Jodie Lechapelain

Over the last decade we have witnessed the emergence of large language models which show impressive performance for a range of language technologies. These models are trained on gigantic amounts of data, using seemingly infinite computing resources, which is often not feasible for many LT actors. However, balancing performance of speech processing technologies across varied data types and languages remains quite challenging and many applications do not reap the benefits of entering this new era of AI. Model performance still degrades significantly on types of data unseen in the training phase. This is particularly problematic for military applications where data is scarce, highly confidential, and extremely costly to annotate, whilst needing to process varied types of data such as rare dialects, highly degraded radio signals or communications based on specific phraseologies. Vocapia's quest for frugal AI speech processing aims to implement highly accurate and portable low footprint solutions using reasonable resources, both in the development phase and at runtime. This presentation will address strategies to tackle these frugality challenges on speech processing use cases, and the integration of these advancements in Vocapia's software products. Work presented is partially conducted in context of projects funded by the European Defence Fund with the aim to bridge the gap in performance between civil and military applications.