e2e-keyword-spotting-v3

E2E Keyword Spotting v3 Assets

This repository contains the public v3 assets for https://github.com/bozliu/E2E-Keyword-Spotting.

What is included

• best_kws12_v2_cpu_baseline.pt: our own detector checkpoint for the stable CPU baseline
• keyword_calibration_v2_cpu_baseline.json: keyword calibration for that checkpoint
• release_summary_v3.json: cleaned release summary used by the public README
• self_benchmark_v3.json: self-benchmark comparison table used by the public README

Important note about v3

The strongest v3 path in the public release is an accuracy-first composed runtime on Apple MPS. It combines:

• our detector gate/runtime packaging from this repository
• upstream Hugging Face models superb/wav2vec2-base-superb-ks and MIT/ast-finetuned-speech-commands-v2
• our local mapping, calibration, and realtime validation logic

This means the best public v3 result is not represented by one newly-trained monolithic checkpoint in this repo. Instead, this HF repo publishes:

• our own reusable baseline checkpoint
• the cleaned release metrics
• the metadata needed to understand the v3 public result

Result snapshot

On the current local KWS12 protocol, the public release summary reports:

• offline worst-class precision: 95.44%
• offline worst-class recall: 95.38%
• unknown-to-target rate: 0.43%

See release_summary_v3.json for details and comparability notes.