ROMPAR: Morphological Completion and Demographic Unlearning for Romanian-Accented Speech Recognition

The paper introduces the ROManian PARliamentary Speech Corpus (ROMPAR), a 17.80-hour dataset designed for automated transcription of Romanian and Moldavian parliamentary speech, featuring double-annotated ground truth and labels for reconstructed word fragments. It proposes a multi-task adversarial training framework that incorporates demographic invariance across age, gender, and dialect, along with an exponential decay mechanism for adversarial coefficients and an LLM-guided decoding strategy for morphological completion, achieving a significant reduction in word error rate (WER) and an F1-score of 96.6% in morphological reconstruction. This work is crucial for practitioners aiming to enhance automatic speech recognition systems by addressing demographic biases and improving transcription accuracy in diverse linguistic contexts.