New FADE method improves ASR model quantization for edge devices

Researchers have developed FADE, a novel framework for improving post-training quantization of encoder-decoder Automatic Speech Recognition (ASR) models. This method addresses the issue of error accumulation across layers by assigning adaptive compensation coefficients to each layer. FADE combines intrinsic vulnerability scores from weight geometry with data-driven calibration reliability scores to balance local fidelity and cross-layer error correction. Experiments on models like Whisper and Qwen3-ASR demonstrated consistent improvements in Word Error Rate at 3- and 4-bit precision. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enables more efficient deployment of ASR models on memory-constrained edge devices by improving quantization accuracy.

RANK_REASON This is a research paper detailing a new framework for model quantization.

Read on arXiv cs.CL →

• FADE
• Whisper
• Qwen3-ASR
• Ziyu Zhao
• arXiv

• paper
• infra