Whisper In-Browser Transcriber

PROJECT_ABSTRACT

Whisper In-Browser Transcriber records your voice or takes a dropped audio file and transcribes it with OpenAI Whisper (tiny) entirely on the visitor's own device. Inference runs inside a Web Worker via ONNX Runtime Web, so the UI stays perfectly smooth even while decoding — and the audio never leaves the browser. The int8-quantized onnx-community/whisper-tiny.en model (~39 MB) is fetched once from the Hugging Face Hub and cached for instant reloads. A device/dtype fallback chain walks WebGPU → WASM q8 → WASM fp32 until a session boots. Ships as a pure static SPA.

KEY_CAPABILITIES

• ▶Off-thread inference: Whisper decodes inside a Web Worker, so the UI stays smooth while model downloads and transcription run in the background.
• ▶Live mic waveform capture plus per-segment timestamped transcript streaming copy and TXT export with the int8-quantized ~39 MB model cached after first visit.
• ▶Recorded or uploaded audio is decoded to 16 kHz mono PCM in the browser and never leaves the device.