- Aidacalc: labeled pictures of math equations
- crohme: handwritten mathematical equations
- ImageNet
- Flicker8k
- landlord handwritten name recognition
- Street View Text
- Text with bounding boxes from real images. Dictionary given so other words in the image can be parsed out
- IAM handwriting
- Motion of hand writing. Sample point is position, timestamp, pressure value of pen
- NEOCR: Natural Environment OCR Dataset
- KAIST Scene Text
- MSRA Text Detection with bounding boxes
- Stanford OCR clean subset of words and images in a csv file with the pixel values
- Chars74k Each character is its own image. Masks for character location also provided
- COCO Images with masks of objects to idenity
- EMNIST Handwritten letters (not just digits)
- EgoBody Motion of interacting people from head-mounted devices
- ami: audio recordings of meetings
- cmudict artic voice: recordings with sentence labels
- commonvoice: speach transcriptions
- Speech commands: individual words
- timit: audio transcription with labels at the sentence, word and phenome level
- CallHome talkbank: audio transcriptions of phone calls mid conversation. utterance level labels/timing for audio
- commonsense dialogues json of conversations (4-6 turns between speakers)
- ConvoKit
- persuasionforgood-corpus: Introduction hello, how are you?
- tennis-corpus: Reporter jumps right into conversation. Each question/answer is a conversation and they go in order
- iq2-corpus: Full on debate with speaker introductions
- friends-corpus: Jumps in mid conversation
- gap-corpus: Mid conversation talking about 15 most important items in a hypothetical plane crash. $=laughter
- casino-corpus: Introduction hello, how are you? Campsight neighbors negotiate for food water firewood etc
- Project Gutenberg plain text books Plain text
- OpenWebText replication of OpenAI's WebText
- HotpotQA chain of thought question and answer using search engine
- ReAct combines chain of thought with actions
- OpenAssistant ChatGPT replacement
- WordNet - how do words relate to each other in terms of hierarchy
- ConceptNet - how do words relate to each other in terms of usage (ex: A person can make coffee)
- NeMo ASR, NLP, TTS
- Flowtron text to melspectrogram fed into WaveGlow for waveform sythnesis
- One TTS alignment to rule them all for NeMo and Flowtron