What it does
Classic "robotic" voice cross-synthesis based on convolution
Illustrates the technique of spectral cross-synthesis based on convolution of two audio signals. In this case, the Periodic Telegraph Noise is convolved with a human voice ("wavewarp.wav"), thus imposing the underlying spectral characteristics of the voice on to the random noise.
Try varying the parameters of all components to investigate the behaviour of the vocoder. Experiment with different audio signals (e.g. replace "wavewarp.wav" with "dance.wav" or any other audio file...or try convolving one audio file with another , or indeed, with itself...)
The Convolution block performs pure convolution of the two inputs. Although it uses the Fast Fourier Transform (FFT) for efficiency, appropriate zero-padding and overlapping is applied to eliminate "circular convolution", such that the output is exactly the direct convolution of the two inputs, as if the computations were carried out in the time domain. The only tell-tale sign that the FFT has been used is the inherent "latency". In other words, the output is delayed by the length of the input data buffer.