Organizations: Shanghai Jiao Tong University · Shanghai Innovation Institute · University of Chinese Academy of Sciences · The Chinese University of Hong Kong, Shenzhen · Tsinghua University · Dalian University of Technology · Alibaba Token Foundry · Fudan University · SpeechColab
Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.
Figures & tables
Fig. 1: Overview of GigaSpeech 3 construction and validation. Channel-disjoint manual and web data train dialect-specific VietASR labeling models, whose hypotheses are filtered to form the five-dialect corpus. Independent Zipformer training and Qwen3-ASR adaptation evaluate the resulting weak supervision; labeled MSA is added only to the Multi-Dialect Zipformer and Qwen training recipes.
Corpus
Language coverage
Duration (h)
Supervision type
GigaSpeech [ 34 ]
English
10k
Human-provided transcripts
GigaSpeech 2 [ 23 ]
Thai, Indonesian, and Vietnamese
30k
Refined weak labels
GigaSpeech 3 (ours)
Arabic: DZA, IRQ, SAU, SYR, and EGY
39k
Refined weak labels
TABLE I: Language coverage, duration, and supervision type in the GigaSpeech series.
Dialect
Duration (h)
Human-labeled
Unlabeled
Refined weak-labeled
DZA
61.1
17,812
11,461
IRQ
63.2
15,531
10,162
SAU
58.3
20,135
4,512
SYR
59.0
14,905
4,361
EGY
51.1
12,183
8,533
TABLE II: Training-data duration for the five target dialects (hours).
Resource
Country codes
Scale (h)
Supervision
Casablanca [ 10 ]
DZA, EGY, ARE, JOR, MRT, MAR, PSE, YEM
48
Human labels
CARDAMOM [ 11 ]
EGY, JOR, LBN, MRT, PSE, SAU
40
Human labels
Hamsa [ 12 ]
ARE
12
Human labels
YODAS v3 [ 13 ]
Unspecified
25k
Weak labels
GigaSpeech 3 (ours)
DZA, IRQ, SAU, SYR, EGY
39k
Refined weak labels
TABLE III: Comparison with recent Arabic speech resources.
Dataset
Duration (h)
DZA
IRQ
SAU
SYR
EGY
GigaSpeechBench
20.0
20.4
20.3
20.2
20.5
MASC
0.1
0.7
1.4
6.6
4.5
Casablanca
1.9
–
–
–
2.0
SADA
–
–
19.9
–
–
Total
22.0
21.0
41.6
26.8
27.0
TABLE IV: Duration of the Arabic evaluation sets by dialect (hours). Development and test partitions are merged for MASC, Casablanca, and SADA.
Model
DZA
IRQ
SAU
SYR
EGY
MSA
Labeled only
53.47
42.50
13.17
58.88
43.92
15.98
SSL + FT
38.47
31.35
12.98
15.11
31.81
–
Dialect-Specific
33.64
29.99
12.12
14.25
29.32
–
Multi-Dialect
36.78
28.85
12.81
16.33
32.80
10.83
TABLE V: GigaSpeechBench WER (%) by model and dialect. The final models use the released weak transcriptions; SAU and SYR include four-round NST refinement.
Model
DZA
IRQ
SAU
SYR
EGY
Avg.
Dolphin Base
78.26
65.20
52.35
38.12
85.31
63.85
Dolphin Small
72.44
62.05
50.91
30.03
74.70
58.03
GPT-4o Transcribe
63.14
54.53
42.38
31.67
64.23
51.19
FunASR-Realtime
66.30
53.44
37.67
24.24
63.33
49.00
Whisper Large v3
72.02
51.04
32.79
19.12
69.78
48.95
Deepgram Nova 3
57.90
47.54
25.02
30.61
52.77
42.77
TABLE VI: WER (%) on five GigaSpeechBench v3 dialects [ 7 ] . Third-party systems are ordered by decreasing average WER; italics mark the best third-party value, and bold/underlining mark the best/second-best overall.
Model
MASC
Casablanca
SADA
DZA
IRQ
SAU
SYR
EGY
DZA
EGY
SAU
Avg.
Whisper-CS
59.49
51.96
45.05
28.43
41.45
90.52
51.85
60.39
53.64
QwenCleo
62.26
51.64
47.07
29.48
35.74
88.12
49.08
50.67
51.76
Qwen3-ASR-1.7B (Base)
56.08
41.79
34.81
22.51
40.80
82.89
57.83
40.06
47.10
Qwen3-ASR-1.7B FT w/o LID (ours)
48.61
36.03
24.03
20.95
36.29
62.59
42.76
30.97
37.78
Dialect-Specific (ours)
49.57
37.01
21.23
21.33
34.04
50.06
37.34
24.54
34.39
TABLE VII: Cross-corpus WER (%) on MASC, Casablanca, and SADA. The same final checkpoints are used throughout; “Avg.” is the unweighted mean over eight corpus–variety conditions. Bold and underlining mark the best and second-best configurations.
True \ Pred.
DZA
IRQ
SAU
SYR
EGY
MSA
Acc.
DZA
10707 (64%)
44 (0%)
348 (2%)
18 (0%)
557 (3%)
4985 (30%)
64.3%
IRQ
78 (1%)
10241 (74%)
601 (4%)
127 (1%)
581 (4%)
2249 (16%)
73.8%
SAU
142 (1%)
663 (6%)
6757 (58%)
119 (1%)
465 (4%)
3443 (30%)
58.3%
SYR
216 (3%)
909 (11%)
556 (7%)
5004 (60%)
91 (1%)
1583 (19%)
59.9%
EGY
27 (0%)
205 (1%)
636 (4%)
44 (0%)
10159 (63%)
4961 (31%)
63.4%
MSA
6 (0%)
14 (0%)
14 (0%)
7 (0%)
15 (0%)
4529 (99%)
98.8%
TABLE VIII: LID confusion matrix for the Multi-Dialect Zipformer averaged over epochs 4–8. Each cell gives the number of utterances followed by its row percentage; bold diagonal entries are correct predictions.
Variety
t<3 s
3≤t<10 s
t≥10 s
DZA
39.8 (7,808)
85.0 (8,300)
99.1 (551)
IRQ
51.5 (4,981)
84.6 (7,620)
96.6 (1,276)
SAU
8.4 (2,801)
76.0 (7,436)
64.4 (1,352)
SYR
24.4 (394)
61.7 (5,037)
61.4 (2,928)
EGY
65.6 (7,030)
62.3 (8,252)
54.1 (750)
MSA
97.9 (192)
98.8 (4,385)
100.0 (8)
TABLE IX: Duration-stratified LID accuracy for the Multi-Dialect model averaged over epochs 4–8. Each cell reports accuracy (%) followed by the number of utterances in parentheses.
Dialect
r0 (initial)
r1
r2
r3
r4 (final)
Rel. reduction
SAU
13.13
12.96
12.42
12.32
12.12
7.7%
SYR
14.95
14.69
14.51
14.29
14.25
4.7%
TABLE X: Noisy-student refinement on SAU and SYR (WER %). r0 is the initial weak-label model and r1 – r4 are successive refinements; released data correspond to r4 .
Model
DZA
IRQ
SAU
SYR
EGY
MSA
Avg.
QwenCleo
71.04
55.46
32.67
28.51
44.65
24.79
42.85
Whisper-CS
74.86
57.86
31.02
28.24
58.91
26.56
46.24
Qwen3-ASR-1.7B (Base)
63.43
41.27
25.85
18.50
59.23
22.22
38.42
Qwen3-ASR-1.7B FT w/o LID
47.71
34.48
16.81
15.93
41.02
13.72
28.28
Qwen3-ASR-1.7B FT w/ LID
39.25
30.47
14.63
13.82
32.91
13.53
24.10
TABLE XI: Large-model comparison on GigaSpeechBench (WER %). Fine-tuned Qwen uses weakly labeled dialect speech and manually labeled MGB-2 speech for MSA; Avg. is the unweighted macro-average.
Arabic speech technology has largely focused on Modern Standard Arabic, leaving the living dialects spoken by hundreds of millions under-served. We introduce ALMIEYAR, a culturally grounded ASR benchmark covering 17 Arabic dialects across six families, built entirely from newly recorded speech unseen by existing models. Dialect-community coordinators selected culturally relevant images across 10 topics, and native speakers described them through five structured scenarios, yielding approximately 50 minutes per dialect (13.7 hours total). We benchmark 12 state-of-the-art ASR systems zero-shot, including GPT-4o-transcribe, Voxtral-Mini-4B, Fanar-STT-LF, Whisper, SeamlessM4T-v2, and wav2vec2-based models. GPT-4o-transcribe achieves the lowest overall WER at 35.0%, followed by Voxtral-Mini-4B, Fanar-STT-LF, and Whisper-Large-v3 at 41.1%, 45.9%, and 49.5%, respectively, indicating substantial remaining errors across Arabic dialect communities. Performance varies considerably across dialect groups, with no model performing uniformly best across all groups. WER alone also obscures dialectal ASR behaviour: wav2vec2-based models show large WER/CER gaps, where character-level agreement remains much higher than word-level accuracy, motivating joint WER/CER reporting. ALMIEYAR provides a unified benchmark for culturally grounded Arabic ASR evaluation, including the first published benchmark for Ahwazi Arabic.
Omid Ghahroodi, Anas Madkoor, Dima Faris Al Saudi +37
Arabic automatic speech recognition (ASR) faces unique challenges due to diglossia, extensive regional dialect variation, and limited speech resources. Existing speech datasets often focus on single dialects or large-scale broadcast/web data, leading to trade-offs between linguistic diversity and annotation quality. We present BULBUL, a multi-dialect Arabic ASR dataset collected from 275 speakers in 11 Arab countries. BULBUL includes structured dialect and sub-dialect coverage, as well as recordings of classical Arabic and modern standard Arabic spoken by participants in their native dialectal accents to support accent-aware modeling. The quality of the recordings was ensured through a two-level human verification process. We further benchmark a range of recent ASR systems, establishing strong baselines for modern dialectal and accented Arabic ASR.
We present Ara-BEST-RQ, a family of self-supervised learning (SSL) models specifically designed for multi-dialectal Arabic speech processing. Leveraging 5,640 hours of crawled Creative Commons speech and combining it with publicly available datasets, we pre-train conformer-based BEST-RQ models up to 600M parameters. Our models are evaluated on dialect identification (DID) and automatic speech recognition (ASR) tasks, achieving state-of-the-art performance on the former while using fewer parameters than competing models. We demonstrate that family-targeted pre-training on Arabic dialects significantly improves downstream performance compared to multilingual or monolingual models trained on non-Arabic data. All models, code, and pre-processed datasets will be publicly released to support reproducibility and further research in Arabic speech technologies.
Haroun Elleuch, Ryan Whetten, Salima Mdhaffar +2
ELYADATA, Paris, France · Laboratoire Informatique d’Avignon, Avignon Université, Avignon, France