The engine, running.

A small story model writes a children's tale one token at a time while KittenTTS reads it aloud. Both models arrive as prepared .dlim images; the whole engine runs in your browser, the speech thread and the parallel kernels on Web Workers.

Type the words you wish for and a story model trained to take requests - our own 27M llama on the TinyStoriesInstruct corpus, published on Hugging Face as borisbat/dasllama-stories - writes a children's tale that uses them while Kyutai Pocket TTS reads it aloud in one baked voice, no phoneme front end in the loop. The same engine as the storyteller, a 32 MB story image and the speech model's own file.

Press record and talk for a few seconds. Kyutai Pocket TTS clones your voice from the take and reads the text in the box - a poem to begin with, or whatever you type - in it; Silero VAD ends the take when you go quiet, and you can record again at any time. The recording stays in the tab: nothing is uploaded. The same engine, the speech model's 75 MB file with its codec encoder, and a 1 MB voice-activity model.