Research map / Large language models / Retrieval and knowledge
Stream RAG: Instant and Accurate Spoken Dialogue Systems with Streaming Tool Usage
Siddhant Arora, Haidar Khan, Kai Sun, Xin Luna Dong, Sajal Choudhary, Seungwhan Moon, Xinyuan Zhang, Adithya Sagar, Surya Teja Appini, Kaushik Patnaik, Sanat Sharma, Shinji Watanabe and 5 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add web and KG tool retrieval (open book) to speech-in speech-out model | Qwen-OMNI · AudioCRAG-Synthetic | accuracy %: 11.1 -> 26.3 | Improved |
| add tool retrieval (open book) | Qwen-OMNI · AudioCRAG-Synthetic | first-token latency (s): 1.34 -> 5.90 | Hurt |
| add Model-Triggered Stream RAG post-training and streaming queries over open book | Qwen-OMNI · AudioCRAG-Synthetic | accuracy %: 26.3 -> 34.2 | Improved |
| add Model-Triggered Stream RAG over open book | Qwen-OMNI · AudioCRAG-Human | tool use latency (s): 3.37 -> 1.57 | Improved |
| serve Model-Triggered Stream RAG with vLLM | Qwen-OMNI · AudioCRAG-Human | median first-token latency (s): 3.16 -> 1.36 | Improved |
| post-train sequential RAG instead of streaming RAG (ablation) | Qwen-OMNI · AudioCRAG-Synthetic | accuracy %: 34.9 vs 34.2 | No clear effect |
| post-train with streaming tool query generation instead of sequential tool integration (ablation) | OpusLM · AudioCRAG-Synthetic | accuracy %: 22.4 -> 23.6 | No clear effect |
| use self-cascade (text then speech) for speech output (ablation) | Qwen-OMNI open book · AudioCRAG-Synthetic | accuracy %: 26.3 -> 33.8 | Improved |
| use model-triggered instead of fixed-interval trigger (ablation) | Stream RAG · AudioCRAG | parallel tool threads and latency savings | Improved |
| use fixed-interval Stream RAG without post-training (ablation) | Qwen-OMNI open book · AudioCRAG | tool use latency: 10.7% reduction | Improved |
| use Stream RAG instead of cascaded pipeline | Qwen-OMNI · SLUE-SQA | exact match accuracy: 36.6 -> 55.8 | Improved |
| add speech output instead of text output with Stream RAG | Qwen-OMNI · AudioCRAG-Synthetic | accuracy %: 39.8 -> 34.2 | Hurt |
Closest papers by meaning
- ToolDial: Multi-turn Dialogue Generation Method for Tool-Augmented Language Models (ICLR 2025 · similarity 0.86)
- CoVoMix2: Advancing Zero-Shot Dialogue Generation with Fully Non-Autoregressive Flow Matching (NeurIPS 2025 · similarity 0.85)
- Streaming Video Understanding and Multi-round Interaction with Memory-enhanced Knowledge (ICLR 2025 · similarity 0.85)
- ChatQA: Surpassing GPT-4 on Conversational QA and RAG (NeurIPS 2024 · similarity 0.84)
- PhoStream: Benchmarking Real-World Streaming for Omnimodal Assistants in Mobile Scenarios (ICML 2026 · similarity 0.84)
- CRAG - Comprehensive RAG Benchmark (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
