Vocapia is coordinating the EU EDF project ALADAN. The consortium is comprised of 4 SMEs from 4 EU counties, embarking on a new collaboration with complementary expertise: Crowdee (Germany), Lingea (Czech Republic), Ianus Consulting (Cyprus), and Vocapia (France). The project goals are to design and develop a disruptive framework of development for Artificial Intelligence (AI) based language solutions for defence applications. Four language technologies are being addressed: spoken language identification, speech recognition, spoken term search, and text and speech translation. The proposed novel framework uses open data for model training and only needs small amounts of application-specific data for validation. The ambitious project aims to substantially reduce the costs (time and financial) of developing language technologies for military applications, thereby drastically reducing the barriers for their wider uptake.

Vocapia is a partner of the EU EDF KIOIS project. KOIOS is an interdisciplinary project that develops Frugal Learning methods, i.e., AI techniques to be applied in specific military use cases. The project represents an ambitious investigation initiative that will produce new AI techniques beyond the current state of the Art in different research domains (few shot transfer learning, zero shot leaning, synthetic data for AI, semi or non-supervised learning, domain adaptation, ...).
Vocapia leads the KOIOS activities on applying frugal machine learning techniques to extract information from audio data for further analysis by intelligence analysts. In an ever changing geopolitical situation, the targeted languages, dialects, types and quality of recordings are constantly changing, which directly impacts the requirements for speech technologies with a need for quick development or adaptation. A priority is addressing the robustness of the speech technologies to changes in acoustic conditions and targeted domain(s). For example, to address target data frugality, the acoustic characteristics of the recordings are taken into account to simulate training data.

The PRESERVE project is developing a trustworthy, transparent, and easy-to-use Hybrid C-UAS C2 platform to support Police Authorities with the prevention, early detection and optimal management of operational response against current and emergent threats in drone technology, increasing the security of citizens in public spaces and contributing significantly to the novel EU Counter-UAS policy. By fusing heterogeneous data from an array of physical sensors and online sources, the project aims to tackle the so far unmanageable threat of weaponised consumer drone swarms through open-source knowledge by non-state actors. Vocapia is customizing their speech technologies to meet the user needs of the pilot use cases. In particular, Vocapia contributes to the creation of a tool for the discovery and continuous monitoring of open data from different external sources, leading activities on automatic speech and speaker recognition in multimedia content. The aim is to detect content of interest in audio and video data from the web, e.g., online training camps. They are adapting their existing AI-based language-dependent solutions to cover highly varied types of speech, accents and acoustic conditions. The structured text produced by the automatic speech recognition component enables further downstream processing to flag radicalized speech across large audio/video collections. The speaker recognition component provides robust intra- and inter-document speaker clustering, and capabilities to train speaker-specific models to identify target speakers.

Vocapia is coordinating the DGA TILADI project along with Airbus D&S. The goal of TILADI is to improve the performance and robustness of language technologies for military needs and under a number of very challenging conditions, including dialectal and accentuated speech, under-resourced languages, and very noisy communications. The targeted technologies are language identification, speech transcription and speech translation. For language identification, about 100 languages and 50 dialects are being modeled. An experimentatal platform has been developed integrating all the project results and is available for user trials.