Fisher English Training Speech Part 1 Transcripts represents the first half of a collection of conversational telephone speech (CTS) that was created at LDC in 2003. It contains time-aligned transcript data for 5,850 complete conversations, each lasting up to 10 minutes. In addition to the transcriptions, which are found under the trans directory, there is a complete set of tables describing the speakers, the properties of the telephone calls, and the set of topics that were used to initiate the conversations. The corresponding speech files are contained in Fisher English Training Speech Part 1 Speech (LDC2004S13).
The Fisher telephone conversation collection protocol was created at LDC to address a critical need of developers trying to build robust automatic speech recognition (ASR) systems. Previous collection protocols, such as CALLFRIEND and Switchboard-II and the resulting corpora, have been adapted for ASR research but were in fact developed for language and speaker identification respectively. Although the CALLHOME protocol and corpora were developed to support ASR technology, they feature small numbers of speakers making telephone calls of relatively long duration with narrow vocabulary across the collection. CALLHOME conversations are challengingly natural and intimate. Under the Fisher protocol, a large number of participants each calls an other participant, whom they typically do not know, for a short short period of time to discuss the assigned topics. This maximizes inter-speaker variation and vocabulary breath while also increasing formality.
Previous protocols such as CALLHOME, CALLFRIEND and Switchboard relied upon participant activity to drive the collection. Fisher is unique in being platform driven rather than participant driven. Participants who wish to initiate a call may do so, however, the collection platform initiates the majority of calls. Participants need only answer their phones at the times they specified when registering for the study.
To encourage a broad range of vocabulary, Fisher participants are asked to speak about an assigned topic chosen from a randomly generated list that changes every 24 hours. All participants that day will be assigned subjects from that list. Some topics are inherited or refined from previous Switchboard studies while others were developed specifically for the Fisher protocol.
Data
Overall, about 12% of the conversations were transcribed at LDC, and the rest were transcribed by BBN and WordWave using a significantly different approach to the task. A central goal in both sets was to maximize the speed and economy of the transcription process. This in turn involved certain aspects of mark-up detail and quality control that may have been common in previous, smaller corpora.
The LDC transcripts were based on automatic segmentation of the audio data, to identify the utterance end-points on both channels of each conversation. Given these time stamps, manual transcription was simply a matter of typing in the words for each segment and doing a rudimentary spell-check. No attempt was made to modify the segmentation boundaries manually, or to locate utterances that the segmenter might have missed. Portions of speech where the transcriber could not be sure exactly what was said were marked with double parentheses – (( … )) – and the transcriber could hazard a guess as to what was said, or leave the region between parentheses blank. The LDC transcription process yields one plain-text transcript file per conversation, in which the first two lines show the call-ID and the fact that the transcript was developed at LDC. The remainder of the file contains one utterance per line (with blank lines separating the utterances), with the start-time, end-time, speaker/channel-ID and utterance text.
Data collection and transcription were sponsored by DARPA and the U.S. Department of Defense, as part of the EARS project for research and development in automatic speech recognition.
v. 5.9
