keywords

This dataset consists of keywords extracted from the LibriSpeech recordings, designed to provide conversational keywords for keyword recognition tasks. It serves a similar purpose to the Speech Commands dataset.

The keywords in both the /train/ and /test/ match those found in the Speech Commands /train/ set. As a result, the /test/ set contains more keyword classes than the Speech Commands /test/ set. If a keyword is missing from either /train/ or /test/, this means that it was not found in the LibriSpeech data.

The _background_noise_ folder includes the same noise samples used in the Speech Commands dataset.

The keywords were extracted using the Montreal Force Aligner. In certain cases identified as incorrect keyword segmentations, manual correction was applied.