Efficient Stft Approximation Using A Quantization And Differencing Method

We introduce a novel class of algorithms for efficiently computing approximations to the short-time Fourier transform (STFT) of any given signal. These algorithms may be classified in accordance with the number of quantization levels they use to represent the values in each short-time frame of the input signal. The structure and properties of these algorithms are illustrated in this paper through a specific case which uses a 3-level quantization in each frame. We also present results obtained by applying an implementation of the 3-level algorithm to musical and speech signals. The resulting approximations conform to theoretical expectations of 9dB SNR in the STFT approximations. However, these approximations are found to require an order of magnitude less computation than FFT-based algorithms for the exact STFT. 1 Introduction Real-time calculation of the STFT [1] is important for applications such as spectrogram analysis of speech [2], automatic music transcription [3] and tactile h..