Lightweight Voice Anonymization Based on Data-Driven Optimization of Cascaded Voice Modification Modules

Kai, Hiroto; Takamichi, Shinnosuke; Shiota, Sayaka; Kiya, Hitoshi

doi:10.1109/slt48900.2021.9383535

Cited by 11 publications

(10 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For ASV, we train the xvector model [19] with an AMSoftmax loss [20]. Following prior work [21,3,4], we report the equal error rate (EER) as our privacy metric (high EER implies good privacy-preserving representations). IEMOCAP (ER).…”

Section: Tasks Datasets and Modelsmentioning

confidence: 99%

“…Existing literature on local speech anonymization generally follows two lines: voice conversion to change the speaker embeddings while retaining the content (i.e., transforming the source voice to a target voice or target speaker) [25,3,26,27] or voice modification (i.e., apply different signal processing techniques directly to the source voice to get a perturbed output voice) [4,28,5]. In this study, we select two voice conversion methods, namely the Cycle-GAN [29] and the Assem-VC [30], as the baselines for the first line of anonymization approach.…”

Section: Baseline Modelsmentioning

confidence: 99%

“…In this study, we select two voice conversion methods, namely the Cycle-GAN [29] and the Assem-VC [30], as the baselines for the first line of anonymization approach. For the second line, we use the signal processing-based methods from Kai et al [4] as our baseline.…”

Section: Baseline Modelsmentioning

confidence: 99%

“…Several methods attempting to anonymize speakers while preserving the contents have been proposed, including voice modification [4,5], voice conversion [6], and adversarial learning [7]. The majority of these methods rely on sharing data in the form of audio, which is not required for machine learning.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Towards Privacy-Preserving Speech Representation for Client-Side Data Sharing

Tran¹,

Soleymani²

2022

Preprint

View full text Add to dashboard Cite

Privacy and security are major concerns when sharing and collecting speech data for cloud services such as automatic speech recognition (ASR) and speech emotion recognition (SER). Existing solutions for client-side privacy mainly focus on voice conversion or voice modification to convert a raw utterance into another one with similar content but different, or no, identity-related information. However, an alternative approach to share speech data under the form of privacypreserving representations has been largely under-explored. To fill this gap, we propose a speech anonymization framework that provides formal privacy guarantees via noise perturbation to a selected subset of the high-utility representations extracted using a pre-trained speech encoder. The subset is chosen with a Transformer-based privacy-risk saliency estimator. We validate our framework on four tasks, namely, Automatic Speaker Verification (ASV), ASR, SER and Intent Classification (IC) for privacy and utility assessment. Experimental results show that our approach is able to achieve a competitive, or even better, utility compared to the baselines that use voice conversion and voice modification, providing the same level of privacy. Moreover, the easily-controlled amount of perturbation allows our framework to have a flexible range of privacy-utility trade-offs without re-training any components.

show abstract

Section: Tasks Datasets and Modelsmentioning

confidence: 99%

Section: Baseline Modelsmentioning

confidence: 99%

Section: Baseline Modelsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Towards Privacy-Preserving Speech Representation for Client-Side Data Sharing

Tran¹,

Soleymani²

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…The privacy is often achieved at the expense of utility, and an important question is how to set up a proper threshold between privacy and utility (Li & Li, 2009). When developing anonymization methods, a joint optimization of utility gain and privacy loss can be performed by incorporating them into the criterion for training anonymization models (Kai et al, 2021).…”

Section: Open Questions and Future Directionsmentioning

confidence: 99%

The VoicePrivacy 2020 Challenge: Results and findings

Tomashenko,

Wang,

Vincent

et al. 2021

Preprint

View full text Add to dashboard Cite

This paper presents the results and analyses stemming from the first VoicePrivacy 2020 Challenge which focuses on developing anonymization solutions for speech technology. We provide a systematic overview of the challenge design with an analysis of submitted systems and evaluation results. In particular, we describe the voice anonymization task and datasets used for system development and evaluation. Also, we present different attack models and the associated objective and subjective evaluation metrics. We introduce two anonymization baselines and provide a summary description of the anonymization systems developed by the challenge participants. We report objective and subjective evaluation results for baseline and submitted systems. In addition, we present experimental results for alternative privacy metrics and attack models developed as a part of the post-evaluation analysis. Finally, we summarise our insights and observations that will influence the design of the next VoicePrivacy challenge edition and some directions for future voice anonymization research.

show abstract