bozliu/e2e-keyword-spotting-demo

Public Keyword Spotting Demo

This Hugging Face Space exposes a free browser-microphone keyword spotting demo for public testing.

It is intentionally not the same operating mode as the local v3 desktop demo:

• —local flagship demo: accuracy-first, Apple MPS, external AST + SUPERB ensemble
• —this hosted demo: browser mic, hosted CPU, public lightweight checkpoint

What this demo does

• —lets users stream speech continuously from the browser microphone
• —runs CPU inference with the current public lightweight checkpoint
• —shows live listening / match state, confidence, and hard-word confusion hints
• —visualizes the keyword wheel used in the desktop demo

Important notes

• —This Space is the free public browser path, not the highest-accuracy local v3 path.
• —This hosted demo follows the local realtime flow more closely, but still runs inside a browser rather than the desktop sounddevice loop.
• —Cloud CPU latency may differ from local Apple Silicon latency.
• —Hard words such as left, on, and down use stricter guardrails to reduce confusion.

Model source

The app downloads public checkpoints from the GitHub release assets of the main repository instead of bundling weights into the Space.

Repository: bozliu/E2E-Keyword-Spotting