From Awareness to Adherence: Bridging the Context Gap in Spoken Dialogue Systems via Context-Aware Decoding

The article introduces a novel Context-Aware Decoding (CAD) approach for improving adherence to context in end-to-end spoken dialogue systems. By utilizing internal attention mechanisms to prioritize relevant historical dialogue during decoding, the method enhances multimodal context signals, leading to improved performance on the Audio MultiChallenge benchmark, particularly in Semantic Memory and Self Coherence tasks. This advancement is crucial for practitioners as it addresses the challenge of context retention in multi-turn conversations, potentially leading to more coherent and contextually aware dialogue systems.