Voice-AI in Newcastle

ADAPT-VA (Acoustic and Dialectal Adaltation to a Voice Assistant) is situated at the intersection of sociolinguistics and speech technology and is conducted in collaboration with Prof. Kevin Tang, Prof. Jochen Steffens, and Prof. Karen Corrigan. It aims to investigate how users adapt their speech when interacting with dialect-sensitive versus neutral voice assistants (VAs). Specifically, it compares interactions with two distinct types of VAs: a dialect-sensitive system, which recognises and speaks with the Geordie accent; and a system without dialect sensitivity, which represents a neutral approach to speech recognition and speaks SSBE. The main variables of interest are: 1) salient dialectal features that are retained in interactions with VAs; 2) communicative strategies participants employ depending on the system's perceived capabilities. 

Overall, the aim is to identify the sociolinguistic factors that influence how individuals adapt their speech when interacting with a VA. Moreover, I am interested in determining which features are more prominent in human-computer interaction, i.e., which features are more likely to be retained in their regional realisation, and which tend to get erased or standardised. These findings have important implications for understanding how speech technologies may contribute to processes of social and cultural erasure in AI-mediated communication, particularly with respect to the marginalisation of non-mainstream identities.

For the experimental setup, pre-recorded human speech (by a voice actor who is a native speaker of Newcastle English but can also speak SSBE) was used to simulate a system that understands and speaks the user's dialect. In order to control the flow of the interaction, the participants are offered to communicate with the system by means of two types of guided prompts that appear on the screen in front of them (see the images below). The first type is of interest from a sociolinguistic perspective as these prompts focus on eliciting local phonological and morphosyntactic features. The second type of prompts tests whether different situational contexts, such as levels of formality and valence, would have an additional effect on the realisation of linguistic features. For instance, the experiment includes a medical emergency scenario, a vacation planning scenario, a job interview scenario, and others. With this setup, it is possible to predict the participants' inquiries without compromising the naturalness of their speech, allowing us to systematically elicit the linguistic features of interest while maintaining experimental control over the interaction.

The goal is to quantify changes in linguistic features based on experimental condition (dialectal vs. mainstream VA) and situational context (professional vs. neutral, high vs. low valence situational scenarios). The correlation between the realisation of the linguistic features and sociolinguistic metadata, such as age, gender, educational background, and previous experience with and exposure to VAs will also be tested. Overall, the aim is to identify the sociolinguistic factors that influence how individuals adapt their speech when interacting with a VA. Moreover, I am interested in determining which features are more prominent in human-computer interaction, i.e., which features are more likely to be retained in their regional realisation, and which tend to get erased or standardised. These findings have important implications for understanding how speech technologies may contribute to processes of social and cultural erasure in AI-mediated communication, particularly with respect to the marginalisation of non-mainstream identities.

Key output

Serditova, D., Tang, K., Steffens, J. (2025) Automatic Speech Recognition Biases in Newcastle English: an Error Analysis. Proc. Interspeech 2025, 3204-3208, doi: 10.21437/Interspeech.2025-1973

Serditova, D., & Tang, K. (2026). A Sociolinguistic Analysis of Automatic Speech Recognition Bias in Newcastle English. arXiv preprint arXiv:2603.24549.