DSB-DG evaluates voice agents answering spoken questions from professional documents. Its 1,636 verified question-answer pairs cover 50 documents in five domains and test increasing context length, retention across dialogue turns, context re-injection, hallucination, and latency.
The cascaded ASR-LLM-TTS pipeline achieved the highest grounding accuracy, with Gemini-Live and GPT-Realtime close behind. Open-weight speech-to-speech systems showed abrupt context-capacity collapse and stronger multi-turn grounding decay. Across architectures, greater context and conversational load reduced factual fidelity, and systems often generated unsupported answers rather than abstaining.