PKU-KWS

We collect a new dataset called PKU-KWS. The PKU-KWS dataset is collected in a relatively quiet acoustic environment with a camera recording at the speed of 25 frames per second. The video resolution is 1080 × 1920, and the audio is synchronously recorded at the sampling frequency of 16000Hz, with 16 bits for each sampling. We define five wake words commonly used in supermarket shopping. Different from other datasets, the PKU-KWS dataset contains 500 single-speaker conversations, 300 double-speaker conversations, and 200 three-speaker conversations. The duration of each conversation is not equal and the multi-speaker conversations may or may not contain a wake word.