Tag
ELARSpeech introduces a 30,202-hour speech dataset covering 675 underrepresented languages curated from the Endangered Languages Archive, along with trained ASR/TTS models and a full data pipeline to advance long-tail multilingual speech technologies.
The paper identifies 'phonological interference' as a systematic failure mode in multilingual phoneme-level speech models, where models impose the phonology of a single assumed language and override phoneme decisions on code-switched or unseen-language input. It proposes windowed language estimation (WLE), an inference-time fix that reduces 34%–69% of interference without harming monolingual performance.
The Bairong system presents a dynamic question-aware evidence routing approach for multilingual conversational spoken question answering, achieving improved performance by effectively integrating transcript and audio evidence in the MLC-SLM 2026 Challenge.
This paper presents the SEAR system for multilingual speech MCQ, achieving 90.92% accuracy on the MLC-SLM Challenge through segment-evidence-aware routing, supervised fine-tuning, and reinforcement learning.