Electrical Engineering And Systems Science

In this paper we demonstrate the effectiveness of non-causal context for mitigating the effects of reverberation in deep-learning-based automatic speech recognition (ASR) systems. First, the value of non-causal context using a non-causal FIR filter is shown by comparing the contributions of previous vs. future information. Second, MLP- and LSTM-based dereverberation networks were trained to confirm the effects of causal and non-causal context when used in ASR systems trained with clean speech. The non-causal deep-learning-based dereverberation provides a 45% relative reduction in word error rate (WER) compared to the popular weighted prediction error (WPE) method in experiments with clean training in the REVERB challenge. Finally, an expanded multicondition training procedure used in combination with a semi-enhanced test utterance generation based on combinations of reverberated and dereverberated signals is proposed to reduce any artifacts or distortion that may be introduced by the non-causal dereverberation methods. The combination of both approaches provided average relative reductions in WER equal to 10.9% and 6.0% when compared to the baseline system obtained with the most recent REVERB challenge recipe without and with WPE, respectively.

In this work, we explore the constant-Q transform (CQT) for speech emotion recognition (SER). The CQT-based time-frequency analysis provides variable spectro-temporal resolution with higher frequency resolution at lower frequencies. Since lower-frequency regions of speech signal contain more emotion-related information than higher-frequency regions, the increased low-frequency resolution of CQT makes it more promising for SER than standard short-time Fourier transform (STFT). We present a comparative analysis of short-term acoustic features based on STFT and CQT for SER with deep neural network (DNN) as a back-end classifier. We optimize different parameters for both features. The CQT-based features outperform the STFT-based spectral features for SER experiments. Further experiments with cross-corpora evaluation demonstrate that the CQT-based systems provide better generalization with out-of-domain training data.

We introduce a novel method for emotion conversion in speech that does not require parallel training data. Our approach loosely relies on a cycle-GAN schema to minimize the reconstruction error from converting back and forth between emotion pairs. However, unlike the conventional cycle-GAN, our discriminator classifies whether a pair of input real and generated samples corresponds to the desired emotion conversion (e.g., A to B) or to its inverse (B to A). We will show that this setup, which we refer to as a variational cycle-GAN (VC-GAN), is equivalent to minimizing the empirical KL divergence between the source features and their cyclic counterpart. In addition, our generator combines a trainable deep network with a fixed generative block to implement a smooth and invertible transformation on the input features, in our case, the fundamental frequency (F0) contour. This hybrid architecture regularizes our adversarial training procedure. We use crowd sourcing to evaluate both the emotional saliency and the quality of synthesized speech. Finally, we show that our model generalizes to new speakers by modifying speech produced by Wavenet.

This paper extends recent work on nonlinear Independent Component Analysis (ICA) by introducing a theoretical framework for nonlinear Independent Subspace Analysis (ISA) in the presence of auxiliary variables. Observed high dimensional acoustic features like log Mel spectrograms can be considered as surface level manifestations of nonlinear transformations over individual multivariate sources of information like speaker characteristics, phonological content etc. Under assumptions of energy based models we use the theory of nonlinear ISA to propose an algorithm that learns unsupervised speech representations whose subspaces are independent and potentially highly correlated with the original non-stationary multivariate sources. We show how nonlinear ICA with auxiliary variables can be extended to a generic identifiable model for subspaces as well while also providing sufficient conditions for the identifiability of these high dimensional subspaces. Our proposed methodology is generic and can be integrated with standard unsupervised approaches to learn speech representations with subspaces that can theoretically capture independent higher order speech signals. We evaluate the gains of our algorithm when integrated with the Autoregressive Predictive Decoding (APC) model by showing empirical results on the speaker verification and phoneme recognition tasks.

We previously proposed a method that allows for nonparallel voice conversion (VC) by using a variant of generative adversarial networks (GANs) called StarGAN. The main features of our method, called StarGAN-VC, are as follows: First, it requires no parallel utterances, transcriptions, or time alignment procedures for speech generator training. Second, it can simultaneously learn mappings across multiple domains using a single generator network and thus fully exploit available training data collected from multiple domains to capture latent features that are common to all the domains. Third, it can generate converted speech signals quickly enough to allow real-time implementations and requires only several minutes of training examples to generate reasonably realistic-sounding speech. In this paper, we describe three formulations of StarGAN, including a newly introduced novel StarGAN variant called "Augmented classifier StarGAN (A-StarGAN)", and compare them in a nonparallel VC task. We also compare them with several baseline methods.

Recent work has shown that it is feasible to use generative adversarial networks (GANs) for speech enhancement, however, these approaches have not been compared to state-of-the-art (SOTA) non GAN-based approaches. Additionally, many loss functions have been proposed for GAN-based approaches, but they have not been adequately compared. In this study, we propose novel convolutional recurrent GAN (CRGAN) architectures for speech enhancement. Multiple loss functions are adopted to enable direct comparisons to other GAN-based systems. The benefits of including recurrent layers are also explored. Our results show that the proposed CRGAN model outperforms the SOTA GAN-based models using the same loss functions and it outperforms other non-GAN based systems, indicating the benefits of using a GAN for speech enhancement. Overall, the CRGAN model that combines an objective metric loss function with the mean squared error (MSE) provides the best performance over comparison approaches across many evaluation metrics.

In this paper, we exploit the properties of mean absolute error (MAE) as a loss function for the deep neural network (DNN) based vector-to-vector regression. The goal of this work is two-fold: (i) presenting performance bounds of MAE, and (ii) demonstrating new properties of MAE that make it more appropriate than mean squared error (MSE) as a loss function for DNN based vector-to-vector regression. First, we show that a generalized upper-bound for DNN-based vector- to-vector regression can be ensured by leveraging the known Lipschitz continuity property of MAE. Next, we derive a new generalized upper bound in the presence of additive noise. Finally, in contrast to conventional MSE commonly adopted to approximate Gaussian errors for regression, we show that MAE can be interpreted as an error modeled by Laplacian distribution. Speech enhancement experiments are conducted to corroborate our proposed theorems and validate the performance advantages of MAE over MSE for DNN based regression.

Audio event localization and detection (SELD) have been commonly tackled using multitask models. Such a model usually consists of a multi-label event classification branch with sigmoid cross-entropy loss for event activity detection and a regression branch with mean squared error loss for direction-of-arrival estimation. In this work, we propose a multitask regression model, in which both (multi-label) event detection and localization are formulated as regression problems and use the mean squared error loss homogeneously for model training. We show that the common combination of heterogeneous loss functions causes the network to underfit the data whereas the homogeneous mean squared error loss leads to better convergence and performance. Experiments on the development and validation sets of the DCASE 2020 SELD task demonstrate that the proposed system also outperforms the DCASE 2020 SELD baseline across all the detection and localization metrics, reducing the overall SELD error (the combined metric) by approximately 10% absolute.

Recent advances in self-supervised learning through contrastive training have shown that it is possible to learn a competitive speech recognition system with as little as 10 minutes of labeled data. However, these systems are computationally expensive since they require pre-training followed by fine-tuning in a large parameter space. We explore the performance of such systems without fine-tuning by training a state-of-the-art speech recognizer on the fixed representations from the computationally demanding wav2vec 2.0 framework. We find performance to decrease without fine-tuning and, in the extreme low-resource setting, wav2vec 2.0 is inferior to its predecessor. In addition, we find that wav2vec 2.0 representations live in a low dimensional subspace and that decorrelating the features of the representations can stabilize training of the automatic speech recognizer. Finally, we propose a bidirectional extension to the original wav2vec framework that consistently improves performance.

The advent of learning-based methods in speech enhancement has revived the need for robust and reliable training features that can compactly represent speech signals while preserving their vital information. Time-frequency domain features, such as the Short-Term Fourier Transform (STFT) and Mel-Frequency Cepstral Coefficients (MFCC), are preferred in many approaches. While the MFCC provide for a compact representation, they ignore the dynamics and distribution of energy in each mel-scale subband. In this work, a speech enhancement system based on Generative Adversarial Network (GAN) is implemented and tested with a combination of Audio FingerPrinting (AFP) features obtained from the MFCC and the Normalized Spectral Subband Centroids (NSSC). The NSSC capture the locations of speech formants and complement the MFCC in a crucial way. In experiments with diverse speakers and noise types, GAN-based speech enhancement with the proposed AFP feature combination achieves the best objective performance while reducing memory requirements and training time.