{"paper":{"title":"VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"A multi-stage training method allows large language models to handle vision and speech together for near real-time interaction.","cross_cats":["cs.SD","eess.AS"],"primary_cat":"cs.CV","authors_text":"Caifeng Shan, Chaoyou Fu, Haojia Lin, Haoyu Cao, Heting Gao, Ke Li, Long Ma, Ran He, Rongrong Ji, Xiaoyu Liu, Xiawu Zheng, Xing Sun, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Zuwei Long","submitted_at":"2025-01-03T18:59:52Z","abstract_excerpt":"Recent Multimodal Large Language Models (MLLMs) have typically focused on integrating visual and textual modalities, with less emphasis placed on the role of speech in enhancing interaction. However, speech plays a crucial role in multimodal dialogue systems, and implementing high-performance in both vision and speech tasks remains a significant challenge due to the fundamental modality differences. In this paper, we propose a carefully designed multi-stage training methodology that progressively trains LLM to understand both visual and speech information, ultimately enabling fluent vision and"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By comparing our method against state-of-the-art counterparts across benchmarks for image, video, and speech tasks, we demonstrate that our model is equipped with both strong visual and speech capabilities, making near real-time vision and speech interaction.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The multi-stage training can be balanced so that speech capabilities are added without degrading the pre-existing vision-language capacity, an assumption stated in the description of the progressive training methodology.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"VITA-1.5 integrates vision and speech into a single LLM through multi-stage training, delivering competitive benchmark results on image, video, and speech tasks with near real-time response speed.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A multi-stage training method allows large language models to handle vision and speech together for near real-time interaction.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e8c6731de283f6b18da3eeb7bc841fc5dddec0485d965e8d46b1193eb80b5464"},"source":{"id":"2501.01957","kind":"arxiv","version":4},"verdict":{"id":"76a7ab1c-81c8-4eb8-9f5c-10000558b2e6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T21:03:49.526514Z","strongest_claim":"By comparing our method against state-of-the-art counterparts across benchmarks for image, video, and speech tasks, we demonstrate that our model is equipped with both strong visual and speech capabilities, making near real-time vision and speech interaction.","one_line_summary":"VITA-1.5 integrates vision and speech into a single LLM through multi-stage training, delivering competitive benchmark results on image, video, and speech tasks with near real-time response speed.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The multi-stage training can be balanced so that speech capabilities are added without degrading the pre-existing vision-language capacity, an assumption stated in the description of the progressive training methodology.","pith_extraction_headline":"A multi-stage training method allows large language models to handle vision and speech together for near real-time interaction."},"references":{"count":77,"sample":[{"doi":"","year":2023,"title":"Visual Instruction Tuning","work_id":"68be622d-a6dc-4a13-82de-e3054a3dc509","ref_index":1,"cited_arxiv_id":"2304.08485","is_internal_anchor":true},{"doi":"","year":2024,"title":"AnyGPT: Unified Multimodal LLM with Discrete Sequence Modeling","work_id":"647f3087-e96f-4aaa-a599-2dbafb5ac024","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding","work_id":"555cf04a-49a7-44b8-9019-a83ce85ace95","ref_index":3,"cited_arxiv_id":"2306.02858","is_internal_anchor":true},{"doi":"","year":2025,"title":"Speechact: Towards generating whole-body motion from speech.IEEE TVCG","work_id":"5c873038-1603-44c8-b3a6-34e24f353959","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"video-salmonn: Speech-enhanced audio-visual large language models","work_id":"14b6b6ca-3c9e-434a-8eb7-ebdcbf08efc6","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":77,"snapshot_sha256":"d9c993562d3e8372ceb4bee179753bc295ea9a4e1ab57f406c8829e487e9d3c9","internal_anchors":28},"formal_canon":{"evidence_count":2,"snapshot_sha256":"56bf5d2546ea0fcddd50ec47bfd0c6b37324b79cc0badd7ccd93ee39d25c0cf4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}