Abstract
Far-field speaker verification (SV) has always been critical but challenging. Data augmentation is commonly used to overcome the problems arising from far-field microphones, such as high background noise levels and reverberation effects. On top of data augmentation, this paper tackles these problems by introducing a UNet-based speech enhancement (SE) module as a front-end processor for the speaker embedding module. To prevent the SE module from distorting speaker information, we propose two improvements to the speech enhancement-speaker embedding pipeline. (1) A UNet-DenseNet joint training scheme in which the UNet is optimized by both the MSE and speaker classification losses. (2) A semi-joint training scheme that stops the UNet training but continues the DenseNet training when overfitting of the UNet is detected. Extensive experiments on noise-contaminated Voxceleb1 and the VOiCES Challenge 2019 demonstrate the effectiveness of the two training schemes.
Original language | English |
---|---|
Pages (from-to) | 3714-3718 |
Number of pages | 5 |
Journal | Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH |
Volume | 2022-September |
DOIs | |
Publication status | Published - Sept 2022 |
Event | 23rd Annual Conference of the International Speech Communication Association, INTERSPEECH 2022 - Incheon, Korea, Republic of Duration: 18 Sept 2022 → 22 Sept 2022 |
Keywords
- DenseNet
- Far-field speaker verification
- speaker embedding
- speech enhancement
- UNet
ASJC Scopus subject areas
- Language and Linguistics
- Human-Computer Interaction
- Signal Processing
- Software
- Modelling and Simulation