Three modalities, still no server
The other playgrounds need WebGPU. These don't. transformers.js (ONNX Runtime Web) runs real speech, vision, and embedding models on your CPU via WebAssembly — so speech-to-text, zero-shot image understanding, and semantic search all work in any modern browser, with nothing leaving the tab.
Whisper transcribes audio, locally.
Play the built-in sample, upload an audio file, or record from your mic. It's decoded to 16 kHz mono in the browser and transcribed by Xenova/whisper-tiny.en. The audio never leaves the tab.
CLIP scores how well an image matches each label.
Zero-shot image classification: no training for these labels — Xenova/clip-vit-base-patch32 embeds the image and each candidate phrase into the same space and ranks them by similarity. Pick a sample or upload your own, edit the labels, and run.
Candidate labels
One phrase per line. These are the classes CLIP will rank the image against.
Search by meaning, not keywords.
Every line of the corpus is embedded once into a 384-dim vector with Xenova/all-MiniLM-L6-v2; your query is embedded the same way and ranked by cosine similarity. No LLM, no keywords — this is the retrieval half of the RAG playground, on its own.
The corpus → vectors
Each line is embedded into a 384-dim unit vector. The strip is a fingerprint of the first 48 dimensions; after a search, each line also shows its cosine similarity to your query.
Ask the corpus
Try “how do machines understand images?” or “a fast way to reduce model size”. Every line is scored above; the best match is highlighted.