Each model opens in a live pipeline: the stages light up as your input moves through, on the device it actually runs on.
On-device ASR for the ESP32-S3: log-mel front-end, Conformer encoder, RNN-T decoder in a 6 MB footprint.
OpenAI whisper-tiny running locally through Transformers.js and WebAssembly. Audio never leaves the tab.