Research map / Large language models / Language, safety and interpretability
DualSpeechLM: Towards Unified Speech Understanding and Generation via Dual Speech Token Modeling with Large Language Models
Yuanyuan Wang, Dongchao Yang, Yiwen Shao, Hangting Chen, Jiankun Zhao, Zhiyong Wu, Helen Meng, Xixin Wu
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 |
|---|---|---|---|
| use USToken instead of HuBERT as input token (ablation) | DualSpeechLM (Phi3.5-3B) · LibriSpeech test-clean ASR | WER: 5.56 -> 4.22 | Improved |
| use USToken instead of HuBERT as input token (ablation) | DualSpeechLM (Phi3.5-3B) · TTS LibriTTS clean | WER: 20.80 -> 9.25 | Improved |
| use USToken instead of HuBERT as input token (ablation) | DualSpeechLM (Vicuna-7B) · TTS LibriTTS clean | WER: 13.58 -> 12.63 | Improved |
| add AcousticGPT dual-token output (vs single-token baseline) | HuBERT-input Phi3.5-3B · TTS LibriTTS clean | WER: 21.72 -> 20.80 | Improved |
| add understanding-driven loss in USTokenizer (ablation) | DualSpeechLM · SQA | BLEU4: 37.67 -> 44.38 | Improved |
| add understanding-driven loss in USTokenizer (ablation) | DualSpeechLM · TTS LibriTTS clean | WER: 8.59 -> 9.25 | No clear effect |
| add reconstruction loss in USTokenizer (ablation) | DualSpeechLM · TTS LibriTTS clean | WER: 52.50 -> 9.25 | Improved |
| add reconstruction loss in USTokenizer (ablation) | DualSpeechLM · T2ST Es2En | BLEU4: 32.43 -> 26.77 | Hurt |
| add semantic supervision loss (ablation) | DualSpeechLM · TTS LibriTTS clean | WER: 167.56 -> 9.25 | Improved |
| add semantic supervision loss (ablation) | DualSpeechLM · SQA | BLEU4: 43.86 -> 44.38 | No clear effect |
| add Chain of Condition (ablation) | DualSpeechLM · VC VCTK | WER: 13.06 -> 10.16 | Improved |
| switch backbone from Phi3.5-3B to Vicuna-7B (ablation) | DualSpeechLM (USToken) · LibriSpeech test-clean ASR | WER: 4.22 -> 4.15 | No clear effect |
Closest papers by meaning
- SpeechTokenizer: Unified Speech Tokenizer for Speech Language Models (ICLR 2024 · similarity 0.90)
- NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction (ICML 2025 · similarity 0.88)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.88)
- UniWav: Towards Unified Pre-training for Speech Representation Learning and Generation (ICLR 2025 · similarity 0.87)
- Long-Form Speech Generation with Spoken Language Models (ICML 2025 · similarity 0.87)
- FastLongSpeech: Enhancing Large Speech-Language Models for Efficient Long-Speech Processing (NeurIPS 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.
