Soundwave: Less is More for Speech-Text Alignment in LLMs
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_ | 1866910834726273024 |
|---|---|
| author | Zhang, Yuhao Liu, Zhiheng Bu, Fan Zhang, Ruiyu Wang, Benyou Li, Haizhou |
| author_facet | Zhang, Yuhao Liu, Zhiheng Bu, Fan Zhang, Ruiyu Wang, Benyou Li, Haizhou |
| contents | Existing end-to-end speech large language models (LLMs) usually rely on large-scale annotated data for training, while data-efficient training has not been discussed in depth. We focus on two fundamental problems between speech and text: the representation space gap and sequence length inconsistency. We propose Soundwave, which utilizes an efficient training strategy and a novel architecture to address these issues. Results show that Soundwave outperforms the advanced Qwen2-Audio in speech translation and AIR-Bench speech tasks, using only one-fiftieth of the training data. Further analysis shows that Soundwave still retains its intelligence during conversation. The project is available at https://github.com/FreedomIntelligence/Soundwave. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_12900 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Soundwave: Less is More for Speech-Text Alignment in LLMs Zhang, Yuhao Liu, Zhiheng Bu, Fan Zhang, Ruiyu Wang, Benyou Li, Haizhou Computation and Language Artificial Intelligence Sound Existing end-to-end speech large language models (LLMs) usually rely on large-scale annotated data for training, while data-efficient training has not been discussed in depth. We focus on two fundamental problems between speech and text: the representation space gap and sequence length inconsistency. We propose Soundwave, which utilizes an efficient training strategy and a novel architecture to address these issues. Results show that Soundwave outperforms the advanced Qwen2-Audio in speech translation and AIR-Bench speech tasks, using only one-fiftieth of the training data. Further analysis shows that Soundwave still retains its intelligence during conversation. The project is available at https://github.com/FreedomIntelligence/Soundwave. |
| title | Soundwave: Less is More for Speech-Text Alignment in LLMs |
| topic | Computation and Language Artificial Intelligence Sound |
| url | https://arxiv.org/abs/2502.12900 |