audio¶

aeon performs a series of customizable transformations on the audio before provisioning the data to the model.

1. Probabilistically add noise to raw audio waveform (with the probability controlled by the add_noise_probability parameter). Noise is chosen from a random file in noise_index_file and added at a random offset and scaling (controlled by the noise_level parameter). Noise addition happens only in the time domain (as opposed to the spectral domain).
2. If time_scale_fraction has been set to a pair of values (tmin, tmax), then the spectral frames will be randomly squeezed or stretched in the time domain according to a random value chosen uniformly from the interval between tmin and tmax. For example:
Parameter Default Description
noise_index_file (string)   File of pathnames to noisy audio files, one per line.
noise_level (tuple(float, float)) (0.0, 0.5) How much noise to add (a value of 1 would be 0 dB SNR). Each clip applies its own value chosen randomly from with the given bounds.