| 研究課題/領域番号 |
20J20580
|
|---|
| 研究機関 | 北陸先端科学技術大学院大学 |
|---|
研究代表者 |
MAWALIM CANDY OLIVIA 北陸先端科学技術大学院大学, 先端科学技術研究科, 特別研究員(DC1)
|
|---|
| 研究期間 (年度) |
2020-04-24 – 2023-03-31
|
| キーワード | information hiding / speech coding / voice privacy / speaker anonymization |
|---|
| 研究実績の概要 |
In the first year, the investigation on audio information hiding (AIH) for secure speech communication systems was conducted. AIH often lacks robustness in dealing with speech codecs, which are efficiently used in a speech communication system. Accordingly, the experiment on a code-excited linear prediction (CELP) codec was carried out to analyze the robust features for AIH. The principal investigator found out that the line spectral frequencies are robust features against the CELP codec. The results of this experiment were reported in the APSIPA proceeding 2020. Besides, the investigation of a method to secure voice privacy based on the information hiding concept was also conducted. The results were reported in the Interspeech proceeding 2020 and a journal (submitted).
|
| 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. The CELP codec was analyzed thoroughly at this stage, and some robust features were investigated as information hiding mediums. The principal investigator presented the experiments on modification of LSFs (one of the robust features) in a paper at APSIPA 2020 conference. The information hiding concept was also considered to hide speaker individuality for securing voice privacy. The technique for hiding speaker individuality is based on the VoicePrivacy Challenge 2020 (VP2020), namely the speaker anonymization approach. The principal investigator participated in the VP2020, and the results were reported in the Interspeech 2020 conference. The extended results were also submitted in the special issue of voice privacy as a journal paper.
|
| 今後の研究の推進方策 |
Despite the promising results, there were still some remaining problems in AIH using LSFs modification. The main reason is due to LSB that fragile for detection. As one of future works, the phase modulation on LSFs will be investigated to deal with this problem. For speaker anonymization, the trade-off issue between speech intelligibility and speaker verifiability in the speaker anonymization method is existing. The consideration of the current proposed method framework will be improved by modifying the model and parameters. In addition to this trade-off issue, the evaluation of speaker anonymization is still somehow inadequate. As future work, the more appropriate subjective and objective evaluation will be carried out by considering the phenomena in human auditory perception.
|
