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
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 TASTE text-aligned tokens instead of text-only reconstruction | TASTE tokenizer · LibriSpeech test-clean | WER (%): 5.9% -> 4.4% | Improved |
| TASTE vs text-only on duration consistency | TASTE tokenizer · LibriSpeech test-clean | duration consistency: 0.57 -> 0.91 | Improved |
| TASTE vs high-bitrate tokenizers on ViSQOL | TASTE tokenizer · LibriSpeech test-clean | ViSQOL: 3.05 vs 3.65 (SpeechTokenizer 2000 bps) | Hurt |
| use TASTE tokens instead of S3 tokens for spoken LM | TASLM 1B · speech continuation | GPT-4o MOS: 1.37 -> 3.08 | Improved |
| use TASTE tokens instead of S3 tokens for spoken LM | TASLM 1B · StoryCloze | accuracy: 58.7 -> 76.5 | Improved |
| TASLM embedding vs token variant | TASLM 1B · SALMON | accuracy: 60.8 -> 57.7 | Hurt |
| TASLM embedding vs token variant | TASLM 1B · speech continuation human MOS | human MOS: 3.93 -> 4.16 | Improved |
| add few-shot spoken QA via TASLM relative to base text LLM | LLaMA3.2-1B · LLaMA-Questions | accuracy: 51.0 -> 57.6 | Improved |
| add aggregator to encoder (ablation) | TASTE tokenizer · S3 token top-5 accuracy | top-5 accuracy: 0.98 -> 0.88 (rate 50Hz -> ~3Hz) | Hurt |
| add quantizer after aggregator (ablation) | TASTE tokenizer · S3 token top-5 accuracy | top-5 accuracy: 0.88 -> 0.76 | Hurt |
| use shallow hidden states instead of last hidden state as aggregator values (ablation) | TASTE tokenizer · S3 token top-5 accuracy | top-5 accuracy: 0.78 -> 0.88 | Improved |
| compare to cascade with same LLM | TASLM 1B (embed.) · speech continuation GPT-4o MOS | GPT-4o MOS: 3.15 vs 3.16 | No clear effect |
Closest papers by meaning
- MATS: An Audio Language Model under Text-only Supervision (ICML 2025 · similarity 0.88)
- A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization (NeurIPS 2026 · similarity 0.86)
- Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation (NeurIPS 2024 · similarity 0.86)
- Scaling Speech-Text Pre-training with Synthetic Interleaved Data (ICLR 2025 · similarity 0.86)
- Is Text All You Need? Text as a Universal Information Bottleneck for Speech LLMs (NeurIPS 2026 · similarity 0.86)
- Model-Aware Tokenizer Transfer (NeurIPS 2026 · 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.
