| 研究実績の概要 |
Last year's work consists of three parts: 1) after analysis of original and anonymized speaker vector spaces, a new speaker (or voice) anonymization algorithm combining Householder transformation and deep learning was proposed. Through experiments on standard voice anonymization datasets, the new method was shown to provide better performance in terms of privacy protection utility (speech intelligibility and naturalness) than all our previous methods. This new work was published in a top IEEE journal. The proposed voice anonymization algorithm is also used to anonymize voice in TV broadcasting.
2) The proposed voice anonymization algorithm was applied to speech database anonymization, a new research topic that has yet to be studied in the speech field. Most speech databases contain speech recordings from real human speakers, and there is the risk that the speakers' identities in the speech databases are illegally copied and misused. The proposed voice anonymization algorithm was used to anonymize all the speakers in one database. This new work was published at a top IEEE conference.
3) With the know-how on voice anonymization in this project, the 3rd Voice Privacy Challenge is organized. New evaluation toolkits with easier-to-use APIs are created, new utility metrics on speech emotion, and new baselines using neural speech codec and general adversarial neural network are included. A workshop will be held in August 2024 to summarize the findings from the Voice Privacy Challenge.
|
