Catherine Sotillo

The DCIEM map task corpus: spontaneous dialogue under sleep deprivation and drug treatment

Abstract This paper describes a resource designed for the general study of spontaneous speech under the stress of sleep deprivation. It is a corpus of 216 unscripted task-oriented dialogues produced by normal adults in the course of a major sleep deprivation study. The study itself examined continuous task performance through baseline, sleepless and recovery periods by groups treated with placebo or one of two drugs (Modafinil, d-amphetamine) reputed to counter the effects of sleep deprivation. The dialogues were all produced while carrying out the route communication task used in the HCRC Map Task Corpus. Pairs of talkers collaborated to reproduce on one partners schematic map a route preprinted on the others. Controlled differences between the maps and use of labelled imaginary locations limit genre, vocabulary and effects of real-world knowledge. The designs for the construction of maps and the allocation of subjects to maps make the corpus a controlled elicitation experiment. Each talker participated in 12 dialogues over the course of the study. Preliminary examinations of dialogue length and task performance measures indicate effects of drug treatment, sleep deprivation and number of conversational partners. The corpus is available to researchers interested in all levels of speech and dialogue analysis, in both normal and stressed conditions.

Limited visual control of the intelligibility of speech in face-to-face dialogue

Speakers are thought to articulate individual words in running speech less carefully whenever additional nonacoustic information can help listeners recognize what is said (Fowler & Housum, 1987; Lieberman, 1963). Comparing single words excerpted from spontaneous dialogues and control tokens of the same words read by the same speakers in lists, Experiment 1 yielded a significant but general effect of visual context: Tokens introducing 71 new entities in dialogues in which participants could see one another’s faces were more degraded (less intelligible to 54 naive listeners) than were tokens of the same words from dialogues with sight lines blocked. Loss of clarity was not keyed to momentto-moment visual behavior. Subjects with clear sight lines looked at each other too rarely to account for the observed effect. Experiment 2 revealed that tokens of 60 words uttered while subjects were looking at each other were significantly less degraded (in length and in intelligibility to 72 subjects) vis-à-vis controls than were spontaneous tokens of the same words produced when subjects were looking elsewhere. Intelligibility loss was mitigated only when listeners looked at speakers. Two separate visual effects are discussed, one of the global availability and the other of the local use of the interlocutor’s face.

The HCRC Map Task corpus: natural dialogue for speech recognition

The HCRC Map Task corpus has been collected and transcribed in Glasgow and Edinburgh, and recently published on CD-ROM. This effort was made possible by funding from the British Economic and Social Research Council.The corpus is composed of 128 two-person conversations in both high-quality digital audio and orthographic transcriptions, amounting to 18 hours and 150,000 words respectively.The experimental design is quite detailed and complex, allowing a number of different phonemic, syntactico-semantic and pragmatic contrasts to be explored in a controlled way.The corpus is a uniquely valuable resource for speech recognition research in particular, as we move from developing systems intended for controlled use by familiar users to systems intended for less constrained circumstances and naive or occasional users. Examples supporting this claim are given, including preliminary evidence of the phonetic consequences of second mention and the impact of different styles of referent negotiation on communicative efficacy.

Taking the Hit: Leaving Some Lexical Competition To Be Resolved Post-Lexically.

Natural variations in word pronunciation are not noise but information. Duration, prosodic prominence, vowel centralisation, and phonological reduction or assimilation can indicate whether a word stands alone or forms part of an utterance, whether it lies at the boundary of a major prosodic unit, is predictable in its context, or refers to a Given or a New entity. Though this variation is related to high-level factors, most discussions of lexical access seem to assume that lower level processes—acousticphonetic processing, phonological representation in the mental lexicon, and lexical effects on phonological representations of input—simply overcome variations in natural pronunciation, assuring that the correct word is accessed and ultimately selected, with no shortfall in the process that demands the participation of higher level information. Many of the papers in this volume deal with the architectural detail of this view. This paper summarises work on spontaneous unscripted speech, where variations most naturally occur, which shows why any such approach is counterproductive.