Abstract
Algorithms in speech and audio applications are often evaluated under adverse conditions to evaluate their robustness against additive noise. This research describes a method to generate artificial but perceptually plausible acoustic disturbances, thereby providing a controlled and repeatable context for evaluating algorithms. This allows for control of such noise parameters as coloration, modulation, and amplitude distribution independently of each other, while also providing the means to define the amount of coherence among all the signal channels. Results of a listening test in a monaural setup show no significant difference in naturalness between synthesized and original signal. It is not always obvious how to create natural noise. For example, it was observed that white Gaussian noise is often an inappropriate noise. Frequency-dependent modulations on a short time scale appear to contribute to naturalness. Synthesizing vinyl/shellac, which has a particular type of impulse character, requires a unique approach to synthesis. Rain and applause synthesis proved to be challenging.