%Aigaion2 BibTeX export from Idiap Publications
%Sunday 22 December 2024 07:52:08 AM

@INPROCEEDINGS{Bhattacharjee_SID_2023,
         author = {Bhattacharjee, Mrinmoy and Motlicek, Petr and Iuliia, Nigmatulina and Helmke, Hartmut and Ohneiser, Oliver and Kleinert, Matthias and Ehr, heiko},
       keywords = {Customization of model, Integration of prior knowledge, model adaptation, Rare-word integration, speech recognition},
       projects = {Idiap},
          month = nov,
          title = {Customization of Automatic Speech Recognition Engines for Rare Word Detection Without Costly Model Re-Training},
      booktitle = {Proc. 13th SESAR Innovation Days},
           year = {2023},
       location = {Seville, Spain},
            url = {https://doi.org/10.61009/SID.2023.1.10},
            doi = {10.61009/SID.2023.1.10},
       abstract = {Thanks to Alexa, Siri or Google Assistant automatic speech recognition (ASR) has changed our daily life during the last decade. Prototypic applications in the air traffic management (ATM) domain are available. Recently pre-filling radar label entries by ASR support has reached the technology readiness level before industrialization (TRL6). However, seldom spoken and airspace related words relevant in the ATM context remain a challenge for sophisticated applications. Open-source ASR toolkits or large pre-trained models for experts – allowing to tailor ASR to new domains – can be exploited with a typical constraint on availability of certain amount of domain specific training data, i.e., typically transcribed speech for adapting acoustic and/or language models. In general, it is sufficient for a “universal” ASR engine to reliably recognize a few hundred words that form the vocabulary of the voice communications between air traffic controllers and pilots. However, for each airport some hundred dependent words that are seldom spoken need to be integrated. These challenging word entities comprise special airline designators and waypoint names like “dexon” or “burok”, which only appear in a specific region. When used, they are highly informative and thus require high recognition accuracies. Allowing plug and play customization with a minimum expert manipulation assumes that no additional training is required, i.e., fine-tuning the universal ASR. This paper presents an innovative approach to automatically integrate new specific word entities to the universal ASR system. The recognition rate of these region-specific word entities with respect to the universal ASR increases by a factor of 6.},
            pdf = {https://publications.idiap.ch/attachments/papers/2023/Bhattacharjee_SID_2023.pdf}
}