Privacy-oriented manipulation of speaker representations

Speaker embeddings are ubiquitous, with applications ranging from speaker recognition and diarization to speech synthesis and voice anonymisation. The amount of information held by these embeddings lends them versatility, but also raises privacy concerns. Speaker embeddings have been shown to contain information on age, sex, health and more, which speakers may want to keep private, especially when this information is not required for the target task. In this work, we propose a method for removing and manipulating private attributes from speaker embeddings that leverages a Vector-Quantized Variational Autoencoder architecture, combined with an adversarial classifier and a novel mutual information loss. We validate our model on two attributes, sex and age, and perform experiments with ignorant and fully-informed attackers, and with in-domain and out-of-domain data.

Memory-augmented conformer for improved end-to-end long-form ASR

Conformers have recently been proposed as a promising modelling approach for automatic speech recognition (ASR), outperforming recurrent neural network-based approaches and transformers. Nevertheless, in general, the performance of these end-to-end models, especially attention-based models, is particularly degraded in the case of long utterances. To address this limitation, we propose adding a fully-differentiable memory-augmented neural network between the encoder and decoder of a conformer. This external memory can enrich the generalization for longer utterances since it allows the system to store and retrieve more information recurrently. Notably, we explore the neural Turing machine (NTM) that results in our proposed Conformer-NTM model architecture for ASR. Experimental results using Librispeech train-clean-100 and train-960 sets show that the proposed system outperforms the baseline conformer without memory for long utterances.

A Simple Feature Method for Prosody Rhythm Comparison

Of all components of Prosody, Rhythm has been regarded as the hardest to address, as it is utterly linked to Pitch and Intensity. Nevertheless, Rhythm is a very good indicator of a speaker's fluency in a foreign language or even of some diseases. Canonical ways to measure Rhythm, such as $\Delta C$ or $\%V$, involve a cumbersome process of segment alignment, often leading to modest and questionable results. Perceptively, however, rhythm does not sound as difficult, as humans can grasp it even when the text is not fully intelligible. In this work, we develop an empirical and unsupervised method of rhythm assessment, which does not rely on the content. We have created a fixed-length representation of each utterance, Peak Embedding (PE), which codifies the proportional distance between peaks of the chosen Low-Level Descriptors. Clustering pairs of small sentence-like units, we have attained averages of 0.444 for Silhouette Coefficient using PE with Loudness, and 0.979 for Global Separability Index with a combination of PE with Pitch and Loudness. Clustering same-structure words, we have attained averages of 0.196 for Silhouette Coefficient and 0.864 for Global Separability Index for PE with Loudness.

Privacy-preserving Automatic Speaker Diarization

Automatic Speaker Diarization (ASD) is an enabling technology with numerous applications, which deals with recordings of multiple speakers, raising special concerns in terms of privacy. In fact, in remote settings, where recordings are shared with a server, clients relinquish not only the privacy of their conversation, but also of all the information that can be inferred from their voices. However, to the best of our knowledge, the development of privacy-preserving ASD systems has been overlooked thus far. In this work, we tackle this problem using a combination of two cryptographic techniques, Secure Multiparty Computation (SMC) and Secure Modular Hashing, and apply them to the two main steps of a cascaded ASD system: speaker embedding extraction and agglomerative hierarchical clustering. Our system is able to achieve a reasonable trade-off between performance and efficiency, presenting real-time factors of 1.1 and 1.6, for two different SMC security settings.

Towards End-to-End Private Automatic Speaker Recognition

The development of privacy-preserving automatic speaker verification systems has been the focus of a number of studies with the intent of allowing users to authenticate themselves without risking the privacy of their voice. However, current privacy-preserving methods assume that the template voice representations (or speaker embeddings) used for authentication are extracted locally by the user. This poses two important issues: first, knowledge of the speaker embedding extraction model may create security and robustness liabilities for the authentication system, as this knowledge might help attackers in crafting adversarial examples able to mislead the system; second, from the point of view of a service provider the speaker embedding extraction model is arguably one of the most valuable components in the system and, as such, disclosing it would be highly undesirable. In this work, we show how speaker embeddings can be extracted while keeping both the speaker's voice and the service provider's model private, using Secure Multiparty Computation. Further, we show that it is possible to obtain reasonable trade-offs between security and computational cost. This work is complementary to those showing how authentication may be performed privately, and thus can be considered as another step towards fully private automatic speaker recognition.

Using Self-Supervised Feature Extractors with Attention for Automatic COVID-19 Detection from Speech

The ComParE 2021 COVID-19 Speech Sub-challenge provides a test-bed for the evaluation of automatic detectors of COVID-19 from speech. Such models can be of value by providing test triaging capabilities to health authorities, working alongside traditional testing methods. Herein, we leverage the usage of pre-trained, problem agnostic, speech representations and evaluate their use for this task. We compare the obtained results against a CNN architecture trained from scratch and traditional frequency-domain representations. We also evaluate the usage of Self-Attention Pooling as an utterance-level information aggregation method. Experimental results demonstrate that models trained on features extracted from self-supervised models perform similarly or outperform fully-supervised models and models based on handcrafted features. Our best model improves the Unweighted Average Recall (UAR) from 69.0\% to 72.3\% on a development set comprised of only full-band examples and achieves 64.4\% on the test set. Furthermore, we study where the network is attending, attempting to draw some conclusions regarding its explainability. In this relatively small dataset, we find the network attends especially to vowels and aspirates.

Domain Adaptation in Dialogue Systems using Transfer and Meta-Learning

Current generative-based dialogue systems are data-hungry and fail to adapt to new unseen domains when only a small amount of target data is available. Additionally, in real-world applications, most domains are underrepresented, so there is a need to create a system capable of generalizing to these domains using minimal data. In this paper, we propose a method that adapts to unseen domains by combining both transfer and meta-learning (DATML). DATML improves the previous state-of-the-art dialogue model, DiKTNet, by introducing a different learning technique: meta-learning. We use Reptile, a first-order optimization-based meta-learning algorithm as our improved training method. We evaluated our model on the MultiWOZ dataset and outperformed DiKTNet in both BLEU and Entity F1 scores when the same amount of data is available.

FoolHD: Fooling speaker identification by Highly imperceptible adversarial Disturbances

Speaker identification models are vulnerable to carefully designed adversarial perturbations of their input signals that induce misclassification. In this work, we propose a white-box steganography-inspired adversarial attack that generates imperceptible adversarial perturbations against a speaker identification model. Our approach, FoolHD, uses a Gated Convolutional Autoencoder that operates in the DCT domain and is trained with a multi-objective loss function, in order to generate and conceal the adversarial perturbation within the original audio files. In addition to hindering speaker identification performance, this multi-objective loss accounts for human perception through a frame-wise cosine similarity between MFCC feature vectors extracted from the original and adversarial audio files. We validate the effectiveness of FoolHD with a 250-speaker identification x-vector network, trained using VoxCeleb, in terms of accuracy, success rate, and imperceptibility. Our results show that FoolHD generates highly imperceptible adversarial audio files (average PESQ scores above 4.30), while achieving a success rate of 99.6% and 99.2% in misleading the speaker identification model, for untargeted and targeted settings, respectively.

Pathological speech detection using x-vector embeddings

The potential of speech as a non-invasive biomarker to assess a speaker's health has been repeatedly supported by the results of multiple works, for both physical and psychological conditions. Traditional systems for speech-based disease classification have focused on carefully designed knowledge-based features. However, these features may not represent the disease's full symptomatology, and may even overlook its more subtle manifestations. This has prompted researchers to move in the direction of general speaker representations that inherently model symptoms, such as Gaussian Supervectors, i-vectors and, x-vectors. In this work, we focus on the latter, to assess their applicability as a general feature extraction method to the detection of Parkinson's disease (PD) and obstructive sleep apnea (OSA). We test our approach against knowledge-based features and i-vectors, and report results for two European Portuguese corpora, for OSA and PD, as well as for an additional Spanish corpus for PD. Both x-vector and i-vector models were trained with an out-of-domain European Portuguese corpus. Our results show that x-vectors are able to perform better than knowledge-based features in same-language corpora. Moreover, while x-vectors performed similarly to i-vectors in matched conditions, they significantly outperform them when domain-mismatch occurs.

Attentive Filtering Networks for Audio Replay Attack Detection

An attacker may use a variety of techniques to fool an automatic speaker verification system into accepting them as a genuine user. Anti-spoofing methods meanwhile aim to make the system robust against such attacks. The ASVspoof 2017 Challenge focused specifically on replay attacks, with the intention of measuring the limits of replay attack detection as well as developing countermeasures against them. In this work, we propose our replay attacks detection system - Attentive Filtering Network, which is composed of an attention-based filtering mechanism that enhances feature representations in both the frequency and time domains, and a ResNet-based classifier. We show that the network enables us to visualize the automatically acquired feature representations that are helpful for spoofing detection. Attentive Filtering Network attains an evaluation EER of 8.99$\%$ on the ASVspoof 2017 Version 2.0 dataset. With system fusion, our best system further obtains a 30$\%$ relative improvement over the ASVspoof 2017 enhanced baseline system.