This paper reports on ongoing work on developing and evaluating speech recognition models for the Welsh language using data from the Common Voice project and two popular open development kits – HuggingFace wav2vec2 and coqui STT. Activities for ensuring the growth and improvement of the Welsh Common Voice dataset are described. Two applications have been developed – a voice assistant and an online transcription service that allow users and organisations to use the new models in a practical and useful context, but which have also helped source additional test data for better evaluation of recognition accuracy and establishing the optimal selection and configurations of models. Test results suggest that in transcription good accuracy can be achieved for read speech, but further data and research is required for improving recognition results of freely spoken formal and informal speech. Meanwhile a limited domain language model provides excellent accuracy for a voice assistant. All code, data and models produced from this work are freely available.

This paper presents the design, collection and verification of a bilingual text-to-speech synthesis corpus for Welsh and English. The ever expanding voice collection currently contains almost 10 hours of recordings from a bilingual, phonetically balanced text corpus. The speakers consist of a professional voice actor and three amateur contributors, with male and female accents from north and south Wales. This corpus provides audio-text pairs for building and training high-quality bilingual Welsh-English neural based TTS systems. We describe the process by which we created a phonetically balanced prompt set and the challenges of attempting to collate such a dataset during the COVID-19 pandemic. Our initial findings in validating the corpus via the implementation of a state-of-the-art TTS models are presented. This corpus represents the first open-source Welsh language corpus large enough to capitalise on neural TTS architectures.

This paper reports on the development of a voice assistant mobile app for speakers of a lesser resourced language – Welsh. An assistant with a smaller set of effective but useful skills is both desirable and urgent for the wider Welsh speaking community. Descriptions of the app’s skills, architecture, design decisions and user interface is provided before elaborating on the most recent research and activities in open source speech technology for Welsh. The paper reports on the progress to date on crowdsourcing Welsh speech data in Mozilla Common Voice and of its suitability for training Mozilla’s DeepSpeech speech recognition for a voice assistant application according to conventional and transfer learning methods. We demonstrate that with smaller datasets of speech data, transfer learning and a domain specific language model, acceptable speech recognition is achievable that facilitates, as confirmed by beta users, a practical and useful voice assistant for Welsh speakers. We hope that this work informs and serves as a model to researchers and developers in other lesser-resourced linguistic communities and helps bring into being voice assistant apps for their languages.