Giampiero Salvi

Xinwei Cao, Mengxuan Lu, Torbjørn Svendsen et al.

We address the problem of out-of-distribution (OOD) detection for target observations embedded in a subspace of the high dimensional data space. Using continuous normalizing flows (CNFs), we propose a Lagrangian sub-flow (LSF) framework designed to isolate and estimate the density for the relevant components in the representation and using the remaining components as context. Through experimentation with models for speech synthesis, we show that CNFs, similarly to other deep generative models (DGMs), are susceptible to the "likelihood paradox", where high likelihood is erroneously assigned to OOD samples. This is attributed to the inductive bias of DGMs that prioritize low-level structural details over high-level semantic coherence. To mitigate this phenomenon, we propose a number of geometric diagnostic signals based on the velocity field over the sub-flow trajectory. Based on these signals, we design metrics for the challenging task of zero-shot phoneme-level mispronunciation detection. Finally, we demonstrate the superiority of these metrics compared to likelihood-based methods on a real-world mispronunciation detection benchmark.

Mohammad Adiban, Kalin Stefanov, Sabato Marco Siniscalchi et al. · gatech

We propose a multi-layer variational autoencoder method, we call HR-VQVAE, that learns hierarchical discrete representations of the data. By utilizing a novel objective function, each layer in HR-VQVAE learns a discrete representation of the residual from previous layers through a vector quantized encoder. Furthermore, the representations at each layer are hierarchically linked to those at previous layers. We evaluate our method on the tasks of image reconstruction and generation. Experimental results demonstrate that the discrete representations learned by HR-VQVAE enable the decoder to reconstruct high-quality images with less distortion than the baseline methods, namely VQVAE and VQVAE-2. HR-VQVAE can also generate high-quality and diverse images that outperform state-of-the-art generative models, providing further verification of the efficiency of the learned representations. The hierarchical nature of HR-VQVAE i) reduces the decoding search time, making the method particularly suitable for high-load tasks and ii) allows to increase the codebook size without incurring the codebook collapse problem.

Mohammad Adiban, Kalin Stefanov, Sabato Marco Siniscalchi et al. · gatech

We address the video prediction task by putting forth a novel model that combines (i) a novel hierarchical residual learning vector quantized variational autoencoder (HR-VQVAE), and (ii) a novel autoregressive spatiotemporal predictive model (AST-PM). We refer to this approach as a sequential hierarchical residual learning vector quantized variational autoencoder (S-HR-VQVAE). By leveraging the intrinsic capabilities of HR-VQVAE at modeling still images with a parsimonious representation, combined with the AST-PM's ability to handle spatiotemporal information, S-HR-VQVAE can better deal with major challenges in video prediction. These include learning spatiotemporal information, handling high dimensional data, combating blurry prediction, and implicit modeling of physical characteristics. Extensive experimental results on four challenging tasks, namely KTH Human Action, TrafficBJ, Human3.6M, and Kitti, demonstrate that our model compares favorably against state-of-the-art video prediction techniques both in quantitative and qualitative evaluations despite a much smaller model size. Finally, we boost S-HR-VQVAE by proposing a novel training method to jointly estimate the HR-VQVAE and AST-PM parameters.

Giampiero Salvi

This paper describes the use of connectionist techniques in phonetic speech recognition with strong latency constraints. The constraints are imposed by the task of deriving the lip movements of a synthetic face in real time from the speech signal, by feeding the phonetic string into an articulatory synthesiser. Particular attention has been paid to analysing the interaction between the time evolution model learnt by the multi-layer perceptrons and the transition model imposed by the Viterbi decoder, in different latency conditions. Two experiments were conducted in which the time dependencies in the language model (LM) were controlled by a parameter. The results show a strong interaction between the three factors involved, namely the neural network topology, the length of time dependencies in the LM and the decoder latency.

Giampiero Salvi

This paper is concerned with automatic continuous speech recognition using trainable systems. The aim of this work is to build acoustic models for spoken Swedish. This is done employing hidden Markov models and using the SpeechDat database to train their parameters. Acoustic modeling has been worked out at a phonetic level, allowing general speech recognition applications, even though a simplified task (digits and natural number recognition) has been considered for model evaluation. Different kinds of phone models have been tested, including context independent models and two variations of context dependent models. Furthermore many experiments have been done with bigram language models to tune some of the system parameters. System performance over various speaker subsets with different sex, age and dialect has also been examined. Results are compared to previous similar studies showing a remarkable improvement.

Giampiero Salvi

This article investigates the possibility to use the class entropy of the output of a connectionist phoneme recogniser to predict time boundaries between phonetic classes. The rationale is that the value of the entropy should increase in proximity of a transition between two segments that are well modelled (known) by the recognition network since it is a measure of uncertainty. The advantage of this measure is its simplicity as the posterior probabilities of each class are available in connectionist phoneme recognition. The entropy and a number of measures based on differentiation of the entropy are used in isolation and in combination. The decision methods for predicting the boundaries range from simple thresholds to neural network based procedure. The different methods are compared with respect to their precision, measured in terms of the ratio between the number C of predicted boundaries within 10 or 20 msec of the reference and the total number of predicted boundaries, and recall, measured as the ratio between C and the total number of reference boundaries.

Yaroslav Getman, Tamás Grósz, Mikko Kurimo et al.

This paper presents the "Non-native Children's Automatic Speech Assessment" (NOCASA) - a data competition part of the IEEE MLSP 2025 conference. NOCASA challenges participants to develop new systems that can assess single-word pronunciations of young second language (L2) learners as part of a gamified pronunciation training app. To achieve this, several issues must be addressed, most notably the limited nature of available training data and the highly unbalanced distribution among the pronunciation level categories. To expedite the development, we provide a pseudo-anonymized training data (TeflonNorL2), containing 10,334 recordings from 44 speakers attempting to pronounce 205 distinct Norwegian words, human-rated on a 1 to 5 scale (number of stars that should be given in the game). In addition to the data, two already trained systems are released as official baselines: an SVM classifier trained on the ComParE_16 acoustic feature set and a multi-task wav2vec 2.0 model. The latter achieves the best performance on the challenge test set, with an unweighted average recall (UAR) of 36.37%.

Xinwei Cao, Zijian Fan, Torbjørn Svendsen et al.

Mispronunciation detection and diagnosis (MDD) is a significant part in modern computer aided language learning (CALL) systems. Within MDD, phoneme-level pronunciation assessment is key to helping L2 learners improve their pronunciation. However, most systems are based on a form of goodness of pronunciation (GOP) which requires pre-segmentation of speech into phonetic units. This limits the accuracy of these methods and the possibility to use modern CTC-based acoustic models for their evaluation. In this study, we first propose self-alignment GOP (GOP-SA) that enables the use of CTC-trained ASR models for MDD. Next, we define a more general alignment-free method that takes all possible alignments of the target phoneme into account (GOP-AF). We give a theoretical account of our definition of GOP-AF, an implementation that solves potential numerical issues as well as a proper normalization which makes the method applicable with acoustic models with different peakiness over time. We provide extensive experimental results on the CMU Kids and Speechocean762 datasets comparing the different definitions of our methods, estimating the dependency of GOP-AF on the peakiness of the acoustic models and on the amount of context around the target phoneme. Finally, we compare our methods with recent studies over the Speechocean762 data showing that the feature vectors derived from the proposed method achieve state-of-the-art results on phoneme-level pronunciation assessment.

Jerome Abdelnour, Jean Rouat, Giampiero Salvi

The goal of the Acoustic Question Answering (AQA) task is to answer a free-form text question about the content of an acoustic scene. It was inspired by the Visual Question Answering (VQA) task. In this paper, based on the previously introduced CLEAR dataset, we propose a new benchmark for AQA, namely CLEAR2, that emphasizes the specific challenges of acoustic inputs. These include handling of variable duration scenes, and scenes built with elementary sounds that differ between training and test set. We also introduce NAAQA, a neural architecture that leverages specific properties of acoustic inputs. The use of 1D convolutions in time and frequency to process 2D spectro-temporal representations of acoustic content shows promising results and enables reductions in model complexity. We show that time coordinate maps augment temporal localization capabilities which enhance performance of the network by ~17 percentage points. On the other hand, frequency coordinate maps have little influence on this task. NAAQA achieves 79.5% of accuracy on the AQA task with ~4 times fewer parameters than the previously explored VQA model. We evaluate the perfomance of NAAQA on an independent data set reconstructed from DAQA. We also test the addition of a MALiMo module in our model on both CLEAR2 and DAQA. We provide a detailed analysis of the results for the different question types. We release the code to produce CLEAR2 as well as NAAQA to foster research in this newly emerging machine learning task.

Mohammad Adiban, Arash Safari, Giampiero Salvi

In this study, we introduce a novel unsupervised countermeasure for smart grid power systems, based on generative adversarial networks (GANs). Given the pivotal role of smart grid systems (SGSs) in urban life, their security is of particular importance. In recent years, however, advances in the field of machine learning, have raised concerns about cyber attacks on these systems. Power systems, among the most important components of urban infrastructure, have, for example, been widely attacked by adversaries. Attackers disrupt power systems using false data injection attacks (FDIA), resulting in a breach of availability, integrity, or confidential principles of the system. Our model simulates possible attacks on power systems using multiple generators in a step-by-step interaction with a discriminator in the training phase. As a consequence, our system is robust to unseen attacks. Moreover, the proposed model considerably reduces the well-known mode collapse problem of GAN-based models. Our method is general and it can be potentially employed in a wide range of one of one-class classification tasks. The proposed model has low computational complexity and outperforms baseline systems about 14% and 41% in terms of accuracy on the highly imbalanced publicly available industrial control system (ICS) cyber attack power system dataset.

Jerome Abdelnour, Giampiero Salvi, Jean Rouat

We introduce the new task of Acoustic Question Answering (AQA) to promote research in acoustic reasoning. The AQA task consists of analyzing an acoustic scene composed by a combination of elementary sounds and answering questions that relate the position and properties of these sounds. The kind of relational questions asked, require that the models perform non-trivial reasoning in order to answer correctly. Although similar problems have been extensively studied in the domain of visual reasoning, we are not aware of any previous studies addressing the problem in the acoustic domain. We propose a method for generating the acoustic scenes from elementary sounds and a number of relevant questions for each scene using templates. We also present preliminary results obtained with two models (FiLM and MAC) that have been shown to work for visual reasoning.

Giovanni Saponaro, Lorenzo Jamone, Alexandre Bernardino et al.

We propose a developmental approach that allows a robot to interpret and describe the actions of human agents by reusing previous experience. The robot first learns the association between words and object affordances by manipulating the objects in its environment. It then uses this information to learn a mapping between its own actions and those performed by a human in a shared environment. It finally fuses the information from these two models to interpret and describe human actions in light of its own experience. In our experiments, we show that the model can be used flexibly to do inference on different aspects of the scene. We can predict the effects of an action on the basis of object properties. We can revise the belief that a certain action occurred, given the observed effects of the human action. In an early action recognition fashion, we can anticipate the effects when the action has only been partially observed. By estimating the probability of words given the evidence and feeding them into a pre-defined grammar, we can generate relevant descriptions of the scene. We believe that this is a step towards providing robots with the fundamental skills to engage in social collaboration with humans.

Jerome Abdelnour, Giampiero Salvi, Jean Rouat

We introduce the task of acoustic question answering (AQA) in the area of acoustic reasoning. In this task an agent learns to answer questions on the basis of acoustic context. In order to promote research in this area, we propose a data generation paradigm adapted from CLEVR (Johnson et al. 2017). We generate acoustic scenes by leveraging a bank elementary sounds. We also provide a number of functional programs that can be used to compose questions and answers that exploit the relationships between the attributes of the elementary sounds in each scene. We provide AQA datasets of various sizes as well as the data generation code. As a preliminary experiment to validate our data, we report the accuracy of current state of the art visual question answering models when they are applied to the AQA task without modifications. Although there is a plethora of question answering tasks based on text, image or video data, to our knowledge, we are the first to propose answering questions directly on audio streams. We hope this contribution will facilitate the development of research in the area.

Cheng Zhang, Cengiz Öztireli, Stephan Mandt et al.

The convergence speed of stochastic gradient descent (SGD) can be improved by actively selecting mini-batches. We explore sampling schemes where similar data points are less likely to be selected in the same mini-batch. In particular, we prove that such repulsive sampling schemes lowers the variance of the gradient estimator. This generalizes recent work on using Determinantal Point Processes (DPPs) for mini-batch diversification (Zhang et al., 2017) to the broader class of repulsive point processes. We first show that the phenomenon of variance reduction by diversified sampling generalizes in particular to non-stationary point processes. We then show that other point processes may be computationally much more efficient than DPPs. In particular, we propose and investigate Poisson Disk sampling---frequently encountered in the computer graphics community---for this task. We show empirically that our approach improves over standard SGD both in terms of convergence speed as well as final model performance.

Giampiero Salvi, Luis Montesano, Alexandre Bernardino et al.

We address the problem of bootstrapping language acquisition for an artificial system similarly to what is observed in experiments with human infants. Our method works by associating meanings to words in manipulation tasks, as a robot interacts with objects and listens to verbal descriptions of the interactions. The model is based on an affordance network, i.e., a mapping between robot actions, robot perceptions, and the perceived effects of these actions upon objects. We extend the affordance model to incorporate spoken words, which allows us to ground the verbal symbols to the execution of actions and the perception of the environment. The model takes verbal descriptions of a task as the input and uses temporal co-occurrence to create links between speech utterances and the involved objects, actions, and effects. We show that the robot is able form useful word-to-meaning associations, even without considering grammatical structure in the learning process and in the presence of recognition errors. These word-to-meaning associations are embedded in the robot's own understanding of its actions. Thus, they can be directly used to instruct the robot to perform tasks and also allow to incorporate context in the speech recognition task. We believe that the encouraging results with our approach may afford robots with a capacity to acquire language descriptors in their operation's environment as well as to shed some light as to how this challenging process develops with human infants.

Giovanni Saponaro, Lorenzo Jamone, Alexandre Bernardino et al.

A growing field in robotics and Artificial Intelligence (AI) research is human-robot collaboration, whose target is to enable effective teamwork between humans and robots. However, in many situations human teams are still superior to human-robot teams, primarily because human teams can easily agree on a common goal with language, and the individual members observe each other effectively, leveraging their shared motor repertoire and sensorimotor resources. This paper shows that for cognitive robots it is possible, and indeed fruitful, to combine knowledge acquired from interacting with elements of the environment (affordance exploration) with the probabilistic observation of another agent's actions. We propose a model that unites (i) learning robot affordances and word descriptions with (ii) statistical recognition of human gestures with vision sensors. We discuss theoretical motivations, possible implementations, and we show initial results which highlight that, after having acquired knowledge of its surrounding environment, a humanoid robot can generalize this knowledge to the case when it observes another agent (human partner) performing the same motor actions previously executed during training.

Kalin Stefanov, Jonas Beskow, Giampiero Salvi

This paper presents a self-supervised method for visual detection of the active speaker in a multi-person spoken interaction scenario. Active speaker detection is a fundamental prerequisite for any artificial cognitive system attempting to acquire language in social settings. The proposed method is intended to complement the acoustic detection of the active speaker, thus improving the system robustness in noisy conditions. The method can detect an arbitrary number of possibly overlapping active speakers based exclusively on visual information about their face. Furthermore, the method does not rely on external annotations, thus complying with cognitive development. Instead, the method uses information from the auditory modality to support learning in the visual domain. This paper reports an extensive evaluation of the proposed method using a large multi-person face-to-face interaction dataset. The results show good performance in a speaker dependent setting. However, in a speaker independent setting the proposed method yields a significantly lower performance. We believe that the proposed method represents an essential component of any artificial cognitive system or robotic platform engaging in social interactions.

Akash Kumar Dhaka, Giampiero Salvi

We propose the application of a semi-supervised learning method to improve the performance of acoustic modelling for automatic speech recognition based on deep neural net- works. As opposed to unsupervised initialisation followed by supervised fine tuning, our method takes advantage of both unlabelled and labelled data simultaneously through mini- batch stochastic gradient descent. We tested the method with varying proportions of labelled vs unlabelled observations in frame-based phoneme classification on the TIMIT database. Our experiments show that the method outperforms standard supervised training for an equal amount of labelled data and provides competitive error rates compared to state-of-the-art graph-based semi-supervised learning techniques.

Akash Kumar Dhaka, Giampiero Salvi

We present a systematic analysis on the performance of a phonetic recogniser when the window of input features is not symmetric with respect to the current frame. The recogniser is based on Context Dependent Deep Neural Networks (CD-DNNs) and Hidden Markov Models (HMMs). The objective is to reduce the latency of the system by reducing the number of future feature frames required to estimate the current output. Our tests performed on the TIMIT database show that the performance does not degrade when the input window is shifted up to 5 frames in the past compared to common practice (no future frame). This corresponds to improving the latency by 50 ms in our settings. Our tests also show that the best results are not obtained with the symmetric window commonly employed, but with an asymmetric window with eight past and two future context frames, although this observation should be confirmed on other data sets. The reduction in latency suggested by our results is critical for specific applications such as real-time lip synchronisation for tele-presence, but may also be beneficial in general applications to improve the lag in human-machine spoken interaction.