Research map / Large language models / Language, safety and interpretability
MGM-Omni: Scaling Omni LLMs to Personalized Long-Horizon Speech
Chengyao Wang, Zhisheng Zhong, Bohao PENG, Senqiao Yang, Yuqi Liu, Haokun GUI, Bin Xia, Jingyao Li, Bei Yu, Jiaya Jia
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 chunk-based decoding to SpeechLM for long-form speech (ablation) | MGM-Omni SpeechLM · Long-TTS-Eval EN | WER: 31.84 -> 4.98 | Improved |
| add chunk-based decoding to SpeechLM for long-form speech (ablation) | MGM-Omni SpeechLM · Long-TTS-Eval ZH | CER: 8.97 -> 5.64 | Improved |
| fuse Qwen2-Audio and Belle-Whisper encoders via information mining instead of Qwen2-Audio only (ablation) | MGM-Omni audio encoder · CommonVoice EN | WER: 13.0 -> 9.1 | Improved |
| fuse encoders via information mining instead of Belle-Whisper only (ablation) | MGM-Omni audio encoder · CommonVoice EN | WER: 21.7 -> 9.1 | Improved |
| fuse encoders via information mining instead of Qwen2-Audio only (ablation) | MGM-Omni audio encoder · CommonVoice ZH | CER: 3.9 -> 3.5 | Improved |
| increase parallel decoding size from 1 to 4 (ablation) | MGM-Omni SpeechLM · Seed-TTS-Eval EN | WER: 1.86 -> 2.28 | Hurt |
| increase parallel decoding size from 1 to 4 (ablation) | MGM-Omni SpeechLM · Long-TTS-Eval samples | RTF: 0.57 -> 0.19 | Improved |
| increase parallel decoding size from 1 to 2 (ablation) | MGM-Omni SpeechLM · Seed-TTS-Eval EN | WER: 1.86 -> 2.02 | Hurt |
| scale SpeechLM from 0.6B to 4B parameters | MGM-Omni-TTS · Seed-TTS-Eval EN | WER: 2.48 -> 2.22 | Improved |
| scale MLLM from 7B to 32B | MGM-Omni · LibriSpeech test-other | WER: 3.6 -> 3.2 | Improved |
| scale MLLM from 7B to 32B | MGM-Omni · TextVQA-Speech | accuracy: 81.7 -> 78.2 | Hurt |
Closest papers by meaning
- OpenOmni: Advancing Open-Source Omnimodal Large Language Models with Progressive Multimodal Alignment and Real-time Emotional Speech Synthesis (NeurIPS 2025 · similarity 0.87)
- UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities (NeurIPS 2026 · similarity 0.87)
- LLaMA-Omni: Seamless Speech Interaction with Large Language Models (ICLR 2025 · similarity 0.87)
- SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation (NeurIPS 2025 · similarity 0.87)
- VITA-1.5: Towards GPT-4o Level Real-Time Vision and Speech Interaction (NeurIPS 2025 · similarity 0.87)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.86)
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.
