SNuC: The Sheffield Numbers Spoken Language Corpus

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1978–1984 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1978 SNuC: The Sheffield Numbers Spoken Language Corpus Emma Barker∗, Jon Barker∗, Robert Gaizauskas∗, Ning Ma∗, Monica Lestari Paramita† ∗Department of Computer Science, †Information School University of Sheffield {e.barker, j.p.barker, r.gaizauskas, n.ma, m.paramita}@sheffield.ac.uk Abstract We present SNuC, the first published corpus of spoken alphanumeric identifiers of the sort typically used as serial and part numbers in the manufacturing sector. The dataset contains recordings and transcriptions of over 50 native British English speakers, speaking over 13,000 multi-character alphanumeric sequences and totalling almost 20 hours of recorded speech. We describe requirements taken into account in the designing the corpus and the methodology used to construct it. We present summary statistics describing the corpus contents, as well as a preliminary investigation into errors in spoken alphanumeric identifiers. We validate the corpus by showing how it can be used to adapt a deep learning neural network based ASR system, resulting in improved recognition accuracy on the task of spoken alphanumeric identifier recognition. Finally, we discuss further potential uses for the corpus and for the tools developed to construct it. Keywords: spoken language corpora, corpus creation methodology, spoken alphanumeric identifier recognition 1. Introduction While large vocabulary automatic speech recogni- tion (ASR) has made tremendous progress in the last few years, there remain many challenges, for exam- ple high accuracy recognition of non-standard vocab- ularies used in specific applications and high accuracy recognition in noisy environments. One application that exhibits these characteristics and is of consider- able interest to a range of manufacturing and equip- ment servicing industries, is the recognition of spoken alphanumeric identifiers, such as serial and part num- bers. Speech recognition is of interest in these areas because workers who are assembling, disassembling or servicing complex machinery, such as jet engines, need their hands and eyes to get on with the job but also need to record or access information related to the components they are working on. Here, very high ac- curacy recognition of serial or part numbers is critical as these numbers are the key identifiers; furthermore the acoustic environments where these tasks are car- ried out, e.g. factories or service centres, are frequently noisy or highly reverberant, adding to the recognition challenge. In current practice workers are often forced to in- terrupt their work to shift to a keyboard and screen where information is entered or retrieved; or notes are scribbled on paper to be entered later. These are time- consuming and error-prone practices. One solution to identifying components is scanning. In some applica- tions this can work, but in many cases, either due to legacy approaches or because RFID tags are unwork- able (e.g., turbine blades in a jet engine), identifiers are stamped in metal and must be read from such. In such cases, spoken language data entry and retrieval is at- tractive, as it exploits the strengths of the human per- …

Content truncated to 500 words