Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR
LoRA-adapted Whisper systems for NADI 2026 Arabic ASR reach 57.1%, 46.7%, and 14.49% WER on three subtasks.
Itgan submitted one LoRA-adapted Whisper recipe for the three NADI 2026 Arabic ASR subtasks. With the dialect label available, per-dialect specialists continued from a pooled adapter reached 57.1% country-average WER; a post-evaluation probe without the label recovered 44% of the oracle-routing gain. Mixed-dialect ASR reached 46.7% WER once system combination included a decorrelated member, with base-model choice mattering more than adapter capacity. The Tunisian code-switched system placed second at 14.49% WER and 5.38% CER, with the last 0.60 WER coming mostly from an exact weight-space average of independent runs plus ROVER.
- All three systems adapt Whisper with LoRA on consumer GPUs.
- Country-level ASR reached 57.1% average WER using per-dialect specialists.
- Mixed-dialect combination reached 46.7% WER after a decorrelated member.
- Tunisian code-switched ASR placed second at 14.49% WER and 5.38% CER.
- The last 0.60 WER came from weight averaging and ROVER without more training.
From the original article
1 min read at arxiv.orgWe describe the Itgan systems for the three ASR subtasks of NADI 2026, namely robust country-level ASR (1. 1), mixed-dialect ASR (1. 2), and Tunisian code-switched ASR (1.
By Ibrahim Almajai · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.