Speech privacy protection by high-quality, invertible, and extendable speech anonymization and de-anonymization

点击翻译按钮获取中文摘要

The second year's work consists of three parts: Part 1) Based on the previous year's work, the second VoicePrivacy challenge was organized by us and other universities. We defined new evaluation frameworks and conducted solid evaluations. In addition to many findings, we found that the new baseline, which was the research outcome of the previous year, outperformed the legacy baseline. We also saw submissions that outperformed the new baseline, which indicates the advancement of the research field brought by the VoicePrivacy challenge.Part 2) Based on the framework of the voice privacy challenge, we did a deep analysis of the common approaches to generating anonymized speaker identity representation (i.e., pseudo speaker embedding). Through a large-scale experiment, we identified good strategies to choose and assign the pseudo-speaker, including random gender selection and utterance-level anonymization. We also found that a simple percentile-based pitch conversion reduced the risk against the strongest (Semi-Informed) attacker. These findings were published in a top IEEE journal.Part 3) We followed the research plan and extended the language-independent speaker anonymization framework. Although the framework is language-independent, its performance degrades when processing unseen languages. We found that using multilingual training data for the waveform generator was helpful. We also proposed a correlation-alignment-based strategy to alleviate channel mismatch. Additionally, we extended the framework to hide gender information. Both works were published in top conferences.