Abstract:We propose a semi-supervised singing synthesizer, which is able to learn new voices from audio data only, without any annotations such as phonetic segmentation. Our system is an encoder-decoder model with two encoders, linguistic and acoustic, and one (acoustic) decoder. In a first step, the system is trained in a supervised manner, using a labelled multi-singer dataset. Here, we ensure that the embeddings produced by both encoders are similar, so that we can later use the model with either acoustic or linguistic input features. To learn a new voice in an unsupervised manner, the pretrained acoustic encoder is used to train a decoder for the target singer. Finally, at inference, the pretrained linguistic encoder is used together with the decoder of the new voice, to produce acoustic features from linguistic input. We evaluate our system with a listening test and show that the results are comparable to those obtained with an equivalent supervised approach.
Abstract:We present a deep learning based methodology for extracting the singing voice signal from a musical mixture based on the underlying linguistic content. Our model follows an encoder decoder architecture and takes as input the magnitude component of the spectrogram of a musical mixture with vocals. The encoder part of the model is trained via knowledge distillation using a teacher network to learn a content embedding, which is decoded to generate the corresponding vocoder features. Using this methodology, we are able to extract the unprocessed raw vocal signal from the mixture even for a processed mixture dataset with singers not seen during training. While the nature of our system makes it incongruous with traditional objective evaluation metrics, we use subjective evaluation via listening tests to compare the methodology to state-of-the-art deep learning based source separation algorithms. We also provide sound examples and source code for reproducibility.
Abstract:We propose a sequence-to-sequence singing synthesizer, which avoids the need for training data with pre-aligned phonetic and acoustic features. Rather than the more common approach of a content-based attention mechanism combined with an autoregressive decoder, we use a different mechanism suitable for feed-forward synthesis. Given that phonetic timings in singing are highly constrained by the musical score, we derive an approximate initial alignment with the help of a simple duration model. Then, using a decoder based on a feed-forward variant of the Transformer model, a series of self-attention and convolutional layers refines the result of the initial alignment to reach the target acoustic features. Advantages of this approach include faster inference and avoiding the exposure bias issues that affect autoregressive models trained by teacher forcing. We evaluate the effectiveness of this model compared to an autoregressive baseline, the importance of self-attention, and the importance of the accuracy of the duration model.
Abstract:There are many use cases in singing synthesis where creating voices from small amounts of data is desirable. In text-to-speech there have been several promising results that apply voice cloning techniques to modern deep learning based models. In this work, we adapt one such technique to the case of singing synthesis. By leveraging data from many speakers to first create a multispeaker model, small amounts of target data can then efficiently adapt the model to new unseen voices. We evaluate the system using listening tests across a number of different use cases, languages and kinds of data.
Abstract:This paper summarizes some recent advances on a set of tasks related to the processing of singing using state-of-the-art deep learning techniques. We discuss their achievements in terms of accuracy and sound quality, and the current challenges, such as availability of data and computing resources. We also discuss the impact that these advances do and will have on listeners and singers when they are integrated in commercial applications.
Abstract:We present a new model for singing synthesis based on a modified version of the WaveNet architecture. Instead of modeling raw waveform, we model features produced by a parametric vocoder that separates the influence of pitch and timbre. This allows conveniently modifying pitch to match any target melody, facilitates training on more modest dataset sizes, and significantly reduces training and generation times. Our model makes frame-wise predictions using mixture density outputs rather than categorical outputs in order to reduce the required parameter count. As we found overfitting to be an issue with the relatively small datasets used in our experiments, we propose a method to regularize the model and make the autoregressive generation process more robust to prediction errors. Using a simple multi-stream architecture, harmonic, aperiodic and voiced/unvoiced components can all be predicted in a coherent manner. We compare our method to existing parametric statistical and state-of-the-art concatenative methods using quantitative metrics and a listening test. While naive implementations of the autoregressive generation algorithm tend to be inefficient, using a smart algorithm we can greatly speed up the process and obtain a system that's competitive in both speed and quality.