Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR

Ibrahim Almajai

cs.CLOct 7, 2026

Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR

Authors: Ibrahim Almajai

Organizations: Independent Researcher

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.

Figures & tables

Country	Backbone	Dev $\Delta$	Test WER
Mauritania	turbo + spec.	$-6.5$	77
Algeria	turbo + spec.	$-3.1$	72
Morocco	turbo + spec.	$-2.9$	60
Egypt	turbo + spec.	$-1.9$	58
Yemen	large-v3	$-2.8^{*}$	63
Jordan	large-v3	$-2.7^{*}$	44

Table 1: Subtask 1.1 submitted system, all figures in %. Dev

\Delta

is the full-development WER change in WER points from the pooled adapter to that country’s specialist; starred rows give the change from turbo to large-v3 on the subset that decided the backbone. Every specialist gain is significant but Jordan’s (Appendix C ). Deltas are scored without the length guard; the guarded progression of Table 4 is smaller on Mauritania and Yemen, where the guard does part of the specialist’s work (Appendix H ).

System	Trainable	WER
turbo , zero-shot	n/a	60.04
turbo + LoRA (narrow)	6.6M	56.54
turbo + LoRA (wide)	27.9M	55.51
large-v3 , zero-shot (int8)	n/a	57.21
large-v3 + LoRA (wide, $r$ 16)	28.8M	52.88
large-v3 + LoRA (wide)	57.7M	51.76

Table 2: Subtask 1.2, 615 held-out utterances (%). Differences quoted in the text are paired-bootstrap estimates (Appendix C ) and can differ by about 0.01 from differences of the values printed here.

		Validation (731)		Practice test (842)		Blind test (2,123)
System	Trainable	WER	CER	WER	CER	WER / CER
Zero-shot whisper-large-v3 †	n/a	85.00	59.84	87.57	64.40	n/a
LoRA ( q,v ), 3k steps	15.7M	21.62	9.11	24.68	11.79	n/a
LoRA (wide), 5k steps	57.7M	19.84	8.26	21.74	10.16	n/a
+ dropout 0.15, SpecAugment	57.7M	19.46	8.29	21.19	10.08	15.09 / 5.56
+ adapter soup ( $N{=}4$ )	n/a	17.19	6.60	20.14	9.51	14.55 / 5.40

Table 3: Subtask 1.3 results (%). † the organizers’ baseline notebook; under the beam-5 decode that all trained rows use, the checkpoint loops and scores 103.2 / 117.8. The final system reduces blind-test WER by 0.60 points over the best single trained model.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Country-average WER (%)
pooled adapter, greedy	57.59
+ length guard	56.58
+ beam 5	55.31
all eight specialists + guard	54.67
+ beam 5	53.55

Appendix

Table 4: Post-evaluation development results for Subtask 1.1. The length guard applies from the second row down, unlike the per-country deltas of Table 1 , which are unguarded. The only training beyond Table 1 is the four one-epoch specialists described above; the guard and the beam are free.

Comparison (validation)	$\Delta$ WER	$P(\text{better})$
q,v $\rightarrow$ wide+reg	$-2.17$	99%
wide $\rightarrow$ wide+reg	$-0.40$	78%
wide+reg $\rightarrow$ soup ( $N{=}3$ )	$-0.81$	97%
wide+reg $\rightarrow$ soup ( $N{=}4$ )	$\mathbf{-2.26}$	100%
soup ( $N{=}4$ ) $\rightarrow$ soup ( $N{=}6$ )	$+0.61$	21%
same-run pair $\rightarrow$ two-seed pair	$-2.38$	100%

Appendix

Table 5: Paired bootstrap over utterances, 4,000 resamples.

\Delta

WER is in WER points. Above the rule, the ablations behind Section 6 ; below it, the post-evaluation seed control of Section 6.1 .

Country	$n$	$\Delta$ WER	$P(\text{better})$
Mauritania	1,600	$-6.53$	100%
Yemen	1,183	$-3.90$	100%
Algeria	726	$-3.12$	100%
Morocco	1,600	$-2.88$	99%
Egypt	1,600	$-1.90$	99%
UAE	1,600	$-1.60$	100%

Appendix

Table 6: Subtask 1.1 specialists, in WER points.

Script	Ref. words	Share	Error
Arabic	7,887	66.7%	20.2%
Latin (Fr/En)	3,946	33.3%	12.5%

Appendix

Table 7: Word error (substitutions + deletions) by script of the reference word, best single model, practice test split.

Country	Steps	Spec.	Control	Spec. only
Algeria	300	-3.1	-0.4	-2.7
Egypt	300	-1.8	-0.0	-1.7
Mauritania	300	-2.8	+0.1	-2.9
Morocco	300	-2.2	+0.6	-2.8
Jordan	100	-0.6	+0.2	-0.9
Palestine	100	-1.1	+0.2	-1.3

Appendix

Table 8: Each specialist against a pooled control given the same number of optimizer steps. Guarded full-development WER change from the pooled adapter, in WER points. Spec. only is Spec. minus Control .

Spec.	Egy	Jor	Alg	UAE	Mor	Mau	Yem	Pal
Egy	-1.8	+2.3	-0.3	+1.8	+2.2	+0.6	+0.7	+2.1
Jor	+2.0	-0.6	+1.4	+1.5	+1.5	+0.2	+1.8	+1.1
Alg	+1.4	+1.9	-3.1	+1.0	+4.0	+0.6	+1.3	+1.8
UAE	+0.8	+1.4	-1.1	-1.6	+1.2	+0.3	-0.4	+1.2
Mor	+1.5	+2.9	+1.3	+1.9	-2.2	+1.2	+1.7	+2.4
Mau	+2.7	+3.7	+0.6	+2.5	+3.4	-2.8	+2.5	+2.0

Appendix

Table 9: Change in guarded development WER, in WER points, from the pooled adapter when each specialist (row) decodes each country (column). Bold is the specialist’s own dialect.

System	Country-avg WER	$P(\text{better})$
pooled adapter	56.56	—
soup of all eight	56.31	99%
routed by the probe	55.73	100%
oracle routing	54.66	100%

Appendix

Table 10: Subtask 1.1 development, guarded, in %.

P(\text{better})

is against the pooled adapter, paired bootstrap over utterances with 4,000 resamples. Scored by the per-utterance routing pass, which reproduces Appendix B ’s pooled figure to within 0.02 points.

Explore similar work

NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task
Sep 22, 2026Peter Sullivan, Bashar Talafha, Ahmed Ashraf +11Arabic Natural Language Processing Arabic
The University of British Columbia · King Fahd University of Petroleum & Minerals · Avignon Université +4
CARDAMOM: A Micro-Dialectal Arabic Speech Dataset for ASR
Sep 28, 2026Bashar Talafha, Samar M. Magdy, Aisha Alansari +36Dialects Multilingual Automatic Speech Recognition
The University of British Columbia · King Fahd University of Petroleum and Minerals · Al al-Bayt University +13
Almieyar: A Culturally Grounded Benchmark for Multi-Dialect Arabic Speech Recognition
Sep 28, 2026Omid Ghahroodi, Anas Madkoor, Dima Faris Al Saudi +37Arabic Natural Language Processing Arabic
QCRI, HBKU · Qatar University · UDST +15

Sep 22, 2026, cs.CLJ/K move · Enter open · S save

NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task

Peter Sullivan, Bashar Talafha, Ahmed Ashraf, Fethi Bougares+10

The University of British Columbia · King Fahd University of Petroleum & Minerals · Avignon Université+4

Arabic Natural Language Processing Arabic

Open paper AskPDF arXiv

NADI 2026 is the seventh edition of the Nuanced Arabic Dialect Identification (NADI) shared task series and the second dedicated to multidialectal Arabic speech processing. This edition comprises five tasks and eight subtasks spanning Automatic Speech Recognition (ASR), Spoken Dialect Identification (SDID), Text-to-Speech (TTS), Spoken Language Translation (SLT), and Spoken Language Understanding (SLU). NADI 2026 emphasizes realistic evaluation through low-bandwidth, mixed-dialect, code-switched, out-of-domain, and zero-shot settings, while introducing TTS, SLT, and SLU to the series for the first time. The shared task attracted 21 participating teams from at least 13 countries, with 48 test-phase submissions and 14 submitted system-description papers. Results show that out-of-domain generalization remains a major bottleneck and highlight the effectiveness of recent Arabic-specialized speech models, multimodal dialect identification approaches, and ensemble methods. Overall, NADI 2026 provides a broader and more challenging benchmark for robust Arabic dialect speech processing.

Abstract

Explore similar work

NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task

CARDAMOM: A Micro-Dialectal Arabic Speech Dataset for ASR

Almieyar: A Culturally Grounded Benchmark for Multi-Dialect Arabic Speech Recognition