AI Paper Atlas

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

TASTE: Text-Aligned Speech Tokenization and Embedding for Spoken Language Modeling

Liang-Hsuan Tseng, Yi-Chang Chen, Kuan Yi Lee, Da-shan Shiu, Hung-yi Lee

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Sign and cross-lingual translation · 18 citations (OpenAlex)

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 TASTE text-aligned tokens instead of text-only reconstructionTASTE tokenizer · LibriSpeech test-cleanWER (%): 5.9% -> 4.4%Improved
TASTE vs text-only on duration consistencyTASTE tokenizer · LibriSpeech test-cleanduration consistency: 0.57 -> 0.91Improved
TASTE vs high-bitrate tokenizers on ViSQOLTASTE tokenizer · LibriSpeech test-cleanViSQOL: 3.05 vs 3.65 (SpeechTokenizer 2000 bps)Hurt
use TASTE tokens instead of S3 tokens for spoken LMTASLM 1B · speech continuationGPT-4o MOS: 1.37 -> 3.08Improved
use TASTE tokens instead of S3 tokens for spoken LMTASLM 1B · StoryClozeaccuracy: 58.7 -> 76.5Improved
TASLM embedding vs token variantTASLM 1B · SALMONaccuracy: 60.8 -> 57.7Hurt
TASLM embedding vs token variantTASLM 1B · speech continuation human MOShuman MOS: 3.93 -> 4.16Improved
add few-shot spoken QA via TASLM relative to base text LLMLLaMA3.2-1B · LLaMA-Questionsaccuracy: 51.0 -> 57.6Improved
add aggregator to encoder (ablation)TASTE tokenizer · S3 token top-5 accuracytop-5 accuracy: 0.98 -> 0.88 (rate 50Hz -> ~3Hz)Hurt
add quantizer after aggregator (ablation)TASTE tokenizer · S3 token top-5 accuracytop-5 accuracy: 0.88 -> 0.76Hurt
use shallow hidden states instead of last hidden state as aggregator values (ablation)TASTE tokenizer · S3 token top-5 accuracytop-5 accuracy: 0.78 -> 0.88Improved
compare to cascade with same LLMTASLM 1B (embed.) · speech continuation GPT-4o MOSGPT-4o MOS: 3.15 vs 3.16No clear effect

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