Academic literature on the topic 'Speaker anonymization'

Sharing real-world speech utterances is key to the training and deployment of voice-based services. However, it also raises privacy risks as speech contains a wealth of personal data. Speaker anonymization aims to remove speaker information from a speech utterance while leaving its linguistic and prosodic attributes intact. State-of-the-art techniques operate by disentangling the speaker information (represented via a speaker embedding) from these attributes and re-synthesizing speech based on the speaker embedding of another speaker. Prior research in the privacy community has shown that anonymization often provides brittle privacy protection, even less so any provable guarantee. In this work, we show that disentanglement is indeed not perfect: linguistic and prosodic attributes still contain speaker information. We remove speaker information from these attributes by introducing differentially private feature extractors based on an autoencoder and an automatic speech recognizer, respectively, trained using noise layers. We plug these extractors in the state-of-the-art anonymization pipeline and generate, for the first time, private speech utterances with a provable upper bound on the speaker information they contain. We evaluate empirically the privacy and utility resulting from our differentially private speaker anonymization approach on the LibriSpeech data set. Experimental results show that the generated utterances retain very high utility for automatic speech recognition training and inference, while being much better protected against strong adversaries who leverage the full knowledge of the anonymization process to try to infer the speaker identity.

Speech is a crucial source of personal information, and the risk of attackers using such information increases day by day. Speaker privacy protection is crucial, and various approaches have been proposed to hide the speaker’s identity. One approach is voice anonymization, which aims to safeguard speaker identity while maintaining speech content through techniques such as voice conversion or spectral feature alteration. The significance of voice anonymization has grown due to the necessity to protect personal information in applications such as voice assistants, authentication, and customer support. Building upon the S3PRL-VC toolkit and on pre-trained speech and speaker representation models, this paper introduces a feature disentanglement approach to improve the de-identification performance of the state-of-the-art anonymization approaches based on voice conversion. The proposed approach achieves state-of-the-art speaker de-identification and causes minimal impact on the intelligibility of the signal after conversion.

This study reports the results of focus groups with school nurses and teachers from elementary, middle, and high schools to explore their perceptions of child and adolescent oral health. Participants included 14 school nurses and 15 teachers (83% female; 31% Hispanic; 21% White; 21% Asian; 14% African American; and 13% Others). Respondents were recruited from Los Angeles County schools and scheduled by school level for six one-hour focus groups using Zoom. Audio recordings were transcribed, reviewed, and saved with anonymization of speaker identities. NVivo software (QSR International, Melbourne, Australia) was used to facilitate content analysis and identify key themes. The nurses’ rate of “Oral Health Education” comments statistically exceeded that of teachers, while teachers had higher rates for “Parental Involvement” and “Mutual Perception” comments. “Need for Care” was perceived to be more prevalent in immigrants to the United States based on student behaviors and complaints. “Access to Care” was seen as primarily the nurses’ responsibilities. Strong relationships between community clinics and schools were viewed by some as integral to students achieving good oral health. The results suggest dimensions and questions important to item development for oral health surveys of children and parents to address screening, management, program assessment, and policy planning.

The advent use of smart devices has enabled the emergence of many applications that facilitate user interaction through speech. However, speech reveals private and sensitive information about the user’s identity, posing several security risks. For example, a speaker’s speech can be acquired and used in speech synthesis systems to generate fake speech recordings that can be used to attack that speaker’s verification system. One solution is to anonymize the speaker’s identity from speech before using it. Existing anonymization schemes rely on using a pool of real speakers’ identities for anonymization, which may result in associating a speaker’s speech with an existing speaker. Hence, this paper investigates the use of Generative Adversarial Networks (GAN) to generate a pool of fake identities that are used for anonymization. Several GAN types were considered for this purpose, and the Conditional Tabular GAN (CTGAN) showed the best performance among all GAN types according to different metrics that measure the naturalness of the anonymized speech and its linguistic content.

L'essor de l'utilisation d'assistants vocaux, présents dans les téléphones, automobiles et autres, a augmenté la quantité de données de parole collectées et stockées. Bien que cette collecte de données soit cruciale pour entrainer les modèles qui traitent la parole, cette collecte soulève également des préoccupations de protection de la vie privée. Des technologies de pointe traitant la parole, telles que le clonage vocal et la reconnaissance d'attributs personnels (telles que l'identité, l'émotion, l'âge, le genre, etc.), peuvent être exploitées pour accéder et utiliser des informations personnelles. Par exemple, un malfaiteur pourrait utiliser le clonage vocal pour se faire passer pour une autre personne afin d'obtenir un accès non autorisé à ses informations bancaires par téléphone. Avec l'adoption croissante des assistants vocaux tels qu'Alexa, Google Assistant et Siri, et la facilité avec laquelle les données peuvent être collectées et stockées, le risque d'utilisation abusive de technologies telles que le clonage vocal et la reconnaissance d'attributs personnels augmente. Il est donc important pour les entreprises et les organisations de prendre en compte ces risques et de mettre en place des mesures appropriées pour protéger les données des utilisateurs, en conformité avec les réglementations juridiques telles que le Règlement Général sur la Protection des Données (RGPD). Pour répondre aux enjeux liés à la protection de la vie privée, cette thèse propose des solutions permettant d'anonymiser la parole. L'anonymisation désigne ici le processus consistant à rendre les signaux de parole non associables à une identité spécifique, tout en préservant leur utilité, c'est-à-dire ne pas modifier le contenu linguistique du message. L'objectif est de préserver la vie privée des individus en éliminant ou en rendant floues toutes les informations personnellement identifiables (PPI) contenues dans le signal acoustique, telles que l'accent ou le style de parole d'une personne. Les informations linguistiques personnelles telles que numéros de téléphone ou noms de personnes ne font pas partie du champ d'étude de cette thèse. Notre recherche s'appuie sur les méthodes d'anonymisation existantes basées sur la conversion de la voix et sur des protocoles d'évaluation existants. Nous commençons par identifier et expliquer plusieurs défis auxquels les protocoles d'évaluation doivent faire face afin d'évaluer de manière précise le niveau de protection de la vie privée. Nous clarifions comment les systèmes d'anonymisation doivent être configurés pour être correctement évalués, en soulignant le fait que de nombreuses configurations ne permettent pas une évaluation adéquate de non-asociabilité d'un signal a une identité. Nous étudions et examinons également le système d'anonymisation basé sur la conversion de la voix le plus courant, identifions ses points faibles, et proposons de nouvelles méthodes pour en améliorer les performances. Nous avons isolé tous les composants du système d'anonymisation afin d'évaluer le niveau de PPI encodé par chaque composant. Ensuite, nous proposons plusieurs méthodes de transformation de ces composants dans le but de réduire autant que possible les PPI encodées, tout en maintenant l'utilité. Nous promouvons les algorithmes d'anonymisation basés sur l'utilisation de la quantification en alternative à la méthode la plus utilisée et la plus connue basée sur le bruit. Enfin, nous proposons une nouvelle méthode d'évaluation qui vise à inverser l'anonymisation, créant ainsi une nouvelle manière d'étudier les systèmes d'anonymisation<br>The growing use of voice user interfaces, from telephones to remote controls, automobiles, and digital assistants, has led to a surge in the collection and storage of speech data. While data collection allows for the development of efficient tools powering most speech services, it also poses serious privacy issues for users as centralized storage makes private personal speech data vulnerable to cyber threats. Advanced speech technologies, such as voice-cloning and personal attribute recognition, can be used to access and exploit sensitive information. Voice-cloning technology allows an attacker to take a recording of a person's voice and use it to generate new speech that sounds like it is coming from that person. For example, an attacker could use voice-cloning to impersonate a person's voice to gain unauthorized access to his/her financial information over the phone. With the increasing use of voice-based digital assistants like Amazon's Alexa, Google's Assistant, and Apple's Siri, and with the increasing ease with which personal speech data can be collected and stored, the risk of malicious use of voice-cloning and speaker/gender/pathological/etc. recognition technologies have increased. Companies and organizations need to consider these risks and implement appropriate measures to protect user data in order to prevent misuse of speech technologies and comply with legal regulations (e.g., General Data Protection Regulation (GDPR)). To address these concerns, this thesis proposes solutions for anonymizing speech and evaluating the degree of the anonymization. In this work, anonymization refers to the process of making personal speech data unlinkable to an identity, while maintaining the usefulness (utility) of the speech signal (e.g., access to the linguistic content). The goal is to protect the privacy of individuals by removing or obscuring any Personally Identifiable Information (PPI) from the acoustic of speech. PPI includes things like a person's voice, accent, and speaking style; other personal information in the speech content like, phone number, person name, etc., is out of the scope of this thesis. Our research is built on top of existing anonymization methods based on voice conversion and existing evaluation protocols. We start by identifying and explaining several challenges that evaluation protocols need to consider to evaluate the degree of privacy protection properly. We clarify how anonymization systems need to be configured for evaluation purposes and highlight the fact that many practical deployment configurations do not permit privacy evaluation. Furthermore, we study and examine the most common voice conversion-based anonymization system and identify its weak points, before suggesting new methods to overcome some limitations. We isolate all components of the anonymization system to evaluate the degree of speaker PPI associated with each of them. Then, we propose several transformation methods for each component to reduce as much as possible speaker PPI while maintaining utility. We promote anonymization algorithms based on quantization-based transformation as an alternative to the most-used and well-known noise-based approach. Finally, we endeavor a new attack method to invert the anonymization, creating a new threat. In this thesis, we openly work on sharing anonymization systems and evaluation protocols to aid organizations in facilitating the preservation of privacy rights for individuals

L'émergence et la généralisation des interfaces vocales présentesdans les téléphones, les applications mobiles et les assistantsnumériques ont permis de faciliter la communication entre les citoyens,utilisateurs d'un service, et les prestataires de services. Citons àtitre d'exemple l'utilisation de mots de passe vocaux pour lesopérations bancaires, des haut-parleurs intelligents personnalisés, etc.Pour réaliser ces innovations, la collecte massive de données vocalesest essentielle aux entreprises comme aux chercheurs. Mais le stockagecentralisé à grande échelle des données vocales pose de graves menaces àla vie privée des locuteurs. En effet, le stockage centralisé estvulnérable aux menaces de cybersécurité qui, lorsqu'elles sont combinéesavec des technologies vocales avancées telles que le clonage vocal, lareconnaissance du locuteur et l'usurpation d'identité peuvent conférer àune entité malveillante la capacité de ré-identifier les locuteurs et devioler leur vie privée en accédant à leurs caractéristiques biométriquessensibles, leurs états émotionnels, leurs attributs de personnalité,leurs conditions pathologiques, etc.Les individus et les membres de la société civile du monde entier, etparticulièrement en Europe, prennent conscience de cette menace. Avecl'entrée en vigueur du règlement général sur la protection des données(RGPD), plusieurs initiatives sont lancées, notamment la publication delivres blancs et de lignes directrices, pour sensibiliser les masses etréguler les données vocales afin que la vie privée des citoyens soitprotégée.Cette thèse constitue un effort pour soutenir de telles initiatives etpropose des solutions pour supprimer l'identité biométrique deslocuteurs des signaux de parole, les rendant ainsi inutiles pourré-identifier les locuteurs qui les ont prononcés.Outre l'objectif de protéger l'identité du locuteur contre les accèsmalveillants, cette thèse vise à explorer les solutions qui le font sansdégrader l'utilité de la parole.Nous présentons plusieurs schémas d'anonymisation basés sur des méthodesde conversion vocale pour atteindre ce double objectif. La sortie detels schémas est un signal vocal de haute qualité qui est utilisablepour la publication et pour un ensemble de tâches en aval.Tous les schémas sont soumis à un protocole d'évaluation rigoureux quiest l'un des apports majeurs de cette thèse.Ce protocole a conduit à la découverte que les approches existantes neprotègent pas efficacement la vie privée et a ainsi directement inspirél'initiative VoicePrivacy qui rassemble les individus, l'industrie et lacommunauté scientifique pour participer à la construction d'un schémad'anonymisation robuste.Nous introduisons une gamme de schémas d'anonymisation dans le cadre del'initiative VoicePrivacy et prouvons empiriquement leur supériorité entermes de protection de la vie privée et d'utilité.Enfin, nous nous efforçons de supprimer l'identité résiduelle dulocuteur du signal de parole anonymisé en utilisant les techniquesinspirées de la confidentialité différentielle. De telles techniquesfournissent des garanties analytiques démontrables aux schémasd'anonymisation proposés et ouvrent des portes pour de futures recherches.En pratique, les outils développés dans cette thèse sont un élémentessentiel pour établir la confiance dans tout écosystème logiciel où lesdonnées vocales sont stockées, transmises, traitées ou publiées. Ilsvisent à aider les organisations à se conformer aux règles mandatées parles gouvernements et à donner le choix aux individus qui souhaitentexercer leur droit à la vie privée<br>Large-scale centralized storage of speech data poses severe privacy threats to the speakers. Indeed, the emergence and widespread usage of voice interfaces starting from telephone to mobile applications, and now digital assistants have enabled easier communication between the customers and the service providers. Massive speech data collection allows its users, for instance researchers, to develop tools for human convenience, like voice passwords for banking, personalized smart speakers, etc. However, centralized storage is vulnerable to cybersecurity threats which, when combined with advanced speech technologies like voice cloning, speaker recognition, and spoofing, may endow a malicious entity with the capability to re-identify speakers and breach their privacy by gaining access to their sensitive biometric characteristics, emotional states, personality attributes, pathological conditions, etc.Individuals and the members of civil society worldwide, and especially in Europe, are getting aware of this threat. With firm backing by the GDPR, several initiatives are being launched, including the publication of white papers and guidelines, to spread mass awareness and to regulate voice data so that the citizens' privacy is protected.This thesis is a timely effort to bolster such initiatives and propose solutions to remove the biometric identity of speakers from speech signals, thereby rendering them useless for re-identifying the speakers who spoke them.Besides the goal of protecting the speaker's identity from malicious access, this thesis aims to explore the solutions which do so without degrading the usefulness of speech.We present several anonymization schemes based on voice conversion methods to achieve this two-fold objective. The output of such schemes is a high-quality speech signal that is usable for publication and a variety of downstream tasks.All the schemes are subjected to a rigorous evaluation protocol which is one of the major contributions of this thesis.This protocol led to the finding that the previous approaches do not effectively protect the privacy and thereby directly inspired the VoicePrivacy initiative which is an effort to gather individuals, industry, and the scientific community to participate in building a robust anonymization scheme.We introduce a range of anonymization schemes under the purview of the VoicePrivacy initiative and empirically prove their superiority in terms of privacy protection and utility.Finally, we endeavor to remove the residual speaker identity from the anonymized speech signal using the techniques inspired by differential privacy. Such techniques provide provable analytical guarantees to the proposed anonymization schemes and open up promising perspectives for future research.In practice, the tools developed in this thesis are an essential component to build trust in any software ecosystem where voice data is stored, transmitted, processed, or published. They aim to help the organizations to comply with the rules mandated by civil governments and give a choice to individuals who wish to exercise their right to privacy