LiteRT.js Whisper

Browser-based Whisper benchmark using LiteRT.js 2.5.3 and Transformers.js with onnx-community/whisper-base. Both engines run float32 English transcription on WebGPU. The primary action transcribes once with LiteRT.js. The comparison runs 20 sequential samples per engine and compares their median inference times. Model download and initialization are not timed.

pnpm install
pnpm dev

Open https://localhost:5176. The LiteRT model is downloaded from Hugging Face and stored in the browser Cache Storage. Transformers.js downloads and caches the ONNX model independently.