| 研究実績の概要 |
The major milestone in FY2021 is developing a framework to improve the security of speaker anonymization. Speaker anonymization aims to address the voice privacy issue by suppressing the original speaker's personally identified information (PII). The output anonymized speech should be able to authenticate by the authorized parties. However, since the mapping between speaker and pseudo-speaker is not necessarily one-to-one correspondence, recognizing genuine anonymized speech is difficult. To deal with this issue, the proposed framework integrates the information hiding approach to simultaneously secure PII and verify the content via an embedded watermark. The related publications consist of one international conference and two journals.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
The progress of this study is going well as planned. At this stage, the proposed framework has been developed by integrating the information hiding approach to protecting content and securing speaker individuality information. It consists of an encoder and a decoder. The encoder aims to protect the speaker's identity by using an anonymization approach while embedding a parameter that represents a watermark. The decoder seeks to protect the authentication of the speech by accurately detecting the embedded watermarks. An extensive evaluation has been conducted to validate the proposed framework's performance compared to the existing methods. The results of this study in FY2021 were reported in APSIPA Proceeding 2021, MDPI Entropy Journal 2021, and Computer Speech and Language Journal 2022.
|
| 今後の研究の推進方策 |
In future work, the remaining issues, especially those related to subjective and objective evaluations for intelligibility and naturalness requirements, will be addressed. The results obtained by using existing objective evaluations could give general information about a speaker anonymization method, but it is still inadequate to show the significance of each method. Besides, x-vector-based information hiding and the investigation of other prospective speech features will be considered. By controlling the less significant eigenstructure of the x-vector, we expect better protection for speech signals. Finally, the workflow for the real application will be investigated for speech tampering and spoofing countermeasure.
|
