Nouveau Recherche PDF, HTML, DOCX et bien d'autres formats.

Prépublication · 2026

TID System for Speech-COSER Challenge 2026: ASR Track

Fernando López, Jordi Luque — Monde

Source : arXiv

Résumé

We describe the system submitted by Telefónica Innovación Digital to the Speech-COSER Challenge 2026, which features spontaneous speech from elderly speakers recorded in rural Spain. For ASR, we perform CTC-based pseudo-forced alignment scores audio-text match quality to guide a two-stage LoRA curriculum, followed by post-hoc magnitude pruning of the LoRA matrices to enhance generalization. An ensemble of seven systems combined via ROVER achieves 36.25\% unnormalized and 22.17\% normalized WER.

Actions

Citation