Recent approaches in text-to-speech (TTS) synthesis employ neural network strategies to vocode perceptually-informed spectrogram representations directly into listenable waveforms. Such vocoding procedures create a computational bottleneck in modern TTS pipelines. We propose an alternative approach which utilizes generative adversarial networks (GANs) to learn mappings from perceptually-informed spectrograms to simple magnitude spectrograms which can be heuristically vocoded. Through a user study, we show that our approach significantly outperforms naïve vocoding strategies while being hundreds of times faster than neural network vocoders used in state-of-the-art TTS systems. We also show that our method can be used to achieve state-of-the-art results in unsupervised synthesis of individual words of speech.

The phase vocoder is a well-known technique for dividing an audio signal into time-varying sinusoidal components and estimating their frequencies and amplitudes. The accuracy of the frequency estimates is studied here by predicting, and then measuring experimentally, the magnitude of errors due to two factors: (1) interference between different components, and (2) interference due to the presence of noise in the signal. The magnitude of the error depends an the relative amplitudes of the component in question and the disturbing signal, on the size and spacing of the analysis windows, on the window function used, and, in the case where the disturbance is due to another sinusoidal component, on the phase difference between the two. The implications of these results for choosing analysis parameters are discussed. The case of a one-sample spacing between analysis windows is treated in detail. Finally, we compare the phase vocoder with the maximum likelihood frequency estimator.

This demo shows the prototyping environment of the Ircam Signal Processing Workstation. The environment is oriented toward rapid prototyping of DSP and Musical applications.