Project 2: Automatic methods to identify FPD; Corpus-based probabilistic work
Project 2 investigates the distribution of particular types of fine phonetic detail (FPD) in speech. The project will focus on the automatic detection of FPD by means of existing and new statistical and statistical learning techniques. It will support and be supported by Project 1. The main aim is to develop tools for the (relatively) rapid automatic identification of FPD, which can be used by phoneticians and engineers alike.
Manual techniques for the identification of FPD are painstakingly slow, and no techniques for automating its discovery are in common use. However, the development of tools to help find FPD is of critical importance for both linguistics and engineering. For linguistic researchers, automation will vastly speed up the process of determining which aspects of FPD are universal and which are language-specific. For engineers, automating the discovery of FPD may lead to the development of better formalisms for both speech recognition and synthesis. In essence, the search for FPD will force engineers to contemplate trainable architectures which can accommodate those aspects of speech communication which have traditionally been ignored or downplayed in the construction of recognizers, e.g. long-duration effects and systematic interactions between suprasegmental and segmental levels.
Methods In the initial phase of this project, a range of statistical learning approaches favoured in automatic speech recognition (e.g. HMMs, ANNs) will be employed to determine the capabilities and limitations of current approaches in the discovery of FPD. To evaluate the performance of existing methods, systems will be trained using material where salient FPD is already known to exist, and a comparison will be made between automated and manually-derived indicators of the fine phonetic distinction. Pilot studies at Cambridge and Sheffield using HMMs to estimate systematic acoustic differences between supposed homonyms such as ‘missed-mist’ have demonstrated that certain automatically-derived measures (such as differences in spectral energy distributions) can correlate well with those captured by manual work, but these studies also suggest that other differences (e.g. segmental durations) are less easy to estimate.
The results of this evaluation will help to focus effort on those elements where existing approaches perform poorly. In the second phase of the project, newer and more advanced statistical learning techniques, such as dynamic Bayesian networks and multistream architectures, will be employed on the same problems. In addition, statistical techniques used at Nijmegen, and ASR techniques at Trondheim, will trace consequences for speech production and comprehension of a word’s a priori lexical probability estimated from spoken corpora, its probability with respect to it competitors in the lexicon, and its probability with respect to its context. These same techniques will trace the presence of FPD in lexical exemplars.
Young researchers One ESR (Schuppler) is based at Nijmegen, and could visit Sheffield, Trondheim (or Naples if long units are studied) for further computational training; and Sheffield (again), Cambridge or York for FPD training. She is also assisted by the Nijmegen ER working on Project 11.
Links Project 2 will use and be informed by Project 1’s newly-collected speech data and FPD discoveries.
Any reliable methods for automated FPD discovery will feed back into Project 1 and be tested in Projects 3 and 4. Through the replacement of the optimisation criterion universally employed in ASR (namely, minimisation of segment error rates) with one based on a closer match to perceptually-salient FPD in listeners, this project will contribute to Theme IV on forging closer links between ASR and HSR. This project entails very close working relationships between linguists and engineers.
Working on this project: » Dr Mirjam Ernestus » Prof Torbjørn Svendsen » Barbara Schuppler » Dr Odette Scharenborg » Dr Jonas Beskow
