Research map / Architectures and efficiency / Attention and transformers
Scaling Transformers for End-to-End Discrete Audio Tokenization
Yitian Gong, Kuangwei Chen, Zhaoye Fei, Xiaogui Yang, Ke Chen, Yang Wang, Kexin Huang, Mingshu Chen, Ruixiao Li, Qinyuan Cheng, Shimin Li, Xipeng Qiu
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 Progressive Sequence Dropout (p>0) during AR TTS training (ablation) | TAC-TTS · Seed-TTS-Eval, reduced bitrate | SIM and WER | Improved |
| vary dropout probability p among 0.25, 0.5, 1.0 (ablation) | TAC-TTS · Seed-TTS-Eval | SIM and WER | No clear effect |
| increase dropout probability p (ablation) | TAC-TTS · Seed-TTS-Eval | training GPU memory | Improved |
| train without Progressive Sequence Dropout (p=0.0) (ablation) | TAC-TTS · Seed-TTS-Eval, full bitrate | SIM and WER | No clear effect |
| full end-to-end optimization instead of freezing encoder and quantizer (ablation) | TAC · speech reconstruction, 500k steps | SIM, STOI, PESQ | Improved |
| freeze encoder and quantizer (partial optimization) (ablation) | TAC · speech reconstruction | SIM, STOI, PESQ | Hurt |
| increase parameters from 319M to 1169M (ablation) | TAC · speech reconstruction, 0.5-4 kbps | SIM, STOI, PESQ | Improved |
| scale parameters without raising bitrate (1169M at low bitrate) (ablation) | TAC · speech reconstruction, low bitrate | SIM, STOI, PESQ | No clear effect |
| raise bitrate from 750 to 4000 bps (ablation) | TAC · speech reconstruction SIM (EN) | SIM: 0.82 -> 0.97 | Improved |
| use fully Transformer, end-to-end tokenizer vs hybrid baselines at 1000 bps | TAC · speech reconstruction SIM (EN) | SIM: 0.85 -> 0.88 | Improved |
| use TAC tokens for AR TTS vs prior discrete AR TTS | TAC-TTS · Seed-TTS-Eval EN | SIM: 67.7 -> 73.1 | Improved |
Closest papers by meaning
- WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.93)
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.92)
- Scaling Transformers for Low-Bitrate High-Quality Speech Coding (ICLR 2025 · similarity 0.91)
- Generative Audio Language Modeling with Continuous-valued Tokens and Masked Next-Token Prediction (ICML 2025 · similarity 0.90)
- A Unified Audio Language Model with Text-Aligned Factorized Audio Tokenization (NeurIPS 2026 · similarity 0.90)
- TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling (NeurIPS 2025 · similarity 0.88)
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.
