The class of disentangled sequential auto-encoders factorises speech into time-invariant (global) and time-variant (local) representations for speaker identity and linguistic content, respectively. Many of the existing models employ this assumption to tackle zero-shot voice conversion (VC), which converts speaker characteristics of any given utterance to any novel speakers while preserving the linguistic content. However, balancing capacity between the two representations is intricate, as the global representation tends to collapse due to its lower information capacity along the time axis than that of the local representation. We propose a simple and effective dropout technique that applies an information bottleneck to the local representation via multiplicative Gaussian noise, in order to encourage the usage of the global one. We endow existing zero-shot VC models with the proposed method and show significant improvements in speaker conversion in terms of speaker verification acceptance rate and comparable or better intelligibility measured in character error rate.

Disentangled representation learning seeks to align individual dimensions or separate groups of coordinates of latent factors with attributes of observed data such that perturbing certain latent factors uniquely changes particular attributes. A main challenge in unsupervised disentanglement using autoencoders is that strong regularisation, while necessary for consistent disentanglement, comes at the expense of accurate data reconstruction. To address this, we introduce a teacher-student framework that incorporates a variational sequential autoencoder and a Jacobian constraint that regularises the variation of observations relative to latent factors. In real-world audio recordings of musical instruments, our approach outperforms a state-of-the-art method in both sampling quality and unsupervised pitch-timbre disentanglement.

Disentangled sequential autoencoders (DSAEs) represent a class of probabilistic graphical models that describes an observed sequence with dynamic latent variables and a static latent variable. The former encode information at a frame rate identical to the observation, while the latter globally governs the entire sequence. This introduces an inductive bias and facilitates unsupervised disentanglement of the underlying local and global factors. In this paper, we show that the vanilla DSAE suffers from being sensitive to the choice of model architecture and capacity of the dynamic latent variables, and is prone to collapse the static latent variable. As a countermeasure, we propose TS-DSAE, a two-stage training framework that first learns sequence-level prior distributions, which are subsequently employed to regularise the model and facilitate auxiliary objectives to promote disentanglement. The proposed framework is fully unsupervised and robust against the global factor collapse problem across a wide range of model configurations. It also avoids typical solutions such as adversarial training which usually involves laborious parameter tuning, and domain-specific data augmentation. We conduct quantitative and qualitative evaluations to demonstrate its robustness in terms of disentanglement on both artificial and real-world music audio datasets.

Recent advances in automatic music transcription (AMT) have achieved highly accurate polyphonic piano transcription results by incorporating onset and offset detection. The existing literature, however, focuses mainly on the leverage of deep and complex models to achieve state-of-the-art (SOTA) accuracy, without understanding model behaviour. In this paper, we conduct a comprehensive examination of the Onsets-and-Frames AMT model, and pinpoint the essential components contributing to a strong AMT performance. This is achieved through exploitation of a modified additive attention mechanism. The experimental results suggest that the attention mechanism beyond a moderate temporal context does not benefit the model, and that rule-based post-processing is largely responsible for the SOTA performance. We also demonstrate that the onsets are the most significant attentive feature regardless of model complexity. The findings encourage AMT research to weigh more on both a robust onset detector and an effective post-processor.

We propose a flexible framework that deals with both singer conversion and singers vocal technique conversion. The proposed model is trained on non-parallel corpora, accommodates many-to-many conversion, and leverages recent advances of variational autoencoders. It employs separate encoders to learn disentangled latent representations of singer identity and vocal technique separately, with a joint decoder for reconstruction. Conversion is carried out by simple vector arithmetic in the learned latent spaces. Both a quantitative analysis as well as a visualization of the converted spectrograms show that our model is able to disentangle singer identity and vocal technique and successfully perform conversion of these attributes. To the best of our knowledge, this is the first work to jointly tackle conversion of singer identity and vocal technique based on a deep learning approach.

Most of the state-of-the-art automatic music transcription (AMT) models break down the main transcription task into sub-tasks such as onset prediction and offset prediction and train them with onset and offset labels. These predictions are then concatenated together and used as the input to train another model with the pitch labels to obtain the final transcription. We attempt to use only the pitch labels (together with spectrogram reconstruction loss) and explore how far this model can go without introducing supervised sub-tasks. In this paper, we do not aim at achieving state-of-the-art transcription accuracy, instead, we explore the effect that spectrogram reconstruction has on our AMT model. Our proposed model consists of two U-nets: the first U-net transcribes the spectrogram into a posteriorgram, and a second U-net transforms the posteriorgram back into a spectrogram. A reconstruction loss is applied between the original spectrogram and the reconstructed spectrogram to constrain the second U-net to focus only on reconstruction. We train our model on three different datasets: MAPS, MAESTRO, and MusicNet. Our experiments show that adding the reconstruction loss can generally improve the note-level transcription accuracy when compared to the same model without the reconstruction part. Moreover, it can also boost the frame-level precision to be higher than the state-of-the-art models. The feature maps learned by our U-net contain gridlike structures (not present in the baseline model) which implies that with the presence of the reconstruction loss, the model is probably trying to count along both the time and frequency axis, resulting in a higher note-level transcription accuracy.

In this paper, we adapt triplet neural networks (TNNs) to a regression task, music emotion prediction. Since TNNs were initially introduced for classification, and not for regression, we propose a mechanism that allows them to provide meaningful low dimensional representations for regression tasks. We then use these new representations as the input for regression algorithms such as support vector machines and gradient boosting machines. To demonstrate the TNNs' effectiveness at creating meaningful representations, we compare them to different dimensionality reduction methods on music emotion prediction, i.e., predicting valence and arousal values from musical audio signals. Our results on the DEAM dataset show that by using TNNs we achieve 90% feature dimensionality reduction with a 9% improvement in valence prediction and 4% improvement in arousal prediction with respect to our baseline models (without TNN). Our TNN method outperforms other dimensionality reduction methods such as principal component analysis (PCA) and autoencoders (AE). This shows that, in addition to providing a compact latent space representation of audio features, the proposed approach achieves higher performance than the baseline models.

Expressive singing voice correction is an appealing but challenging problem. A robust time-warping algorithm which synchronizes two singing recordings can provide a promising solution. We thereby propose to address the problem by canonical time warping (CTW) which aligns amateur singing recordings to professional ones. A new pitch contour is generated given the alignment information, and a pitch-corrected singing is synthesized back through the vocoder. The objective evaluation shows that CTW is robust against pitch-shifting and time-stretching effects, and the subjective test demonstrates that CTW prevails the other methods including DTW and the commercial auto-tuning software. Finally, we demonstrate the applicability of the proposed method in a practical, real-world scenario.