Whistle Detection Dataset

Data used to train and evaluate the whistle detector shown in 'Neural Network and Prior Knowledge Ensemble for Whistle Recognition'.
The data was recorded on a NAO V6 robot. In the folder Baseline_Data, we provide the labelled dataset we used for the publication as numpy files, while the folder TestBench contains the wave files we used to tune the parameters. The corresponding code is shared on GitHub (<a href="https://github.com/NaoDevils/AudioProcessing">https://github.com/NaoDevils/AudioProcessing</a>).
Due to legal reasons, not all data can be published.

Since the idea is to use the neural network as a pattern recognizer in the frequency domain, the audio samples must be prepared accordingly. This process involves buffering the audio stream to get windows containing 1024 samples. In the second step, we use a window function, the Hamming window, to avoid the Leakage-Effekt of a rectangular window. Following this, we compute the spectrogram of this window. Due to the symmetry of the Fourier transformation, we only need one half of the result plus the Nyquist frequency. After a Fourier transformation, the result contains complex numbers. However, we only use the real part of these complex numbers. For a more informative visualization, the volume/amplitude per frequency is shown in decibel (dB).