Module dsp::stft

Short-Time Fourier Transform + Multi-Resolution STFT loss

stft_forward slides a windowed frame across the signal and FFTs each one (see fft.h — frameSize must be a power of two), producing a spectrogram: stft_num_frames(...) frames of frameSize Complex bins each. stft_inverse reconstructs the time-domain signal via windowed overlap-add, normalized by the accumulated squared analysis window at each output sample — the standard robust OLA method, exact whenever the window/hop combination satisfies the constant-overlap-add condition (e.g. Hann with 50% or 75% overlap) and well-behaved otherwise.

mr_stft_loss computes the "multi-resolution STFT loss" from the neural vocoder literature (Yamamoto et al., Parallel WaveGAN, and others): for each requested frame size, take the spectral convergence (relative Frobenius-norm magnitude error) plus the mean L1 log-magnitude error between two signals' spectrograms, then average across resolutions — a similarity metric that's sensitive to spectral (not just sample-wise time-domain) differences at multiple time/frequency trade-offs at once.