Diagnostic-Driven Layer-Wise Compensation for Post-Training Quantization of Encoder-Decoder ASR Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Wang, Xinyu, Zhao, Ziyu, Luo, Yajie, Wu, Yihong, Ma, Liheng, Tian, Jingrui, Ding, Lei, Chang, Xiao-Wen, Lu, Peng
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914508715327488
author Wang, Xinyu
Zhao, Ziyu
Luo, Yajie
Wu, Yihong
Ma, Liheng
Tian, Jingrui
Ding, Lei
Chang, Xiao-Wen
Lu, Peng
author_facet Wang, Xinyu
Zhao, Ziyu
Luo, Yajie
Wu, Yihong
Ma, Liheng
Tian, Jingrui
Ding, Lei
Chang, Xiao-Wen
Lu, Peng
contents Deploying Automatic Speech Recognition (ASR) models on memory-constrained edge devices requires aggressive low-bit weight quantization. Layer-wise post-training quantization is practical and effective, but it suffers from cross-layer error accumulation. Existing compensation methods typically use a single global strength for all layers, which is ill-suited to encoder-decoder ASR models whose acoustic encoder and linguistic decoder exhibit markedly different sensitivities to quantization noise. We propose FADE, a diagnostic-driven framework that assigns each layer an adaptive compensation coefficient by combining two complementary signals: an intrinsic vulnerability score from weight geometry and a calibration reliability score from the data-driven solution. The resulting layer-wise coefficient balances local quantization fidelity against cross-layer error correction, enabling tailored compensation without retraining or hyperparameter search. Experiments on Whisper, Moonshine, and Qwen3-ASR across four benchmarks show that FADE consistently improves mean Word Error Rate over strong baselines at both 3- and 4-bit precision while substantially reducing run-to-run variance.
format Preprint
id arxiv_https___arxiv_org_abs_2601_02455
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Diagnostic-Driven Layer-Wise Compensation for Post-Training Quantization of Encoder-Decoder ASR Models
Wang, Xinyu
Zhao, Ziyu
Luo, Yajie
Wu, Yihong
Ma, Liheng
Tian, Jingrui
Ding, Lei
Chang, Xiao-Wen
Lu, Peng
Sound
Computation and Language
Audio and Speech Processing
Deploying Automatic Speech Recognition (ASR) models on memory-constrained edge devices requires aggressive low-bit weight quantization. Layer-wise post-training quantization is practical and effective, but it suffers from cross-layer error accumulation. Existing compensation methods typically use a single global strength for all layers, which is ill-suited to encoder-decoder ASR models whose acoustic encoder and linguistic decoder exhibit markedly different sensitivities to quantization noise. We propose FADE, a diagnostic-driven framework that assigns each layer an adaptive compensation coefficient by combining two complementary signals: an intrinsic vulnerability score from weight geometry and a calibration reliability score from the data-driven solution. The resulting layer-wise coefficient balances local quantization fidelity against cross-layer error correction, enabling tailored compensation without retraining or hyperparameter search. Experiments on Whisper, Moonshine, and Qwen3-ASR across four benchmarks show that FADE consistently improves mean Word Error Rate over strong baselines at both 3- and 4-bit precision while substantially reducing run-to-run variance.
title Diagnostic-Driven Layer-Wise Compensation for Post-Training Quantization of Encoder-Decoder ASR Models
topic Sound
Computation and Language
Audio and Speech Processing
url https://arxiv.org/abs/2601.02455