Transcribing Speech: Errors in Corpora and Experimental Settings

Administrations, government organs, judiciary courts always faced the problem of defining limits in transcription practices. Nowadays corpus linguistics and computational linguistics have focused their attention on spoken corpora as indispensable tools for descriptive linguistics, as well as for applied purposes (in speech technologies, such as text-to-speech and speech recognition, in dialogue systems, in natural language processing and information retrieval, etc.). But transcription is not just a meta-linguistic practice serving linguistic analysis, it is, at the same time, a linguistic act itself, governed by its own strategies and tightly linked to other speech acts, and linguistic practices (such as note-taking, listening to spoken language for different purposes, writing following dictation, etc.). Recent literature has often been centred on transcription system design, on reviewing and comparing different transcription systems (Chafe, 1995; Connell and Kowal, 1999; Cook, 1995; Derville, 1997; Edwards and Lampert, 1993; Lapadat, 2000; Leech, Myers and Thomas, 1995; Pallaud, 2003; Romero, O’Connell and Kowal, 2002), and on errors and inconsistencies in linguistic annotation. Furthermore a large tradition in transcription is common in ethnographic studies (Powers, 2005; Vigouroux, 2007) and in conversation analysis (Ashmore and Reed, 2000). Transcription of speech is often driven by different transcribers’ understanding strategies, leading to specific error typologies (Chiari, 2006a; Chiari, 2006b; Lindsay and O’Connell, 1995; Pallaud, 2002; Pallaud, 2003). How does the transcriber contribute to the reconstruction or mis-reproduction of the spoken text? An analysis of common errors derived from experimentally induced transcriptions and from spoken reference corpora of the Italian language are compared and analyzed quantitatively and qualitatively in order to observe different patterns, relative frequencies, and motivations of occurrence.