The paper 'Systematic PTQ Study of Integer and Floating-Point Formats for On-Device Whisper ASR', written by Choi Woo-seok (master's, first author), Lee Do-hyun (undergraduate), Kim Tae-hyung (master's) and Professor Lee Hyuk-jun (corresponding author) of the Intelligent Computer Architecture and Embedded Computing Laboratory (ECL), has been accepted for publication as a regular paper at Interspeech 2026.

▶ The PTQ accuracy–efficiency Pareto front by precision and numeric format

For deploying Whisper-based automatic speech recognition models in mobile and edge environments, this research systematically analyses post-training quantization (PTQ) performance across integer and floating-point formats. It compares more than 80 quantization configurations — INT8/4/3, FP8, FP4, NVFP4, MXFP4 and others — on the Whisper tiny.en and base.en models, and presents efficient quantization formats and hardware design guidance for on-device speech recognition.

[References]

Conference link: https://interspeech2026.org/en-AU