Five Minute Speech Segments

Project Description

This project will bring together social scientists, computer scientists and a mental health clinician to explore the application of automated speech analysis techniques to code expressed emotion in maternal speech samples. We will capitalise upon a unique existing collection of audio recordings of mothers speaking about their children from an existing UK cohort study, which were obtained on 2031 children when they were 5 years old. These recordings have previously been coded by highly trained human raters for mothers' negativity, hostility, emotional over-involvement and warmth towards their child. Machine-learning-based methods will be used to code these emotional attitudes in the mothers' speech samples and then tested to see how well the automated coding compare to the human ratings. To check for potential biases, we will explore whether the automated coding perform similarly well across speech samples from mothers in living in different parts of England and Wales and those from deprived versus wealthy families (to see if the automated analysis works well despite differences in accents/dialects and socio-economic status which can be problematic when using human raters). The children in this cohort were followed up for 13 years and their mental health was assessed at 12 and 18 years of age. We will therefore also investigate whether the automated ratings of mothers' emotional attitudes towards their children are associated with future mental health problems in the children, taking into account pre-existing emotional and behavioural problems in the children (assessed at age 5) and family history of mental illness. Additionally, we will conduct creative workshops with key stakeholders (young people, parents, healthcare and social-work practitioners and policy-makers, and educators) at the beginning and end of the project to determine the main ethical, social and practical challenges to using and sharing maternal speech data and developing and implementing the automated models in practice so that these can inform the approach we take within this project and inform the questions we explore in future related grant applications.

People and Partners

Working with <a href=https://www.kcl.ac.uk/people/johnny-downs>Prof Johnny Down</a>, and Prof Helen Fisher, Dr Nick Cummins and Dr André Bittar. Dr Bahman Mirheidari is the Sheffield Research Associate.

Funders

This project has received funding from the PSYCHIATRY RESEARCH TRUST.