Electrical Engineering And Systems Science

Multi-speaker speech synthesis is a technique for modeling multiple speakers' voices with a single model. Although many approaches using deep neural networks (DNNs) have been proposed, DNNs are prone to overfitting when the amount of training data is limited. We propose a framework for multi-speaker speech synthesis using deep Gaussian processes (DGPs); a DGP is a deep architecture of Bayesian kernel regressions and thus robust to overfitting. In this framework, speaker information is fed to duration/acoustic models using speaker codes. We also examine the use of deep Gaussian process latent variable models (DGPLVMs). In this approach, the representation of each speaker is learned simultaneously with other model parameters, and therefore the similarity or dissimilarity of speakers is considered efficiently. We experimentally evaluated two situations to investigate the effectiveness of the proposed methods. In one situation, the amount of data from each speaker is balanced (speaker-balanced), and in the other, the data from certain speakers are limited (speaker-imbalanced). Subjective and objective evaluation results showed that both the DGP and DGPLVM synthesize multi-speaker speech more effective than a DNN in the speaker-balanced situation. We also found that the DGPLVM outperforms the DGP significantly in the speaker-imbalanced situation.

Target speaker extraction, which aims at extracting a target speaker's voice from a mixture of voices using audio, visual or locational clues, has received much interest. Recently an audio-visual target speaker extraction has been proposed that extracts target speech by using complementary audio and visual clues. Although audio-visual target speaker extraction offers a more stable performance than single modality methods for simulated data, its adaptation towards realistic situations has not been fully explored as well as evaluations on real recorded mixtures. One of the major issues to handle realistic situations is how to make the system robust to clue corruption because in real recordings both clues may not be equally reliable, e.g. visual clues may be affected by occlusions. In this work, we propose a novel attention mechanism for multi-modal fusion and its training methods that enable to effectively capture the reliability of the clues and weight the more reliable ones. Our proposals improve signal to distortion ratio (SDR) by 1.0 dB over conventional fusion mechanisms on simulated data. Moreover, we also record an audio-visual dataset of simultaneous speech with realistic visual clue corruption and show that audio-visual target speaker extraction with our proposals successfully work on real data.

Alzheimer's disease is estimated to affect around 50 million people worldwide and is rising rapidly, with a global economic burden of nearly a trillion dollars. This calls for scalable, cost-effective, and robust methods for detection of Alzheimer's dementia (AD). We present a novel architecture that leverages acoustic, cognitive, and linguistic features to form a multimodal ensemble system. It uses specialized artificial neural networks with temporal characteristics to detect AD and its severity, which is reflected through Mini-Mental State Exam (MMSE) scores. We first evaluate it on the ADReSS challenge dataset, which is a subject-independent and balanced dataset matched for age and gender to mitigate biases, and is available through DementiaBank. Our system achieves state-of-the-art test accuracy, precision, recall, and F1-score of 83.3% each for AD classification, and state-of-the-art test root mean squared error (RMSE) of 4.60 for MMSE score regression. To the best of our knowledge, the system further achieves state-of-the-art AD classification accuracy of 88.0% when evaluated on the full benchmark DementiaBank Pitt database. Our work highlights the applicability and transferability of spontaneous speech to produce a robust inductive transfer learning model, and demonstrates generalizability through a task-agnostic feature-space. The source code is available at this https URL

For many small- and medium-vocabulary tasks, audio-visual speech recognition can significantly improve the recognition rates compared to audio-only systems. However, there is still an ongoing debate regarding the best combination strategy for multi-modal information, which should allow for the translation of these gains to large-vocabulary recognition. While an integration at the level of state-posterior probabilities, using dynamic stream weighting, is almost universally helpful for small-vocabulary systems, in large-vocabulary speech recognition, the recognition accuracy remains difficult to improve. In the following, we specifically consider the large-vocabulary task of the LRS2 database, and we investigate a broad range of integration strategies, comparing early integration and end-to-end learning with many versions of hybrid recognition and dynamic stream weighting. One aspect, which is shown to provide much benefit here, is the use of dynamic stream reliability indicators, which allow for hybrid architectures to strongly profit from the inclusion of visual information whenever the audio channel is distorted even slightly.

In this work, we explore a multimodal semi-supervised learning approach for punctuation prediction by learning representations from large amounts of unlabelled audio and text data. Conventional approaches in speech processing typically use forced alignment to encoder per frame acoustic features to word level features and perform multimodal fusion of the resulting acoustic and lexical representations. As an alternative, we explore attention based multimodal fusion and compare its performance with forced alignment based fusion. Experiments conducted on the Fisher corpus show that our proposed approach achieves ~6-9% and ~3-4% absolute improvement (F1 score) over the baseline BLSTM model on reference transcripts and ASR outputs respectively. We further improve the model robustness to ASR errors by performing data augmentation with N-best lists which achieves up to an additional ~2-6% improvement on ASR outputs. We also demonstrate the effectiveness of semi-supervised learning approach by performing ablation study on various sizes of the corpus. When trained on 1 hour of speech and text data, the proposed model achieved ~9-18% absolute improvement over baseline model.

This paper addresses the extraction of multiple F0 values from polyphonic and a cappella vocal performances using convolutional neural networks (CNNs). We address the major challenges of ensemble singing, i.e., all melodic sources are vocals and singers sing in harmony. We build upon an existing architecture to produce a pitch salience function of the input signal, where the harmonic constant-Q transform (HCQT) and its associated phase differentials are used as an input representation. The pitch salience function is subsequently thresholded to obtain a multiple F0 estimation output. For training, we build a dataset that comprises several multi-track datasets of vocal quartets with F0 annotations. This work proposes and evaluates a set of CNNs for this task in diverse scenarios and data configurations, including recordings with additional reverb. Our models outperform a state-of-the-art method intended for the same music genre when evaluated with an increased F0 resolution, as well as a general-purpose method for multi-F0 estimation. We conclude with a discussion on future research directions.

Music source separation is a core task in music information retrieval which has seen a dramatic improvement in the past years. Nevertheless, most of the existing systems focus exclusively on the problem of source separation itself and ignore the utilization of other~---possibly related---~MIR tasks which could lead to additional quality gains. In this work, we propose a novel multitask structure to investigate using instrument activation information to improve source separation performance. Furthermore, we investigate our system on six independent instruments, a more realistic scenario than the three instruments included in the widely-used MUSDB dataset, by leveraging a combination of the MedleyDB and Mixing Secrets datasets. The results show that our proposed multitask model outperforms the baseline Open-Unmix model on the mixture of Mixing Secrets and MedleyDB dataset while maintaining comparable performance on the MUSDB dataset.

Music annotation has always been one of the critical topics in the field of Music Information Retrieval (MIR). Traditional models use supervised learning for music annotation tasks. However, as supervised machine learning approaches increase in complexity, the increasing need for more annotated training data can often not be matched with available data. In this paper, a new self-supervised music acoustic representation learning approach named MusiCoder is proposed. Inspired by the success of BERT, MusiCoder builds upon the architecture of self-attention bidirectional transformers. Two pre-training objectives, including Contiguous Frames Masking (CFM) and Contiguous Channels Masking (CCM), are designed to adapt BERT-like masked reconstruction pre-training to continuous acoustic frame domain. The performance of MusiCoder is evaluated in two downstream music annotation tasks. The results show that MusiCoder outperforms the state-of-the-art models in both music genre classification and auto-tagging tasks. The effectiveness of MusiCoder indicates a great potential of a new self-supervised learning approach to understand music: first apply masked reconstruction tasks to pre-train a transformer-based model with massive unlabeled music acoustic data, and then finetune the model on specific downstream tasks with labeled data.

The analysis of the structure of musical pieces is a task that remains a challenge for Artificial Intelligence, especially in the field of Deep Learning. It requires prior identification of structural boundaries of the music pieces. This structural boundary analysis has recently been studied with unsupervised methods and \textit{end-to-end} techniques such as Convolutional Neural Networks (CNN) using Mel-Scaled Log-magnitude Spectograms features (MLS), Self-Similarity Matrices (SSM) or Self-Similarity Lag Matrices (SSLM) as inputs and trained with human annotations. Several studies have been published divided into unsupervised and \textit{end-to-end} methods in which pre-processing is done in different ways, using different distance metrics and audio characteristics, so a generalized pre-processing method to compute model inputs is missing. The objective of this work is to establish a general method of pre-processing these inputs by comparing the inputs calculated from different pooling strategies, distance metrics and audio characteristics, also taking into account the computing time to obtain them. We also establish the most effective combination of inputs to be delivered to the CNN in order to establish the most efficient way to extract the limits of the structure of the music pieces. With an adequate combination of input matrices and pooling strategies we obtain a measurement accuracy F 1 of 0.411 that outperforms the current one obtained under the same conditions.

High-level musical qualities (such as emotion) are often abstract, subjective, and hard to quantify. Given these difficulties, it is not easy to learn good feature representations with supervised learning techniques, either because of the insufficiency of labels, or the subjectiveness (and hence large variance) in human-annotated labels. In this paper, we present a framework that can learn high-level feature representations with a limited amount of data, by first modelling their corresponding quantifiable low-level attributes. We refer to our proposed framework as Music FaderNets, which is inspired by the fact that low-level attributes can be continuously manipulated by separate "sliding faders" through feature disentanglement and latent regularization techniques. High-level features are then inferred from the low-level representations through semi-supervised clustering using Gaussian Mixture Variational Autoencoders (GM-VAEs). Using arousal as an example of a high-level feature, we show that the "faders" of our model are disentangled and change linearly w.r.t. the modelled low-level attributes of the generated output music. Furthermore, we demonstrate that the model successfully learns the intrinsic relationship between arousal and its corresponding low-level attributes (rhythm and note density), with only 1% of the training set being labelled. Finally, using the learnt high-level feature representations, we explore the application of our framework in style transfer tasks across different arousal states. The effectiveness of this approach is verified through a subjective listening test.