Content preservation, high-quality production and process automation are at the core of the current transformation of Public Service Media (PSM) from its traditional business to the modern digital era. Broadcasters’ archives play a central role in this change, being a valuable source of information to run the business. On the other hand, accessing and retrieving desired content (and metadata) may be a non-trivial task. This is where artificial intelligence comes into play, providing solutions that help users extract knowledge and organise data more efficiently and effectively.

This presentation highlights how emerging AI-based technologies can support the mission of PSMs in this transition by exploiting AI’s potential to facilitate archive organisation and valorisation. Sample tasks in this area are those aimed at AI-based metadata extraction, like Natural Language Processing, geographical landmark recognition or TV celebrity identification.

In this context, we describe a use case based on the main tasks usually performed by media professionals during their everyday job. To validate our vision, we collected opinions from industry experts, which let us rank AI functionalities according to users’ needs for their implementation.

Maurizio Montagnuolo is a Senior Research Engineer working at the R&D Department of the Italian public broadcaster RAI. He graduated in Telecommunications Engineering and holds a PhD in “Business and Management”. His interests are mostly addressed in the context of multimedia data mining and artificial intelligence, fields in which he counts several publications in international journals and conferences. He has been working on several EC-funded projects in the area of digital archiving, automated metadata extraction and cloud technologies.

AI technologies have been used at the RTS for automatic metadata extraction for several years. Currently, we are in the final stages of implementing speaker recognition. This presentation will show how documentalists have been involved and the role they play in the development and use of these new tools.

As a specialist in new information technologies and data management, Janique Sonderegger participates in the implementation of new AI tools at the RTS and in the business analysis around these developments. She also has a role in monitoring the evolution of the documentalist profession in the Data and Archives department of RTS.

The RTÉ archives house vast amounts of analogue audio/-visual material from 20th-century radio and television broadcasting, which are being digitised as part of an ongoing initiative.

These materials are associated with varying degrees of metadata, but in the absence of transcription, the raw data (i.e. audio) cannot be searched. While human transcription remains the gold standard in terms of accuracy, automatic speech recognition (ASR) systems offer a speed and a scale that can be used in tandem with transcribers to improve workflow efficiency.

ASR has been thoroughly researched and developed for major languages like English, but the Irish-language equivalent is still at an early stage. Although impressive results have been achieved for careful read speech, performance is markedly worse for conversational, spontaneous speech, the dominant style found in broadcast materials. This is primarily due to a lack of sufficiently large Irish speech corpora. However, recent advances in self-supervised learning have seen the inclusion of unlabelled data in training corpora i.e. audio materials without transcriptions, significantly reducing the amount of paired audio-transcription data required to achieve good performance. Thus, the application of research-stage Irish ASR systems to unlabelled speech corpora in, e.g., the RTÉ Archives presents an opportunity to accelerate both (i) the training of these systems and (ii) the conversion of these data to a format more accessible for cultural and academic purposes. This, in turn, can contribute to ongoing linguistic research, the results of which may themselves be fed back into speech technology development.

Connor McCabe is a Research Fellow with the ABAIR project. He is a phonetician and phonologist by training, having recently completed a PhD, his thesis for which examined prominence marking in the Munster varieties of Irish. In his current postdoctoral role, he is involved in managing ABAIR’s collection of speech data for linguistic research and technological development. Connor is interested in the research and development potential represented by media archives, having used 1928 archival data from the Doegen Records as key data for his PhD.

—

Liam Lonergan is a PhD student working with the ABAIR project in the Phonetics & Speech Laboratory of Trinity College, Dublin. He completed a B.A. (Mod.) in Computer Science, Linguistics and German at Trinity in 2019, after which he joined ABAIR as a research assistant. His work since then has centred around developing speech recognition technology for the Irish language. For his PhD in particular, he is investigating the role of dialect and language variety for speech recognition in the Irish context.