Project Abstract

The United Kingdom is home to a rich tapestry of regional accents, of which are on the decline. With accent bias still corporeal in our society it must be a conscious effort to preserve these regional accents where we can. In speech technologies especially, those with regional accents (particularly Glaswegian and South Welsh) are forced to code-switch into a more ‘appropriate’ Standard South-English accents in order for their speech to be accurately recognized. The purpose of this research is to create a system which can automatically recognize regional accents, of which in stages passing this research, can be used to create more robust and accent-friendly speech technologies, such as Google Assistant, Alexa, etc. We collected audio data for nine regional accents, of which for each we extract phonological and prosodic features for use in both Logistic Regression Models and Random Forest Models. We also extract the most common words across all data clips and extract phonological features for each word bag. We found a relatively poor performance of 22% for the non-word-isolation approach, however in the word-isolation approach some words gained an accuracy of 50%. We note that either a larger dataset is needed to achieve a high (>=80%) accuracy, or the dataset should be composed of participants reading out identical linguistically-driven sentences to be able to fully assess the phonetic and phonological features of regional accents. With this, a GMM-model with i-Vectors may be a more viable approach for such a task.

Keywords: Machine Learning, Computational Linguistics,

 Conference Details

Session: Poster Session A at Poster Stand 86

Location: Sir Stanley Clarke Auditorium at Tuesday 7th 13:30 – 17:00

Markers: Julian Hough, Eike Neumann

Course: BSc Software Engineering, 3rd Year

Future Plans: I’m undecided