Demo dataset
This directory contains a small demo dataset with real recordings:
-
positive/— 22 recordings of the target phrase «Привет Аврора» (different speeds and volumes, including a few quiet/slurred takes). -
negative_hard/— 26 short non-target phrases, including adversarial ones that are close to the target («привет», «аврора», «это Аврора», «спасибо, Аврора», ...). -
negative_long/— 23 two-second fragments of arbitrary speech and room tone.
All recordings are made by the project maintainer and are published with their consent.
The dataset is intentionally tiny so that ../train_wakeword.py runs in
seconds. It demonstrates the training pipeline and the difficulty of
separating the target phrase from similar speech; a model trained on it
is not production quality.
Format of all clips: 16 kHz, mono, PCM signed 16-bit WAV.
To extend the dataset, add your own recordings in the same format to the
corresponding class directories — any *.wav file is picked up
automatically.