DDSP's FilteredNoise synthesizer shapes white noise with a learned frequency-domain magnitude envelope

The FilteredNoise processor takes a frame-rate tensor of filter bank magnitudes from a neural network. It generates white noise (uniform random signal), converts the magnitude envelope to an FIR impulse response via IRFFT and a Hann window, then applies FFT convolution. This models the aperiodic (noise) component of sound, such as breath noise in a flute or frication in speech. Combined with the Harmonic synthesizer via Add, it creates the classic harmonic-plus-noise model used in DDSP autoencoders. The initial bias parameter (-5.0 by default) pushes the magnitudes toward silence at initialization, which stabilizes early training.

Examples

filtered_noise = ddsp.synths.FilteredNoise(n_samples=64000, window_size=257)
# magnitudes: [batch, n_frames, n_filter_banks]
audio = filtered_noise(magnitudes)

Assessment

Why is the harmonic+noise decomposition (Harmonic + FilteredNoise + Add) a useful model for musical instruments? What does the initial_bias=-5.0 achieve at the start of training?