High-Quality Training Datasets

Massive 3D Models & Scenes Dataset includes 270 million sets of 3D models and 18 million 3D scenes. 3D models cover conventional models, interactive models, and physics-enhanced models with various objects in indoor residential environments. 3D scenes cover indoor home decoration scenarios and commercial space environments. This dataset can be used for tasks like 3D asset generation, virtual environment simulation, AI model training, and industrial design applications.

The INTERSPEECH 2025 MLC-SLM Challenge Dataset, curated by Nexdata, is derived from fifteen proprietary conversational speech corpora. Distinguished by exceptional annotation accuracy and operational reliability, this dataset is engineered to address critical challenges in multilingual automatic speech recognition (ASR) and long-context comprehension. It meticulously replicates real-world complexities including spontaneous interruptions and speaker overlaps across 11 languages (1500 hours total duration), thereby providing robust training resources for developing world-ready ASR systems. All data collection and processing strictly comply with international privacy regulations including GDPR, CCPA and PIPL, with rigorous protocols ensuring participant anonymity and ethical data usage throughout the lifecycle.

Mandarin Full-Duplex Spontaneous Dialogue Speech Dataset, collected from dialogues based on given topics. Transcribed with text content, speaker's ID, gender, age and other attributes. Our dataset was collected from extensive and diversify speakers, geographicly speaking, enhancing model performance in real and complex tasks. Quality tested by various AI companies. We strictly adhere to data protection regulations and privacy standards, ensuring the maintenance of user privacy and legal rights throughout the data collection, storage, and usage processes, our datasets are all GDPR, CCPA, PIPL complied.

This dataset covers 21 languages, with 20,000 to 25,000 images per language. The data includes natural scenes, document photography scenes, and electronic scenes. The data diversity includes various data types, multiple shooting angles, and multiple languages. In terms of annotation, quadrilateral or polygonal at the row (column) level and content transcription at the row (column) level are adopted. This dataset can be use for multilingual optical character recognition (OCR) and text detection tasks.

This dataset contains 200,000 sets of images and bilingual captions (Chinese and English) featuring landmark buildings from over 20 countries, including the United States, United Kingdom, France, Germany, and Russia. Each set includes 1–10 images of a specific landmark, captured from different angles, distances, and time periods. The dataset covers approximately 80,000 domestic landmarks and 120,000 international ones. Types of landmarks include commercial buildings, ancient architecture, monuments, libraries, and scenic spots. Annotations include landmark country, city, location, category, and descriptive captions. This high-quality dataset is ideal for training models in landmark recognition, image classification, multilingual image captioning, and image-to-text retrieval.

The 600 Hours Greek Real-World Speech Dataset includes both casual conversations and monologues, covers self-media, conversation, live, variety show and other generic domains, mirroring real-world interactions. Transcribed with text content, speaker's ID, and other attributes. Our dataset was collected from extensive and diversify speakers, geographicly speaking, enhancing model performance in real and complex tasks. Quality tested by various AI companies. We strictly adhere to data protection regulations and privacy standards, ensuring the maintenance of user privacy and legal rights throughout the data collection, storage, and usage processes, our datasets are all GDPR, CCPA, PIPL complied.

The 600 Hours Norwegian Real-World Speech Dataset includes both casual conversations and monologues, covering domains such as self-media, live shows, and other generic domains, mirroring real-world interactions. Transcribed with text content, speaker's ID, and other attributes. Our dataset was collected from extensive and diversify speakers, geographicly speaking, enhancing model performance in real and complex tasks. Quality tested by various AI companies. We strictly adhere to data protection regulations and privacy standards, ensuring the maintenance of user privacy and legal rights throughout the data collection, storage, and usage processes, our datasets are all GDPR, CCPA, PIPL complied.

This 3D high-fidelity synthetic dataset simulates real-world Driver Monitoring System (DMS) environments using photorealistic 3D scene modeling. It includes multi-modal sensor outputs such as camera images, videos, and point clouds, all generated through simulation. The dataset is richly annotated with object classification, detection, and segmentation labels, as well as human pose data (head, eye, arm, and leg position/orientation), camera parameters, and temporal metadata such as illumination and weather conditions. Ideal for training and evaluating models in autonomous driving, robotics, driver monitoring, computer vision, and synthetic perception tasks.

This dataset is collected from the Japanese OKWAVE Q&A platform and includes large-scale parsed and processed text data suitable for LLM training and Japanese natural language understanding. It contains structured fields such as questions, answers, categories, timestamps, user metadata, and supplementary explanations. As of April 2025, the dataset includes 8.4 million questions with 2.3 billion words, 27 million answers totaling 7.6 billion words, 15.5 million thank-you messages (1.7 billion words), and 2.1 million supplementary replies (360 million words). Continuously updated and rich in user-generated content, this dataset is ideal for building Japanese conversational AI, ChatGPT fine-tuning, question answering systems, text summarization, and semantic parsing models. All data complies with relevant data usage and privacy regulations.

This dataset includes 500 hours of scripted Tamil monologue speech collected using smartphones. Each sample is transcribed with text content and metadata such as speaker ID, gender, and age. The dataset features diverse speakers from various regions, making it highly representative of real-world Tamil language use and suitable for automatic speech recognition (ASR), text-to-speech (TTS), voice activity detection (VAD), and natural language processing (NLP) tasks. Validated by leading AI companies, the dataset is designed to enhance model robustness in multilingual environments and low-resource languages. All data was collected in full compliance with global privacy regulations including GDPR, CCPA, and PIPL, ensuring ethical sourcing and responsible AI development.

Lao Scripted Monologue Smartphone speech dataset, collected from monologue based on given scripts. Transcribed with text content and other attributes. Our dataset was collected from extensive and diversify speakers geographicly speaking, enhancing model performance in real and complex tasks. Quality tested by various AI companies. We strictly adhere to data protection regulations and privacy standards, ensuring the maintenance of user privacy and legal rights throughout the data collection, storage, and usage processes, our datasets are all GDPR, CCPA, PIPL complied.

Chinese Multi-emotional Modal particle and Natural Conversation Speech Synthesis Corpus, is recorded by multiple native Chinese voice actors. It not only includes sentences rich in modal particles that align with daily expression habits, but also encompasses free conversation data on given topics. In each conversation, the audio of each speaker is independently stored in their respective tracks. Professional phoneticians have annotated information such as text content, meeting the precise requirements for speech synthesis research and development to a full extent.