Context Projector: Complementary Keyword and Dialogue Context Embeddings for LLM-based ASR
| Type of publication: | Conference paper |
| Citation: | Villatoro-Tello_INTERSPEECH2026_2026 |
| Publication status: | Accepted |
| Booktitle: | INTERSPEECH 2026 |
| Year: | 2026 |
| Month: | September |
| Abstract: | Contact-center dialogue systems require both accurate transcripts and reliable recognition of business-critical entities. We introduce a hybrid keyword and \textit{context projector} (CP) approach for LLM-based ASR. Rather than appending raw dialogue history, CP encodes previous turns into compact contextual tokens, complemented by automatically extracted keywords. The module is trained while keeping the backbone frozen, enabling parameter-efficient adaptation. Experiments on realistic, multi-domain contact-center data show that naive raw-context prompting can degrade performance, whereas the proposed approach improves the tradeoff between overall and bias-word error, achieving average relative reductions of up to 2.5% overall and 7.2% on bias words. These results show our approach to be an effective and practical enhancement of context-aware LLM-based ASR. |
| Keywords: | bias word error rate, contact-center speech, contextual ASR, entity-aware evaluation, LLM-based ASR, spoken dialogue systems |
| Projects: |
ELOQUENCE UNIPHORE |
| Authors: | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
