Zoran Cvetkovic

Erfan Loweimi, Andrea Carmantini, Peter Bell et al.

In this paper, we analyse the error patterns of the raw waveform acoustic models in TIMIT's phone recognition task. Our analysis goes beyond the conventional phone error rate (PER) metric. We categorise the phones into three groups: {affricate, diphthong, fricative, nasal, plosive, semi-vowel, vowel, silence}, {consonant, vowel+, silence}, and {voiced, unvoiced, silence} and, compute the PER for each broad phonetic class in each category. We also construct a confusion matrix for each category using the substitution errors and compare the confusion patterns with those of the Filterbank and Wav2vec 2.0 systems. Our raw waveform acoustic models consists of parametric (Sinc2Net) or non-parametric CNNs and Bidirectional LSTMs, achieving down to 13.7%/15.2% PERs on TIMIT Dev/Test sets, outperforming reported PERs for raw waveform models in the literature. We also investigate the impact of transfer learning from WSJ on the phonetic error patterns and confusion matrices. It reduces the PER to 11.8%/13.7% on the Dev/Test sets.

Dino Oglic, Zoran Cvetkovic, Peter Sollich et al.

We study the problem of learning robust acoustic models in adverse environments, characterized by a significant mismatch between training and test conditions. This problem is of paramount importance for the deployment of speech recognition systems that need to perform well in unseen environments. First, we characterize data augmentation theoretically as an instance of vicinal risk minimization, which aims at improving risk estimates during training by replacing the delta functions that define the empirical density over the input space with an approximation of the marginal population density in the vicinity of the training samples. More specifically, we assume that local neighborhoods centered at training samples can be approximated using a mixture of Gaussians, and demonstrate theoretically that this can incorporate robust inductive bias into the learning process. We then specify the individual mixture components implicitly via data augmentation schemes, designed to address common sources of spurious correlations in acoustic models. To avoid potential confounding effects on robustness due to information loss, which has been associated with standard feature extraction techniques (e.g., FBANK and MFCC features), we focus on the waveform-based setting. Our empirical results show that the approach can generalize to unseen noise conditions, with 150% relative improvement in out-of-distribution generalization compared to training using the standard risk minimization principle. Moreover, the results demonstrate competitive performance relative to models learned using a training sample designed to match the acoustic conditions characteristic of test utterances.

Jan Rosenzweig, Zoran Cvetkovic, Ivana Rosenzweig

A fundamental bottleneck in utilising complex machine learning systems for critical applications has been not knowing why they do and what they do, thus preventing the development of any crucial safety protocols. To date, no method exist that can provide full insight into the granularity of the neural network's decision process. In the past, saliency maps were an early attempt at resolving this problem through sensitivity calculations, whereby dimensions of a data point are selected based on how sensitive the output of the system is to them. However, the success of saliency maps has been at best limited, mainly due to the fact that they interpret the underlying learning system through a linear approximation. We present a novel class of methods for generating nonlinear saliency maps which fully account for the nonlinearity of the underlying learning system. While agreeing with linear saliency maps on simple problems where linear saliency maps are correct, they clearly identify more specific drivers of classification on complex examples where nonlinearities are more pronounced. This new class of methods significantly aids interpretability of deep neural networks and related machine learning systems. Crucially, they provide a starting point for their more broad use in serious applications, where 'why' is equally important as 'what'.

Jan Rosenzweig, Zoran Cvetkovic, Ivana Rosenzweig

We introduce the new "Goldilocks" class of activation functions, which non-linearly deform the input signal only locally when the input signal is in the appropriate range. The small local deformation of the signal enables better understanding of how and why the signal is transformed through the layers. Numerical results on CIFAR-10 and CIFAR-100 data sets show that Goldilocks networks perform better than, or comparably to SELU and RELU, while introducing tractability of data deformation through the layers.

Jingjing Zhang, Osvaldo Simeone, Zoran Cvetkovic et al.

Quantifying the directionality of information flow is instrumental in understanding, and possibly controlling, the operation of many complex systems, such as transportation, social, neural, or gene-regulatory networks. The standard Transfer Entropy (TE) metric follows Granger's causality principle by measuring the Mutual Information (MI) between the past states of a source signal $X$ and the future state of a target signal $Y$ while conditioning on past states of $Y$. Hence, the TE quantifies the improvement, as measured by the log-loss, in the prediction of the target sequence $Y$ that can be accrued when, in addition to the past of $Y$, one also has available past samples from $X$. However, by conditioning on the past of $Y$, the TE also measures information that can be synergistically extracted by observing both the past of $X$ and $Y$, and not solely the past of $X$. Building on a private key agreement formulation, the Intrinsic TE (ITE) aims to discount such synergistic information to quantify the degree to which $X$ is \emph{individually} predictive of $Y$, independent of $Y$'s past. In this paper, an estimator of the ITE is proposed that is inspired by the recently proposed Mutual Information Neural Estimation (MINE). The estimator is based on variational bound on the KL divergence, two-sample neural network classifiers, and the pathwise estimator of Monte Carlo gradients.

Qi Yu, Wei Dai, Zoran Cvetkovic et al.

Algorithms for learning a dictionary to sparsely represent a given dataset typically alternate between sparse coding and dictionary update stages. Methods for dictionary update aim to minimise expansion error by updating dictionary vectors and expansion coefficients given patterns of non-zero coefficients obtained in the sparse coding stage. We propose a block total least squares (BLOTLESS) algorithm for dictionary update. BLOTLESS updates a block of dictionary elements and the corresponding sparse coefficients simultaneously. In the error free case, three necessary conditions for exact recovery are identified. Lower bounds on the number of training data are established so that the necessary conditions hold with high probability. Numerical simulations show that the bounds approximate well the number of training data needed for exact dictionary recovery. Numerical experiments further demonstrate several benefits of dictionary learning with BLOTLESS update compared with state-of-the-art algorithms especially when the amount of training data is small.

Dino Oglic, Zoran Cvetkovic, Peter Sollich

We investigate the potential of stochastic neural networks for learning effective waveform-based acoustic models. The waveform-based setting, inherent to fully end-to-end speech recognition systems, is motivated by several comparative studies of automatic and human speech recognition that associate standard non-adaptive feature extraction techniques with information loss which can adversely affect robustness. Stochastic neural networks, on the other hand, are a class of models capable of incorporating rich regularization mechanisms into the learning process. We consider a deep convolutional neural network that first decomposes speech into frequency sub-bands via an adaptive parametric convolutional block where filters are specified by cosine modulations of compactly supported windows. The network then employs standard non-parametric 1D convolutions to extract relevant spectro-temporal patterns while gradually compressing the structured high dimensional representation generated by the parametric block. We rely on a probabilistic parametrization of the proposed neural architecture and learn the model using stochastic variational inference. This requires evaluation of an analytically intractable integral defining the Kullback-Leibler divergence term responsible for regularization, for which we propose an effective approximation based on the Gauss-Hermite quadrature. Our empirical results demonstrate a superior performance of the proposed approach over comparable waveform-based baselines and indicate that it could lead to robustness. Moreover, the approach outperforms a recently proposed deep convolutional neural network for learning of robust acoustic models with standard FBANK features.

Enzo De Sena, Huseyin Hacihabiboglu, Zoran Cvetkovic et al.

An acoustic reverberator consisting of a network of delay lines connected via scattering junctions is proposed. All parameters of the reverberator are derived from physical properties of the enclosure it simulates. It allows for simulation of unequal and frequency-dependent wall absorption, as well as directional sources and microphones. The reverberator renders the first-order reflections exactly, while making progressively coarser approximations of higher-order reflections. The rate of energy decay is close to that obtained with the image method (IM) and consistent with the predictions of Sabine and Eyring equations. The time evolution of the normalized echo density, which was previously shown to be correlated with the perceived texture of reverberation, is also close to that of IM. However, its computational complexity is one to two orders of magnitude lower, comparable to the computational complexity of a feedback delay network (FDN), and its memory requirements are negligible.

Matthew Ager, Zoran Cvetkovic, Peter Sollich

Speech representation and modelling in high-dimensional spaces of acoustic waveforms, or a linear transformation thereof, is investigated with the aim of improving the robustness of automatic speech recognition to additive noise. The motivation behind this approach is twofold: (i) the information in acoustic waveforms that is usually removed in the process of extracting low-dimensional features might aid robust recognition by virtue of structured redundancy analogous to channel coding, (ii) linear feature domains allow for exact noise adaptation, as opposed to representations that involve non-linear processing which makes noise adaptation challenging. Thus, we develop a generative framework for phoneme modelling in high-dimensional linear feature domains, and use it in phoneme classification and recognition tasks. Results show that classification and recognition in this framework perform better than analogous PLP and MFCC classifiers below 18 dB SNR. A combination of the high-dimensional and MFCC features at the likelihood level performs uniformly better than either of the individual representations across all noise levels.

Jibran Yousafzai, Zoran Cvetkovic, Peter Sollich et al.

This work proposes a novel support vector machine (SVM) based robust automatic speech recognition (ASR) front-end that operates on an ensemble of the subband components of high-dimensional acoustic waveforms. The key issues of selecting the appropriate SVM kernels for classification in frequency subbands and the combination of individual subband classifiers using ensemble methods are addressed. The proposed front-end is compared with state-of-the-art ASR front-ends in terms of robustness to additive noise and linear filtering. Experiments performed on the TIMIT phoneme classification task demonstrate the benefits of the proposed subband based SVM front-end: it outperforms the standard cepstral front-end in the presence of noise and linear filtering for signal-to-noise ratio (SNR) below 12-dB. A combination of the proposed front-end with a conventional front-end such as MFCC yields further improvements over the individual front ends across the full range of noise levels.

Yaqub Alwan, Zoran Cvetkovic, Michael Curtis

We studied classification of human ECGs labelled as normal sinus rhythm, ventricular fibrillation and ventricular tachycardia by means of support vector machines in different representation spaces, using different observation lengths. ECG waveform segments of duration 0.5-4 s, their Fourier magnitude spectra, and lower dimensional projections of Fourier magnitude spectra were used for classification. All considered representations were of much higher dimension than in published studies. Classification accuracy improved with segment duration up to 2 s, with 4 s providing little improvement. We found that it is possible to discriminate between ventricular tachycardia and ventricular fibrillation by the present approach with much shorter runs of ECG (2 s, minimum 86% sensitivity per class) than previously imagined. Ensembles of classifiers acting on 1 s segments taken over 5 s observation windows gave best results, with sensitivities of detection for all classes exceeding 93%.