. Transcription efforts in 2008 and 2009 will support automatic speech-to-text transcription and speaker segmentation in the meeting domain for RT09. Further information about the RT09 evaluation is available on NIST's RT09 site.

The links below provide additional information about LDC's transcription efforts for Meeting Recognition.

This method produces accurate time-aligned transcripts within a limited timeframe. The QTR method excludes manual segmentation, special markup and multiple quality checks in favor of a single focused transcription pass. LDC followed the QTR method for the transcription of more than ten hours of training data for the RT-07 project.

The careful transcription effort involves multiple passes over the data. Annotators first manually segment speaker turns and (for broadcast data) story boundaries. Annotators then create a verbatim transcript by listening to each segment in turn. A second pass verifies the accuracy of the segment boundaries and transcript itself, revisits difficult sections, and adds information like speaker identity, background noise conditions, plus special markup for mispronounced words, proper names, acronyms, partial words and other phenomena. Further scans over the data identify common errors, conduct spelling and syntax checks, and standardize the spelling of personal, organization and other names across the transcripts. This technique was used for transcription of the evaluation corpora for the Rich Transcription 2005, 2006, and 2007 Spring Meeting Recognition Evaluations.