Question 1

What problem does the article address in TinyML keyword spotting?

Accepted Answer

State-of-the-art TinyML KWS designs require time-consuming data collection and server-side training for each target scenario, creating a bottleneck for rapid prototyping on new devices and environments.

Question 2

How does on-device customization work in the proposed method?

Accepted Answer

A feature extractor is trained offline to output embeddings, then on-device class prototypes are computed from a few recorded examples; at inference, keyword classification is performed by measuring distance between the current embedding and prototypes.

Question 3

What performance results are reported on the Google Speech Command dataset?

Accepted Answer

Using 10 examples, the ResNet15-based model achieves up to 80% accuracy at a 5% false acceptance rate, with results reported as lossless under 8-bit quantization.