Word joins in real-life speech: a large corpus-based study

Lead Research Organisation: University of Oxford
Department Name: Linguistics Philology and Phonetics

Abstract

Abstracts are not currently available in GtR for all funded research. This is normally because the abstract was not required at the time of proposal submission, but may be because it included sensitive information such as personal details.
 
Description By developing and analysing a large colllection (1,477 hours) of recordings of ordinary spoken English, the Audio British National Corpus, we discovered that:
(1) word-final nasal consonants made with the lips or the back of the tongue sometimes change their articulation to match that of the beginning of the following word-initial consonant. This is contrary to findings from earlier, smaller studies, and overturns the predictions of previous
phonological theory. Such an unpredicted phenomenon required a large-scale study in order to detect it, on account of its relative rarity.
(2) Word-final [t] and [d] in consonant clusters (such as [ft] or [ld]) are highly variable, and may often appear to be "deleted" (e.g. ol' man). In previous work this has been
attributed to the class of the word (regular past tense forms vs. irregular pasts vs. uninflected words) - but we found no such conditioning. Instead, we found that most of the variation is due to differences in the frequency of use of the word. This relationship seems to be mediated by word duration: higher frequency words are pronounced
more quickly and their endings are pronounced less "carefully", whereas lower frequency words are pronounced more slowly and carefully. This finding, too, overturns prior views about how final "[t]/[d] deletion" works.
(3) For both nasal assimilation and "[t]/[d] deletion", novel probabilistic models of the respective patterns of variation are required.
(4) The large amount of data we examined, and its natural variability, necessitated creative approaches to acoustic measurement that are automatic and robust.
(5) Quite subtle variations in pronunciation can be studied in colloquial, spontaneous speech using purely acoustic measurements when a sufficiently large sample (roughly 100 or more instances of each item of interest, 50 per gender, in the same context) is employed.
(6) In order to find relevant examples in such a large audio corpus, it is necessary to use automatic methods to align the transcription - the start and end times of each vowel, consonant and word - to the audio, and to link data about the speaker (such as their sex and age) to the time-aligned transcriptions. We used forced alignment to obtain the tens of millions of time-alignments, and evaluated the accuracy of this method: 60% of automatically-assigned word boundaries were within 50 ms of manually-corrected controls, and 80% were within 100 ms.
(7) In a complementary study of "g-dropping" (i.e. variation in word-final -ing or -in', as in droppin'), our US co-investigator Jiahong Yuan developed a technique based upon forced alignment (adaptive discriminative training) to find examples of -ing or -in' pronunciations and to measure the acoustic differences between them.
(8) Given the highly asymmetric distribution of words in spontaneous speech, our corpus of 6.9 million words was barely sufficient for some of our research questions; in
future, speech corpora 10 or 100 times larger will be needed, thanks in part to the diversity of English accents in the British Isles.
Exploitation Route The audio data and transcriptions we have published have already seen (quite rapid) actual
uptake by teachers of English as a foreign language, around the world. During the lifetime of
the project, over 1200 distinct users have registered to access the data. The economic and
societal impacts of this is that UK English, in a wider range of dialects and varieties, is being
used in teaching English. This also helps to advance the BNC "brand", and may help to
promote sales of other resources based upon the BNC, such as English language teaching
dictionaries published by Longman, Oxford University Press, and Chambers.
The OCR school examinations board uses extracts from our BNC conversation
transcriptions in public examinations such as AS and A-Level English Language, and sells
copies of those exam papers.
Sectors Creative Economy,Digital/Communication/Information Technologies (including Software),Education,Healthcare,Culture, Heritage, Museums and Collections
 
Description The audio data and transcriptions we have published have already seen (quite rapid) actual uptake by teachers of English as a foreign language, around the world. During the lifetime of the project, over 1200 distinct users have registered to access the data. The economic and societal impacts of this is that UK English, in a wider range of dialects and varieties, is being used in teaching English. This also helps to advance the BNC "brand", and may help to promote sales of other resources based upon the BNC, such as English language teaching dictionaries published by Longman, Oxford University Press, and Chambers. Some of the user registrations are from commercial researchers in the field of speech technology, interested in using the recordings we have published for training and testing their algorithms. The OCR school examinations board uses extracts from our BNC conversation transcriptions in public examinations such as AS and A-Level English Language, and sells copies of those exam papers.
First Year Of Impact 2011
Sector Digital/Communication/Information Technologies (including Software),Education
Impact Types Societal,Economic
 
Title Audio BNC: the audio edition of the Spoken British National Corpus 
Description Audio recordings from the spoken part of the British National Corpus, digitized from the analogue audio cassette tapes deposited at the British Library Sound Archive, together with associated transcription and annotation files, published as open data on the public internet via http://www.phon.ox.ac.uk/AudioBNC 
Type Of Material Database/Collection of data 
Year Produced 2012 
Provided To Others? Yes  
Impact The audio data and transcriptions we have published have already seen (quite rapid) actual uptake by teachers of English as a foreign language, around the world. During the lifetime of the project, over 1200 distinct users have registered to access the data. The economic and societal impacts of this is that UK English, in a wider range of dialects and varieties, is being used in teaching English. This also helps to advance the BNC "brand", and may help to promote sales of other resources based upon the BNC, such as English language teaching dictionaries published by Longman, Oxford University Press, and Chambers. The OCR school examinations board uses extracts from our BNC conversation transcriptions in public examinations such as AS and A-Level English Language, and sells copies of those exam papers. 
URL http://www.phon.ox.ac.uk/AudioBNC
 
Description Unpredicted assimilations in UK English: Initial results from the Audio BNC 
Form Of Engagement Activity A talk or presentation
Part Of Official Scheme? No
Geographic Reach Local
Primary Audience Postgraduate students
Results and Impact Invited lecture

An early presentation of an output of our research to a scientists outside our own group
Year(s) Of Engagement Activity 2013
URL http://www.ppls.ed.ac.uk/events/the-p-workshop-21