Differentiable Short-Time Fourier Transform: A Time-Frequency Layer with Learnable Parameters

We present a differentiable version of the short-time Fourier transform (STFT), enabling gradient-based optimization of its parameters. This approach integrates with neural networks, allowing joint learning of both STFT and network parameters. Tests on simulated and real data demonstrate an improved time-frequency representation and enhanced performance on downstream tasks, illustrating the potential of our method as a standard for setting spectrogram parameters automatically.