Fine-tuning Whisper for Pashto ASR: strategies and scale

Hanif Rahman · 2026

Read Fine-tuning Whisper for Pashto ASR: strategies and scale by Hanif Rahman (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-2604-06507.txt.

Description

Rahman fine-tunes Whisper for Pashto, missing from its pre-training. Off-the-shelf models print Arabic, Dari, or Urdu script for Pashto audio above 100 percent error. Vanilla fine-tuning wins at 21.22 percent on whisper-base, beating LoRA, frozen-encoder, and Urdu transfer. Frozen-encoder tuning degrades the model; Urdu transfer fails on an unverified checkpoint, mismatch, and thin training. On newer data, whisper-small reaches about 25 percent and large-turbo about 23 percent, making small optimal. Augmentation adds about 7 points; analysis finds suffix confusion and retroflex substitutions dominant.

Catalog record

CREATORHanif Rahman
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50185
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.06507

Browse the full library · About the Tower of Babel