Modern Automatic Speech Recognition (ASR) systems are increasingly deployed in high-stakes settings, including clinical interviews, public services, and educational tools, where equitable performance across speaker groups is essential. While pre-trained speech models like Whisper achieve strong overall accuracy, they often exhibit inconsistent group-level performance that varies across domains. These disparities are not fixed properties of the model, but emerge from the interaction between model, data, and task—posing challenges for fairness interventions designed in-domain.We frame fairness in ASR as a generalisation problem. We fine-tune a Whisper model on the Fair-Speech corpus using four strategies: basic fine-tuning, demographic rebalancing, gender-swapped data augmentation, and a novel contrastive learning objective that encourages gender-invariant representations. We evaluate performance across multiple aspects of fairness and utility, both in-domain and on three out-of-domain test sets: LibriSpeech, EdAcc, and CognoSpeak.Our findings show that the method with the best in-domain fairness performed worst out-of-domain, illustrating that fairness gains do not always generalise. Demographic balancing generalises more consistently, while our contrastive method offers a practical alternative: it achieves stable, cross-domain fairness improvements without requiring changes to the training data distribution, and with minimal accuracy trade-offs.

This paper introduces a new British English speech database, named the homeService corpus, which has been gathered as part of the homeService project. This project aims to help users with speech and motor disabilities to operate their home appliances using voice commands. The audio recorded during such interactions consists of realistic data of speakers with severe dysarthria. The majority of the homeService corpus is recorded in real home environments where voice control is often the normal means by which users interact with their devices. The collection of the corpus is motivated by the shortage of realistic dysarthric speech corpora available to the scientific community. Along with the details on how the data is organised and how it can be accessed, a brief description of the framework used to make the recordings is provided. Finally, the performance of the homeService automatic recogniser for dysarthric speech trained with single-speaker data from the corpus is provided as an initial baseline. Access to the homeService corpus is provided through the dedicated web page at http://mini.dcs.shef.ac.uk/resources/homeservice-corpus/. This will also have the most updated description of the data. At the time of writing the collection process is still ongoing.