LRS-VoxMM
收藏资源简介:
LRS-VoxMM是由韩国科学技术院构建的野外环境音视频语音识别基准数据集,基于多领域YouTube对话语料库VoxMM精选而成。该数据集包含29,146条样本,总时长25.3小时,涵盖12个不同领域的真实对话场景,具有人类标注的精确文本转录和丰富的元数据。通过严格的样本筛选流程,保留1-25秒的单人说话片段,并采用LRS系列标准格式进行音频重采样(16kHz)和视频处理(25fps/224×224分辨率)。其特色在于包含原始测试集和四种合成失真变体(噪声/混响/带宽限制组合),为研究复杂声学条件下的多模态语音识别提供了标准化评估平台,特别适用于验证视觉信息在音频退化场景中的补偿作用。
LRS-VoxMM is a real-world audio-visual speech recognition benchmark dataset developed by the Korea Advanced Institute of Science and Technology (KAIST), curated from the multi-domain YouTube dialogue corpus VoxMM. Comprising 29,146 samples with a total duration of 25.3 hours, this dataset covers real-world dialogue scenarios across 12 distinct domains, and is equipped with human-annotated accurate text transcriptions and rich metadata. Through a rigorous sample filtering pipeline, it retains single-speaker speech segments ranging from 1 to 25 seconds in length, and applies the standard LRS-series formats for audio resampling (16 kHz) and video processing (25 fps, 224×224 resolution). A key characteristic of LRS-VoxMM is that it includes the original test set and four synthetic distortion variants (combinations of noise, reverb, and bandwidth restriction), providing a standardized evaluation platform for researching multimodal speech recognition under complex acoustic conditions, and being particularly suitable for validating the compensatory effect of visual information in audio-degraded scenarios.
- 1LRS-VoxMM: A benchmark for in-the-wild audio-visual speech recognition韩国科学技术院 · 2026年



