Free Shipping over $75
Secured Payment
OEM Quality
Customer Support
Abstract We introduce PREMEPORA-BARONS-01720-PHEVC-WEBDL-BENGALIX (hereafter PBB-PWB), a new multimodal dataset and benchmark designed to advance low-resource language understanding, compressed-video processing, and cross-domain web-derived text alignment. PBB-PWB comprises 17,220 annotated video clips encoded with perceptual HEVC variants (PHEVC), paired with crowd-sourced Bengali and code-switched (Bengali–English) transcripts, time-aligned subtitles, and web-derived metadata. We detail dataset curation, compression-aware preprocessing, and three tasks: (1) robust automatic speech recognition for low-bandwidth PHEVC video, (2) multimodal retrieval linking frames and web metadata, and (3) cross-lingual alignment for Bengali–English code-switching. We propose a baseline multimodal architecture combining compression-robust video encoders, wav2vec-style speech encoders fine-tuned on noisy PHEVC audio, and a cross-attention retrieval head. Extensive evaluations show PBB-PWB exposes performance gaps in current state-of-the-art models: relative WER increases of 28–45% under PHEVC artifacts, retrieval mAP drops of 22% for web-noise metadata, and alignment F1 reductions for code-switch segments. We release benchmarks, evaluation scripts, and baseline models to stimulate research in compression-robust multimodal systems for low-resource languages.
Unmatched, expert support for ANY keyless remote issue. We've got your back!
Need Help?Remote & Transponder Key Programming Instructions for over 5,000 vehicles!
Find Yours
Abstract We introduce PREMEPORA-BARONS-01720-PHEVC-WEBDL-BENGALIX (hereafter PBB-PWB), a new multimodal dataset and benchmark designed to advance low-resource language understanding, compressed-video processing, and cross-domain web-derived text alignment. PBB-PWB comprises 17,220 annotated video clips encoded with perceptual HEVC variants (PHEVC), paired with crowd-sourced Bengali and code-switched (Bengali–English) transcripts, time-aligned subtitles, and web-derived metadata. We detail dataset curation, compression-aware preprocessing, and three tasks: (1) robust automatic speech recognition for low-bandwidth PHEVC video, (2) multimodal retrieval linking frames and web metadata, and (3) cross-lingual alignment for Bengali–English code-switching. We propose a baseline multimodal architecture combining compression-robust video encoders, wav2vec-style speech encoders fine-tuned on noisy PHEVC audio, and a cross-attention retrieval head. Extensive evaluations show PBB-PWB exposes performance gaps in current state-of-the-art models: relative WER increases of 28–45% under PHEVC artifacts, retrieval mAP drops of 22% for web-noise metadata, and alignment F1 reductions for code-switch segments. We release benchmarks, evaluation scripts, and baseline models to stimulate research in compression-robust multimodal systems for low-resource languages.