Meng, Wenjun (2024) Beyond Adult Speech: Exploring SepFormer’s Performance in Child Speech Separation. Master thesis, Voice Technology (VT).
|
PDF
MasterThesisWenjunMeng.pdf Download (591kB) | Preview |
Abstract
This thesis investigates the performance of SepFormer, a state-of-the-art speech separation model, in processing child speech, which has been less explored compared to adult speech. The study aims to evaluate the effectiveness of SepFormer in separating speech in datasets comprising child speech, with the hypothesis that SepFormer’s performance will significantly decline due to the unique acous- tic properties of child speech. The research utilizes the PhonBank database and employs evaluation metrics such as Scale-Invariant Signal-to-Noise Ratio and Signal-to-Distortion Ratio to assess per- formance. The findings are expected to highlight the need for recalibrating existing models or devel- oping child-specific speech separation models. This investigation is crucial for advancing automatic speech recognition systems, ensuring they are inclusive and effective in educational and commu- nicative contexts for children. The research package is made available on Github.
| Item Type: | Thesis (Master) |
|---|---|
| Name supervisor: | Coler, M.L. |
| Date Deposited: | 17 Jul 2024 08:18 |
| Last Modified: | 17 Jul 2024 08:18 |
| URI: | https://campus-fryslan.studenttheses.ub.rug.nl/id/eprint/528 |
Actions (login required)
 |
View Item |