ESRC Society Today

1. Research questions In this project we shall investigate variation in English intonation, insofar as it is represented in an existing corpus containing recordings of nine dialects of English. The acoustic properties of intonation are known to vary according to a large number of interacting factors. We shall construct a statistical computational model of intonation which takes account of variation due to dialect, speaking style, gender and individual speaker habits. (Prior models of intonation focus on a single dialect and style.) We shall therefore examine two principal research questions: 1) What is the contribution of dialect, speaking style, gender and individual speaker habits to intonational variation and how do they interact? (Note that rather than assuming that a group of speakers of a particular dialect is homogenous, we will also look at between-speaker variation.) 2) How can intonation variation be properly represented and modelled? The sort of model we will develop shall be a complete and explicit computational model. 2. Background Lack of quantitative modelling of intonational variation is widely recognised as a serious gap in speech technology. Commercially available speech synthesis systems produce intonation patterns only from Southern British English. A speech-impaired native of Northern Ireland, for example cannot obtain a synthesiser that produces the rising inflections typical of that variety of English. Furthermore, synthesisers do not produce different speaking styles. A similar situation arises in speech recognition. Commercial recognition systems, such as computer-based dictation systems and speech-driven web browsers, do not exploit intonation. Recent research, however, has begun to investigate how intonation can improve recognition. Without appropriate information on dialectal variation in intonation, wide use of speech recognition systems will be limited to Southern British English speakers. Equal access to speech technology for speakers of most varieties of English requires much better treatment of intonation. In turn, this depends on quantitative analyses of intonational variation combined with computational-linguistic models of intonation. There are two main reasons for the lack of work on intonational variation in linguistics and speech technology: 1. Many factors govern intonational variation. Dialect and style are obvious influences. Intonation is also affected by the words spoken, discourse structure, grammar, and the speakers gender, age, and linguistic habits. All these factors interact. As a result, quantitative investigations of intonational variation require computational techniques that linguists often do not command. 2. The computational techniques demand a large body of speech data in which factors such as dialect and style are represented. Such data have become available only very recently in the form of the IViE corpus, collected, acoustically analysed and transcribed on Grabes previous ESRC-funded research project (IViE = Intonational Variation in English, ESRC award R000237145). This corpus contains 36 hours of recordings from 108 speakers living in nine urban areas in the British Isles (Belfast, Bradford, Cambridge, Cardiff, Dublin, Leeds, Liverpool, London, Newcastle). The Bradford data is from speakers of the Punjabi community and the London data are from speakers of Caribbean descent, making the corpus somewhat representative of a wide range of communities in the British Isles. In each location, six male and six female speakers from the same age group performed an identical set of tasks, designed to elicit five speaking styles. These include read speech, goal-directed interactions, and free conversation. 3. Research proposed Two types of data are central to our project. One is the fundamental frequency trace, a sequence of numbers corresponding to the frequency of vocal cord vibration sampled at 10 ms intervals. This acoustic measure provides information about intonation, since it represents the pitch of the voice. Such traces have already been made for all recordings in IViE. The second type of data is machine-readable transcriptions of the intonation. These are sequences of letters that summarize the shape of an intonation pattern at particular points in time. These two types of data are illustrated in Figure 1 in the appended Research Proposal. Our research will take two steps. First, we will statistically characterise the variation in fundamental frequency and in the linguistic transcriptions due to dialect, style, gender and individual speaker habits. We will use such techniques as time-series analysis and normalisation across speakers. The second step is the development of a computational-linguistic model of the intonational variation quantified in the first step. We will use probabilistic finite state transducers (see the Research Proposal, section 1.2 for details) to construct a model that predicts intonation patterns produced by different spe

Award/Grant Outputs and Documents
Click on to download the document.	Number of Documents: 12

1     2      View All

• Back to top