Itgan adapts whisper for better Arabic dialect and code-switching speech recognition
Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR
Computation and LanguageArtificial IntelligenceSound
Summary
Different Arabic dialects and mixing languages make speech recognition tricky. Itgan’s team improved a popular speech recognition model called Whisper by tweaking it efficiently on regular GPUs to handle these challenges. They tested their approach on three tasks involving country-level dialects, mixed dialects, and Tunisian Arabic mixed with other languages. Their method improved accuracy by customizing parts of the model for each dialect and combining different model versions. This work shows practical ways to adapt speech recognition to complex Arabic speech situations.
What this means in practice
- •For speech technology developers: Improve Arabic speech recognition systems by efficiently adapting Whisper models to handle dialect variations and code-switching on affordable hardware.
- •For multilingual service providers: Enhance automatic transcription services in Arabic-speaking regions by supporting mixed dialects and language switching with better-adapted ASR models.$Commercial implications: Enables offering more accurate transcription services in complex linguistic settings common in Arabic markets, improving product competitiveness.
Authors
Ibrahim Almajai
Abstract
We 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 (1.3). All three share one recipe, Whisper adapted with LoRA on consumer GPUs, and each was carried by a different addition to it. On 1.1, where the dialect label is given at test time, per-dialect specialists continued from a pooled adapter gave the largest gain, and the submitted system reached 57.1% country-average WER. A post-evaluation linear probe on frozen encoder features routes utterances without the label and recovers 44% of what oracle routing gives. On 1.2 the choice of base model mattered more than adapter capacity, and system combination helped only once we added a decorrelated member, reaching 46.7% WER. On 1.3 our system placed second at 14.49% WER with the lowest CER among the leading submissions, 5.38%. Its last 0.60 WER points came without further training, mostly from an exact weight-space average of independently trained runs, with ROVER voting adding the remainder. Every comparison carries a paired-bootstrap test, and we report eight directions that did not work.