Time–Frequency Synthesis of Noisy Sounds With Narrow Spectral Components

Authors: Marelli D., Aramaki M., Kronland-Martinet R., Verron C.
Publication Date: November 2010
Journal: IEEE Transactions on Audio, Speech and Language Processing (vol. 18(8), pp. 1929-1940, 2010)

Tags: Environmental Sounds

Abstract

The inverse fast Fourier transform (FFT) method was proposed to alleviate the computational complexity of the additive sound synthesis method in real-time applications, and consists in synthesizing overlapping blocks of samples in the frequency domain. However, its application is limited by its inherent tradeoff between time and frequency resolution. In this paper, we propose an alternative to the inverse FFT method for synthesizing colored noise. The proposed approach uses subband signal processing to generate time–frequency noise with an autocorrelation function such that the noise obtained after converting it to time domain has the desired power spectral density. We show that the inverse FFT method can be interpreted as a particular case of the proposed method, and therefore, the latter offers some extra design flexibility. Exploiting this property, we present experimental results showing that the proposed method can offer a better tradeoff between time and frequency resolution, at the expense of some extra computations.

Colored noise simulating the propagation of noise into a dispersive 75cm tube length:

Original sound generated by a digital waveguide model and several resyntheses of the colored noise using subband method and IFFT method:

Original sound generated by a digital waveguide model

Resynthesis with subband method (M = 128 and C = 438)

Resynthesis with IFFT method (M = 128)

Resynthesis with IFFT method (M = 1024)

Resynthesis with IFFT method (M = 8192)

Note: original and resynthesized sounds were filtered by a lowpass filter with cutoff frequency 6kHz since the subband model was designed to approximate the target PSD up to 6kHz.

Blurry sound effect on a Glass impact sound using: