q8 / int8 quantization via @huggingface/transformers + ONNX Runtime Web (WASM).
Each page has its own IndexedDB namespace, so recordings don't cross between them.
First recording on each page downloads the model (cached thereafter via Cache API).