Levels of transcription in speech-oriented research

 

For the speech community, transcription of a spoken corpus is linked to the notion of labelling. According to Barry & Fourcin (1992:2):

The `labeling' of a recorded utterance involves the temporal definition and naming of its parts with reference to the acoustic signal. These `parts' may be temporarily discrete or over-lapping, and may be defined in acoustic, physiological, phonetic or higher level linguistic terms.

It is clear from this definition that, apart from the orthographic representation, other levels are necessary in this particular domain.

Various levels of labelling have been defined and used in different projects. Two of these will be reviewed here, since they are the basis of the recommendations presented in this document.

Barry & Fourcin (1992) offer a comprehensive system consisting of five levels:

Physical level

in which labels are defined solely with reference to the physically defined events in an utterance (Barry & Fourcin, 1992:2);

Acoustic-phonetic level

based on labels representing well known phonetic categories;

Narrow-phonetic level

using phonetic transcription symbols such as the IPA;

Phonemic level

in which only distinctive sounds (phonemes) in a given language are represented; and

Broad phonetic level:

this constitutes an intermediate level in which only symbols for phonemes are used although these may transcribe non-phonemic continuous speech phenomena.

Moreover, prosodic labels should be added. As Barry & Fourcin (1992:11) point out prosodic labels can be defined as a separate tier at each level since different categories of prosodic events can be found at different levels.

Similar levels are found in other proposals, such as the levels of segmentation and labelling discussed by Tillmann & Pompino-Marschall (1993) and used in the German PHONDAT project. Again, five levels of representation are defined: orthographic, canonical word forms, actual word realizations, sound segments and sub-segmental acoustic-phonetic events. The conventions developed within the German VERBMOBIL project (Kohler et. al, 1994; Hess at al., 1995; more information on the project is found at URL http://www.dfki.uni-sb.de/verbmobil/overview-us.html and at URL http://www.ims.uni-stuttgart.de/projekte/verbmobil/index-en.html), and those proposed by Autesserre et al.(1989), together with the work carried out under the SAM project (chapter 5 of Fourcin et al. (Eds.), 1989) should also be mentioned in this context.

The EAGLES Hanbook on Spoken Language Systems (EAGLES Spoken Language Working Group, 1995) provides a careful discussion of transcription and labelling levels in the chapter devoted to corpus representation. The following types of transcriptions are described:

Orthographic transcription or transliteration

,in which the standard spelling of a given language is used to indicate the spoken words.

Phonemic/phonotypical transcription

,based on the phonemes of a given language.

Allophonic transcription

,in which different symbols are used for a single phoneme when this phoneme occurs in different contexts.

Phonetic transcription

representing the pronunciation of words of individual speakers.

Prosodic transcription

Labelling levels are also proposed by the EAGLES Spoken Language Working Group (1995) and are defined as follows:

Orthographic level

,in which the standard spelling of a given language is used to indicate the spoken words.

Citation-phonemic level

containing the output phoneme string derived from the orthographic form (by lexical access, by letter-to-sound rules, or both.

Broad-phonetic or phonotypical level

, which uses only symbols that have the status of phonemes, marking the output of connected speech processes that insert or delete phonemes, or transform one phoneme into another.

Narrow-phonetic level

, which attempts to represent what the speaker actually said at the time of recording.

Acoustic-phonetic level

, which distinguishes every portion of speech that is recognisably a separate segment of the acoustic waveform or spectrogram.

Physical level

representing acoustic parameters or articulatory data.

Non-linguistic phenomena

including speaker noises, extraneous noises and paralinguistic information.