Rethinking Leveraging Pre-Trained Multi-Layer Representations for Speaker Verification
Fuente:
arXiv
Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866909976644026368 |
|---|---|
| author | Kim, Jin Sob Park, Hyun Joon Shin, Wooseok Han, Sung Won |
| author_facet | Kim, Jin Sob Park, Hyun Joon Shin, Wooseok Han, Sung Won |
| contents | Recent speaker verification studies have achieved notable success by leveraging layer-wise output from pre-trained Transformer models. However, few have explored the advancements in aggregating these multi-level features beyond the static weighted average. We present Layer Attentive Pooling (LAP), a novel strategy for aggregating inter-layer representations from pre-trained speech models for speaker verification. LAP assesses the significance of each layer from multiple perspectives time-dynamically, and employs max pooling instead of averaging. Additionally, we propose a lightweight backend speaker model comprising LAP and Attentive Statistical Temporal Pooling (ASTP) to extract speaker embeddings from pre-trained model output. Experiments on the VoxCeleb benchmark reveal that our compact architecture achieves state-of-the-art performance while greatly reducing the training time. We further analyzed LAP design and its dynamic weighting mechanism for capturing speaker characteristics. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2512_22148 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Rethinking Leveraging Pre-Trained Multi-Layer Representations for Speaker Verification Kim, Jin Sob Park, Hyun Joon Shin, Wooseok Han, Sung Won Sound Artificial Intelligence Recent speaker verification studies have achieved notable success by leveraging layer-wise output from pre-trained Transformer models. However, few have explored the advancements in aggregating these multi-level features beyond the static weighted average. We present Layer Attentive Pooling (LAP), a novel strategy for aggregating inter-layer representations from pre-trained speech models for speaker verification. LAP assesses the significance of each layer from multiple perspectives time-dynamically, and employs max pooling instead of averaging. Additionally, we propose a lightweight backend speaker model comprising LAP and Attentive Statistical Temporal Pooling (ASTP) to extract speaker embeddings from pre-trained model output. Experiments on the VoxCeleb benchmark reveal that our compact architecture achieves state-of-the-art performance while greatly reducing the training time. We further analyzed LAP design and its dynamic weighting mechanism for capturing speaker characteristics. |
| title | Rethinking Leveraging Pre-Trained Multi-Layer Representations for Speaker Verification |
| topic | Sound Artificial Intelligence |
| url | https://arxiv.org/abs/2512.22148 |