Hwyel Stoakes,"Iterative Automatic Phonetic Segmentation of Australian Languages", 6 October 2015
When: 11am-12 noon (followed by lunch), Tues 6th October
Where: Room 78-217 (GP-South), University of Queensland
The increasing amount of speech data being gathered presents a number of data management challenges which are currently being addressed. Australian languages also have had more time dedicated to the collection of high quality speech data that are able to be used for various linguistic analyses. Phonetic investigations require very specific recording conditions in order to reach successful outcomes.
Australian languages are generally thought to be homogenous from a phonetic perspective although there has been relatively little critical analysis of this claim. One of the main impediments to further research has been the lack of cross-linguistic speech data, labelled in a consistent manner. Accurate phonetic analysis relies on time-aligned, phonetically labelled metadata to accurately mark-up speech.
Although there have been great advances in automatic labelling of speech data in recent years, this has relied on the gathering of large-scale, multi-speaker corpora with time-aligned labelling as a basis for developing acoustic models. The construction of such resources is not currently feasible for Australian Languages.
In this presentation I will present a method for automatically segmenting Australian language data requiring a speech signal and an associated orthographic transcription which uses currently available software. Using this method, speech data can be time aligned, segmented and phonetically labelled based on matching the possible phonetic input with acoustic models.
This is an iterative process that initially relies on acoustic models already developed for other languages. The labelling for the first pass is then corrected by a trained phonetician and a new model can be trained based on this adjusted annotation. This is then fed back into the model in an iterative process until a new model is created.
Initially each model remains independent but in future work the viability of an acoustic model that can be successfully applied to under-described Australian languages with no prior phonetic analysis will be investigated.
I will present data from four Australian languages Arrernte, Mawng, Djambarrpuyngu (Yolngu Matha) and Kunwinjku (Bininj Gun-wok).