Research map / Large language models / Language, safety and interpretability
Towards True Speech-to-Speech Models Without Text Guidance
Xingjian Zhao, Zhe Xu, Luozhijie Jin, Yang Wang, Hanfu Chen, Yaozhou Jiang, Ke Chen, Ruixiao Li, Mingshu Chen, Ruiming Wang, Wenbo Zhang, Qinyuan Cheng and 3 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 modality-based layer split (separate final speech/text layers) (ablation) | MOSS-Speech pre-training (NF vs NF--NoSplit) · StoryCloze tS.C. | accuracy: 77.12 -> 77.66 | Improved |
| add modality-based layer split (ablation) | MOSS-Speech pre-training · MMLU | accuracy: 60.97 -> 62.11 | Improved |
| add frozen pre-training of text backbone (ablation) | MOSS-Speech pre-training (FP--Full vs NF) · StoryCloze tS.C. | accuracy: 77.66 -> 85.20 | Improved |
| add frozen pre-training of text backbone (ablation) | MOSS-Speech pre-training · MMLU | accuracy: 62.11 -> 66.50 | Improved |
| unfreeze only shared layers or layer-wise instead of all (ablation) | FP--Full · StoryCloze / MMLU | accuracy: tS.C. 85.20 vs 83.27; MMLU 66.50 vs 67.26 | No clear effect |
| add speech modality with split design and frozen backbone | Qwen3-8B · MMLU | accuracy: 76.60 -> 67.19 | Hurt |
| generate speech directly without text guidance | MOSS-Speech vs GLM-4-Voice · TriviaQA S->S | accuracy: 43.20 -> 28.8 | Hurt |
| generate speech directly without text guidance | MOSS-Speech vs GLM-4-Voice · LlamaQA S->S | accuracy: 65.67 -> 63.67 | No clear effect |
| make speech encoder fully causal (streaming) fine-tuned from GLM-4-Voice tokenizer | speech tokenizer · Librispeech ASR WER overall | WER: 9.17 -> 10.80 | Hurt |
| compress chunk size in flow-matching decoder (12.5hz) | CosyVoice 2 decoder · Seed-TTS-Eval-EN | WER: 4.63 -> 4.14 | Improved |
Closest papers by meaning
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.89)
- SALMONN-omni: A Standalone Speech LLM without Codec Injection for Full-duplex Conversation (NeurIPS 2025 · similarity 0.88)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.88)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.87)
- Language Model Can Listen While Speaking (AAAI 2025 · similarity 0.87)
- NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction (ICML 2025 · similarity 0.87)
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.
