The 2023 Speech Datathon: an overview

The 2023 edition of the “Speech Datathon”, a workshop for speech data archiving, took place from Monday November 13 to Tuesday November 14, 2023 in the DataLab spaces of the Bibliothèque nationale de France.

Goals

This event followed on from a first edition held in November 2021 in Orléans (report available here). The aim of the Datathon is to provide a framework within which to work on the description of oral resources and their context of production, with a view to their digital publication. Topics covered included key aspects of data management plans, such as document classification, file naming conventions, formatting and other steps that lead to the archival deposit, which opens up the data to new uses. The central challenge is to initiate a process of continuous enrichment of corpora, as part of a mutually supportive approach to documentation and research.

Participants and organisers

Calls for participation are open to all researchers and doctoral students with a set of speech data that they wish to open up (for a variety of purposes, including its re-use for new research purposes). All applications were accepted. Following two last-minute cancellations, ten participants were there at the workshop. A range of career stages were represented – 2 PhD students, 3 tenured academics, 4 postdocs and an engineer. In terms of subject areas, linguistics was best represented, but there were also specialists of ethnology and anthropology. The corpora concerned nine languages: Spanish, and eight little-documented languages: Samba Leko (Adamawa group), Gbaya (Ubangian), Yulu (Sara-Bongo-Baguirmi group), Wushi (Bantu), Quechua, Malagasy, Na (Tibeto-Burman), and Aramaic (Chamito-Semitic).

The eight organisers-supervisors from Bibliothèque nationale de France and CNRS were all present along the four half-days. This ensured a high supervisory ratio (close to 1:1), which was conducive to rapid progress by participants in mastering the tools presented. In addition, the diversity of the professions represented within this well-staffed team ensured that each of the topics covered was dealt with by experienced practitioners.

Overview

The detailed programme has been online since June 2023. The present overview aims to highlight the main messages delivered at the event.

Marie Carlin introduced the “BnF DataLab”, which she coordinates. Listening to a library curator explaining how she supports researchers in using BnF’s digital collections allowed participants to get a feel for the research potential offered by properly catalogued and referenced documents. The presentation set the topic of archival data deposit in its most attractive light, encouraging participants to take the plunge. Entering into a deposit process should be self-evident, but has become marginal since the 1990s, with distressing consequences in terms of data loss.

Fabrice Menneteau and Audrey Viault then presented oral corpora management at Bibliothèque nationale de France. Their presentation emphasised the wealth of oral corpora preserved by BnF and the link between the complementary missions of preservation, description and transmission.

The Cocoon platform, which hosts the deposits, was then presented by its designer and manager, Michel Jacobson (Huma-Num).

Through these three presentations, the challenges of good cataloguing were clearly set out. The key notion of metadata was clarified for participants: providing good metadata means explaining the essential parameters of the surveys carried out, and thereby clarifying the collection process, which is key for the scientific enterprise in all its dimensions. Metadata is data about data, but this definition can give rise to the worst suspicions of a dizzying spiral into abstraction, losing touch with the data in the process. A more telling description is that metadata is a love letter to the future, in the words of an American colleague.

In addition to getting an in-depth overview of the workflow, participants were introduced to the prospect of continuing the process beyond the stage of electronic publication of a corpus. Close association between publications, data and tools, which lies at the heart of the Open Science project, was illustrated by a presentation of the concrete possibilities for depositors to enrich their annotations and metadata after depositing, and to reference the resources in their research.

The three introductory presentations, complemented by a round-table discussion, were immediately followed by hands-on activities. The studious atmosphere of the Bibliothèque nationale de France was well suited to a group of highly-motivated participants, and the workshops adopted a fast pace. In view of the progress made by the end of the first day, a change was made to the initial program: the morning workshop on the second day was preceded by a demo (by Michel Jacobson) of tools for batch deposits. A second change consisted in making room for ethical and legal topics. These had deliberately been left out of the programme in order to focus efficiently on technical aspects. But these issues are so important that it was useful to provide relevant information, while clarifying that there are no ready-made solutions for the extremely diverse socio-political and personal situations that are encountered in the field. The main topics mentioned were (i) the precautions to be taken in the context of sensitive data, (ii) the need for authorisation for distribution, and (iii) the implications of difference licenses.

The last half-day session opened with an optimistic message about the benefits of data publication for research. This short presentation by a researcher (Alexis Michaud, CNRS-LACITO) took as an example the Pangloss collection, providing a demo of the tools put in place in the service of corpora of less-studied languages (the world’s linguistic diversity and heritage).

Ample time was then devoted to an assessment, perspectives and a general open discussion, on which the “Perspectives” section below is based.

Perspectives

Finalising deposits:

The work initiated during the workshop is continuing through a dialogue between the participants and the people they met during the Datathon. The results of the 2021 Datathon are very encouraging: since then, many participants have reached the stage of publishing their corpus. The organising team is equally optimistic about the potential of the 2023 participants.

Futures editions:

Clearly, it makes sense to have new editions of the Speech Datathon in a similar format in the future.

An improvement would consist in asking participants to prepare their corpus metadata in advance, so that work on their corpus could go ahead rapidly from the outset of the workshop.

A second wish is to scale up the project, so as to involve all researchers, and in particular PhD students, for whom the adoption of Open Science practices is an essential career challenge. It seems important that future calls for participation should be distributed as widely as possible, and passed on by supervisors and research unit directors.

In closing, the team of organisers would like to reiterate their encouragement to a very pleasant group of participants. We are available for exchanges with organisations and bodies that may be interested in partnerships, such as doctoral schools and research groups.

Acknowledgments

The Datathon was organised by the research network “Linguistique Informatique, Formelle et de Terrain” (GDR LIFT, Institut des Sciences Humaines et Sociales du CNRS), the Collection de Corpus Oraux Numériques (Cocoon) platform and the Bibliothèque nationale de France. It is among the activities supported by the “Corpus, Langues et Interactions” (CORLI) consortium of the Huma-Num Research Infrastructure (IR* Huma-Num) and by the Institute for Language Diversity and Heritage (ILARA) of the École Pratique des Hautes Études. Many thanks to GDR LIFT for financial support, and to LACITO-CNRS for its contribution to the budget and for handling the accounting.

---

---