OptionaldeviceRuntime; default WebAssembly (as fast as WebGPU for these models, and lighter).
OptionalkeepKeep the model loaded between jobs (faster next time, holds ~0.3–1 GB). Default false: the worker closes after each job and its memory goes back to the page; the downloaded files stay cached, so the next job only waits a few seconds.
OptionalmodelModel size; default picks for the device (see chooseModel).
OptionalmodelWhere model files come from — self-host them by mirroring the model repos: files are fetched
from <modelHost>/onnx-community/whisper-<size>_timestamped/resolve/main/<file>.
Default https://huggingface.co.
OptionalruntimeWhere the speech runtime's .wasm files are served (default: next to its JS module).
Options of createOnDeviceTranscriber.