Off-the-shelf video datasets for multimodal and speech AI — interviews, podcasts, lectures, films and real-world footage, many with transcripts. From Shaip.
AAA gameplay videos - approximately 1,000 hours of gameplay footage from AAA and AAA+ titles on PC and games consoles. Listed under gaming datasets, and suitable for game scene understanding, action and event detection, and game-playing agent or reinforcement learning model training.
AI interview videos - approximately 10,000 hours of video in which a single person answers questions put by an AI bot in text form, with participants of mixed ethnicity. Listed under interview and podcast video datasets, and suitable for interview and conversational AI, speech recognition and candidate assessment model training.
Arabic radio and TV interviews (AAUP) - 500 hours of video and audio obtained by permission, reviewed and verified by the vendor team, and transcribed using the vendor's ASR. Listed under interview and podcast video datasets, and suitable for Arabic speech recognition, speaker diarization and conversational AI model training.
Arabic radio and TV interviews - 500 hours of video and audio obtained by permission, reviewed and verified by the vendor team, and transcribed using the vendor's ASR. Listed under interview and podcast video datasets, and suitable for Arabic speech recognition, speaker diarization and conversational AI model training.
Baloch language video dataset - 750 videos in Balochi, a low-resource Asian language. Listed under language and multilingual media datasets, and suitable for automatic speech recognition (ASR), speech-to-text, machine translation and language identification model training.
A Chinese-language media library covering lecture videos with accompanying PPT slides and long-format video, alongside Chinese books, English books, journals and China public policy regulations, plus STEM audio and text (publications, novels and children's material, including Cantonese).
Documentary filmmaker with over 3,000 hours of original, high-quality footage shot across 8 countries (India, Morocco, Mexico, Mongolia, Argentina, Brazil, Saudi Arabia and the United States). The content includes real human interactions and daily life; interviews in local languages (e.g. Hindi, Arabic, Mongolian, Spanish and Portuguese); diverse faces, voices, accents and emotions; rural and urban environments; and horses, landscapes, rituals, street scenes, children and elders, and nature.
Interview and podcast videos featuring Asian participants - 1,000 videos. Listed under interview and podcast video datasets, and suitable for speech recognition, speaker diarization, audio-visual speech and conversational AI model training.
Licensed video library of 1.2 million minutes spanning news, weather, sports, live events, infrastructure, people and real-world environments. Broad scene and subject diversity with mixed demographics, suited to video understanding, action and event recognition, scene classification and multimodal foundation model training.
As a martial arts TV channel in Europe, we have over 1,000 original 60-minute fighting videos (boxing, kick-boxing, MMA and more). We are confident this is very good material and data for video machine learning.
Movies and short films in Italian, Arabic and English - 5,000 videos. Listed under film and documentary video datasets, and suitable for video understanding and captioning, subtitle alignment and dubbing, multilingual speech recognition and multimodal video model training.
Podcast videos in English and Arabic - 100 hours of video. Listed under interview and podcast video datasets, and suitable for speech recognition, speaker diarization, multilingual language model and audio-visual conversational AI training.
Radio and TV interviews with transcripts - 600 hours of video with accompanying text, obtained by permission and reviewed and verified by the vendor team. Listed under interview and podcast video datasets, and suitable for speech recognition, transcript alignment, speaker diarization and multimodal video-text model training.
Asian short films - 200 short film videos. Listed under film and documentary video datasets, and suitable for video understanding, scene and shot detection, subtitle alignment and multimodal video model training.
Short films and interview videos from the USA - approximately 1,000 hours of video. Listed under film and documentary video datasets, and suitable for video understanding, scene detection, speech recognition and multimodal AI model training.
The vendor has a video library of more than 40,000 curated short stock video clips, ranging from 5 to 60 seconds. The dataset includes footage of human activities and emergency scenarios (firefighting, rescue operations, police incidents, crime, natural disasters, floods, wildfires and extreme weather), as well as nature and wildlife scenes, including landscapes, forests, rivers, mountain ranges, rare wildlife, and aerial drone shots of remote and inaccessible terrain across different seasons and climates.
YouTube video dataset - approximately 80,000 hours of video. Listed under web and stock video datasets, and suitable for video understanding and captioning, action and scene recognition, speech recognition and large-scale multimodal model training.
