akhilvssg/magmar-2026-test-asr-embeddings

该数据集包含90个测试集视频的自动语音识别（ASR）转录和每帧视频嵌入，用于MAGMaR 2026 ACL Workshop共享任务（oracle track）。ASR转录来自两个模型：Whisper large-v3和Qwen3-Omni-30B-A3B-Thinking，提供原始语言和英语翻译的转录。视频嵌入来自Qwen3-VL-Embedding-8B，采样率为每秒1帧。数据集不包含原始媒体文件（.mp4和.wav），仅包含衍生的文本和特征数组。语言包括英语、汉语普通话、缅甸语、泰语和尼泊尔语，以及一些无声片段。

This dataset contains automatic speech recognition (ASR) transcripts and per-frame video embeddings for the 90 test-set videos of the MAGMaR 2026 shared task. Transcripts come from two ASR models (Whisper large-v3 and Qwen3-Omni-30B-A3B-Thinking), providing native-language and English translations. Embeddings come from Qwen3-VL-Embedding-8B, sampled at 1 frame per second. The raw .mp4 and .wav media files are NOT included; only derived text and feature arrays are. Languages include English, Mandarin Chinese, Burmese, Thai, and Nepali, plus a few silent clips.