Dynamic Quantization Error Propagation in Encoder-Decoder ASR Quantization
Xinyu Wang 1* , Yajie Luo 2* , Yihong Wu 2* , Liheng Ma 1,5* Ziyu Zhao 1 † \dagger , Jingrui Tian 3 , Lei Ding 4 , Yufei Cui 1,5 and Xiao-Wen Chang 1 · 2026
Read Dynamic Quantization Error Propagation in Encoder-Decoder ASR Quantization by Xinyu Wang 1* , Yajie Luo 2* , Yihong Wu 2* , Liheng Ma 1,5* Ziyu Zhao 1 † \dagger , Jingrui Tian 3 , Lei Ding 4 , Yufei Cui 1,5 and Xiao-Wen Chang 1 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-02455.txt.
Description
Xinyu Wang and six co authors improve post training quantization for speech recognition models, the compression step that shrinks a trained model to three or four bits per weight without retraining. When layers are quantized one by one, each layer sees inputs already distorted by the layers before it. An earlier compensation method used one coefficient for the whole model, mixing the global setting with differences between layers.
Their method, FADE, builds a separate coefficient for each layer from the layer's rounding distortion and a calibrated solver response. With no training and no per model search, and adds nothing at run time. Across seven Whisper, Moonshine, and Qwen ASR models and four English benchmarks, FADE lowered word error rates in most settings compared with the fixed earlier method. Gains were largest where three bit models had collapsed, which the authors frame as damage control rather than usable accuracy.
Catalog record
| CREATOR | Xinyu Wang 1* , Yajie Luo 2* , Yihong Wu 2* , Liheng Ma 1,5* Ziyu Zhao 1 † \dagger , Jingrui Tian 3 , Lei Ding 4 , Yufei Cui 1,5 and Xiao-Wen Chang 1 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-02455