AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs

Yuhan Song, Linhao Zhang, Chuhan Wu, Aiwei Liu, Wei Jia, Houfeng Wang, Zhou Xiao

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Speech and audio language models

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 changedWhereMeasuredEffect
use StableToken instead of best supervised tokenizersupervised semantic tokenizer · FLEURS noisy avgUED: 26.17 -> 10.17Improved
add consensus loss (ablation)StableToken · FLEURS Real (OOD)UED: 17.43 -> 10.96Improved
add noise-aware training (ablation)StableToken · LibriSpeech-otherWER: 5.52 -> 4.88Improved
add multi-branch voting architecture (ablation)StableToken · FLEURS Real (OOD)UED: 24.47 -> 21.51Improved
increase voters from 3 to 5 (ablation)StableToken · FLEURS Gaussian noiseUED: 20.66 -> 18.68Improved
increase voters from 5 to 7 (ablation)StableToken · FLEURS Gaussian noiseUED: 18.68 -> 18.10No clear effect
use StableToken tokenizer in SpeechLLM instead of GLM-4-VoiceQwen2.5-3B SpeechLLM · CHiME-4 test realWER: 51.08 -> 35.90Improved
use StableToken tokenizer in SpeechLLM instead of GLM-4-VoiceQwen2.5-3B SpeechLLM · SEED-TTS ENMOS: 4.19 -> 4.12Mixed
place quantizer at layer 24 instead of 16 (ablation)StableToken · Gaussian UED vs ESD SERUED / SER acc: UED 18.68 -> 13.65; SER 67.38 -> 59.51Mixed
place quantizer at layer 8 instead of 16 (ablation)StableToken · Gaussian UEDUED: 18.68 -> 22.05Hurt
use 4:1 clean-to-noisy branch ratio instead of 3:2 (ablation)StableToken · Gaussian UEDUED: 18.68 -> 20.94Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy