In this publication, we present a deep learning-based method to transform the f0 in speech and singing voice recordings. f0 transformation is performed by training an auto-encoder on the voice signal’s mel-spectrogram and conditioning the auto-encoder on the f0. Inspired by AutoVC/F0, we apply an information bottleneck to it to disentangle the f0 from its latent code. The resulting model successfully applies the desired f0 to the input mel-spectrograms and adapts the speaker identity when necessary, e.g., if the requested f0 falls out of the range of the source speaker/singer. Using the mean f0 error in the transformed mel-spectrograms, we define a disentanglement measure and perform a study over the required bottleneck size. The study reveals that to remove the f0 from the auto-encoder’s latent code, the bottleneck size should be smaller than four for singing and smaller than nine for speech. Through a perceptive test, we compare the audio quality of the proposed auto-encoder to f0 transformations obtained with a classical vocoder. The perceptive test confirms that the audio quality is better for the auto-encoder than for the classical vocoder. Finally, a visual analysis of the latent code for the two-dimensional case is carried out. We observe that the auto-encoder encodes phonemes as repeated discontinuous temporal gestures within the latent code.

No abstract

The use of the mel spectrogram as a signal parameterization for voice generation is quite recent and linked to the development of neural vocoders. These are deep neural networks that allow reconstructing high-quality speech from a given mel spectrogram. While initially developed for speech synthesis, now neural vocoders have also been studied in the context of voice attribute manipulation, opening new means for voice processing in audio production. However, to be able to apply neural vocoders in real-world applications, two problems need to be addressed: (1) To support use in professional audio workstations, the computational complexity should be small, (2) the vocoder needs to support a large variety of speakers, differences in voice qualities, and a wide range of intensities potentially encountered during audio production. In this context, the present study will provide a detailed description of the Multi-band Excited WaveNet, a fully convolutional neural vocoder built around signal processing blocks. It will evaluate the performance of the vocoder when trained on a variety of multi-speaker and multi-singer databases, including an experimental evaluation of the neural vocoder trained on speech and singing voices. Addressing the problem of intensity variation, the study will introduce a new adaptive signal normalization scheme that allows for robust compensation for dynamic and static gain variations. Evaluations are performed using objective measures and a number of perceptual tests including different neural vocoder algorithms known from the literature. The results confirm that the proposed vocoder compares favorably to the state-of-the-art in its capacity to generalize to unseen voices and voice qualities. The remaining challenges will be discussed.

We conduct an investigation on various hyperparameters regarding neural networks used to generate spectral envelopes for singing synthesis. Two perceptive tests, where the first compares two models directly and the other ranks models with a mean opinion score, are performed. With these tests we show that when learning to predict spectral envelopes, 2d-convolutions are superior over previously proposed 1d-convolutions and that predicting multiple frames in an iterated fashion during training is superior over injecting noise to the input data. An experimental investigation whether learning to predict a probability distribution vs. single samples was performed but turned out to be inconclusive. A network architecture is proposed that incorporates the improvements which we found to be useful and we show in our experiments that this network produces better results than other stat-of-the-art methods.

No abstract