PaperlayerView exact revision on arXiv

Source-linked AI summary

GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data

Junzhe Liu, Yujie Tu, Yifan Yang, Peng Wang, Haoran Wang, Guodong Lin, Leyan Liao, Chengquan Zhang, Zhifu Gao, Xiangang Li, Qinyuan Cheng, Xipeng Qiu, Wei-Qiang Zhang, Guoguo Chen, Kai Yu, Xie Chen

arXiv:2611.02476v2cs.CL

TL;DR

Dialectal Arabic speech recognition lacks dialect-resolved transcribed speech at scale, leaving open whether automatically generated transcripts can provide reusable training supervision. GigaSpeech 3 builds a weakly labeled corpus across five dialects; its final dialect-specific models reach 23.86 macro-average WER, a 16.6% relative reduction from the strongest third-party system covering all five dialects.

  • Problem

    Dialect-resolved transcribed Arabic speech remains limited relative to large Arabic speech collections, and whether automatically generated transcripts can form reusable training supervision remains an open question.

    Source: §I · pp. 1–2

  • Method

    GigaSpeech 3 builds a dialect-matched weakly labeled web-speech corpus through ASR-biased labeling, selection, and refinement, then evaluates its utility with independently initialized downstream models.

    Source: §I · p. 2, B. ASR-biased speech pretraining · p. 3

  • Results

    23.86 macro-average WER on GigaSpeechBench, a 16.6% relative reduction from the strongest third-party system covering all five dialects; final dialect-specific models also outperform their labeling models across all five dialects.

    Source: §I · p. 2

  • Takeaways & Limitations

    The results demonstrate utility across dialect-specific and multi-dialect recognition, external domains, and model families.

    Source: §I · p. 2, §IX · p. 13

    Weak-label quantity, confidence, and source diversity are not independently controlled, and web data may inherit demographic, topical, speaker, and transcription biases.

    Source: §VIII · p. 13

Abstract

from arXiv · show

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.

I. INTRODUCTION

GigaSpeech 3 addresses the shortage of large-scale, reliably dialect-labeled Arabic training data with a 39,028-hour weakly labeled corpus across five dialects. Its final models improve over their labeling models and show strong benchmark and cross-domain results.

I. INTRODUCTION

  • Arabic ASR spans regional varieties with substantial pronunciation, vocabulary, morphology, and spelling differences, while existing large collections lack balanced country-level supervision and smaller annotated corpora are not training-scale.The remaining gap is large-scale dialect-specific training data with reliable dialect labels, rather than Arabic speech availability itself.Source: §I · p. 1
  • The corpus uses speech-recognition-biased self-supervised labeling models trained from small manually transcribed sets, then tests whether downstream models can improve without inheriting the labeling models’ pretrained parameters.An online proofreading platform enables community review and correction of weak transcriptions.Source: §I · p. 2
  • GigaSpeech 3 contains 39,028 hours and 10 million weakly transcribed segments across DZA, EGY, IRQ, SAU, and SYR, with 4,361–11,461 hours per dialect.Starting from manually transcribed speech per dialect, the corpus uses dialect-matched web audio, selection, and NST refinement; SAU/SYR partitions are filtered through round four.Source: §I · p. 2
  • Across eight conditions from MASC, Casablanca, and SADA, a GigaSpeech 3-trained Zipformer has the lowest WER among six evaluated models, while Qwen3-ASR fine-tuning improves performance in all eight.These results demonstrate transfer across evaluation domains and model families.Source: §I · p. 2
  • The multi-dialect model reaches 25.51 WER with predicted dialect IDs, compared with 28.62 for the strongest complete third-party system, and also supports MSA.The introduction also reports that predicted-LID routing errors increase WER relative to known-ID routing, especially for EGY and SYR.Source: §I · pp. 1–2
  • 23.86 macro-average WER on GigaSpeechBench, with final dialect-specific models improving over labeling models on all five dialects and a 16.6% relative reduction over the strongest third-party system covering all five.The dialect-specific models lead on DZA, EGY, and SAU, while the strongest third-party system retains an advantage on IRQ and SYR.Source: §I · p. 2

II. RELATED WORK … C. Automatic weak labeling at scale

Prior work develops self-supervised and ASR-biased speech representations and scalable weak labeling; GigaSpeech 3 applies these ideas to Arabic dialects and evaluates the resulting labels with independent downstream recognizers.

A. Low-resource ASR through self-supervised and multilingual learning

  • Self-supervised and multilingual approaches improve representations or parameter sharing, but still require task-specific transcripts during adaptation,,,, [18],.GigaSpeech 3 instead uses self-supervised learning to build dialect-specific labeling models and generate weak labels from dialect-matched unlabeled speech.Source: A. Low-resource ASR through… · p. 2

B. ASR-biased speech pretraining

  • ASR-biased pretraining guides target generation, iterative training, or the loss, including supervision-guided codebooks, ASBERT, Biased SSL, and CTCBERT –.These methods use labeled speech to improve the specificity of masked-prediction targets or reduce sensitivity to target-boundary errors.Source: B. ASR-biased speech pretraining · p. 2
  • VietASR combines ASR-biased iterative targets, an Fbank front end, and a multi-rate Zipformer; its 68M-parameter model trained on 50 labeled hours was competitive with larger systems,.It clusters representations from a small-set RNN-T model, predicts units on masked Fbank sequences, and repeats target extraction and pretraining with an improved encoder.Source: B. ASR-biased speech pretraining · p. 3
  • Independent downstream Zipformer training and cross-architecture Qwen adaptation test whether GigaSpeech 3’s weak labels benefit models beyond the recognizers that generated them.The paper identifies the labeling recipe as a foundation rather than its claimed novelty.Source: B. ASR-biased speech pretraining · p. 3

C. Automatic weak labeling at scale

  • Weak-labeling pipelines refresh recognizer hypotheses and control label errors, adapt multilingual recognizers to target-language audio, or select audio–text pairs using multiple models.Effectiveness depends on decoding, augmentation, and label-error control.Source: C. Automatic weak labeling at scale · p. 3
  • GigaSpeech 2 uses general-purpose models for transcription, alignment, filtering, and iterative refinement, whereas GigaSpeech 3 targets dialects with adapted dialect-specific labelers and independent downstream validation.GigaSpeech 3 also separates gains from its initial weak corpus from those added by filtering and iterative refinement.Source: C. Automatic weak labeling at scale · pp. 3–4
  • YODAS v3 reports roughly 25,000 hours of Arabic but uses an unreliable locale proxy and reports no Arabic ASR results; GigaSpeech 3 curates five dialects and evaluates dialect ASR models.The YODAS release’s locale field is missing for many recordings, and its language directories use caption-track language when available.Source: C. Automatic weak labeling at scale · p. 4

D. Arabic and multi-dialect speech recognition · III. DATA AND PROBLEM SETUP

Arabic ASR resources and methods have broadened dialect coverage, but country-level training data and performance remain uneven. This section motivates dialect-matched supervision and specifies evaluation conditions for the study.

D. Arabic and multi-dialect speech recognition

  • The study selects DZA, EGY, SYR, SAU, and IRQ as representatives of five broad Arabic regional groups, whose boundaries are approximate and internally varied.Source: D. Arabic and multi-dialect speech… · p. 4
  • MGB-2, MGB-3, and MGB-5 expanded dialect-focused evaluation but did not provide large-scale ASR training data for country-level varieties.Source: D. Arabic and multi-dialect speech… · p. 4
  • QASR, SADA, and Casablanca broaden broadcast and regional coverage, but their dialect inventories, scales, and transcription conventions remain uneven.QASR contains 2,000 hours of Al Jazeera speech, SADA 668 hours from Saudi TV programs, and Casablanca conversational speech for eight countries.Source: D. Arabic and multi-dialect speech… · p. 4
  • CARDAMOM and Hamsa prioritize precise dialect-specific supervision, while GigaSpeech 3 scales country-level weak supervision and tests it with an independently initialized recognizer [11].CARDAMOM labels 21 subcountry micro-dialects and overlaps with the EGY and SAU coverage; Hamsa contributes a complementary Emirati gold benchmark, so their results are not directly comparable baselines.Source: D. Arabic and multi-dialect speech… · p. 4
  • Arabic ASR research spans dialect-specific adaptation, joint modeling, country-level routing, and Arabic-specific pretraining, yet gains remain uneven across countries and domains [38].ADI17 illustrates the utility and noise of source-channel dialect assignment with roughly 3,000 hours of weakly labeled YouTube speech and 57 hours of manually refined development and test data.Source: D. Arabic and multi-dialect speech… · p. 4
  • Evaluation uses GigaSpeechBench Arabic test sets with training channels held separate, predicted dialect IDs for the main multi-dialect comparison, and text normalization, manifests, and decoding conditions as protocol [7].A matched comparison with and without ground-truth LID guidance diagnoses routing errors; improving dialect identification is outside the contribution. Results from incompatible protocols are reported separately and do not support direct superiority claims.Source: D. Arabic and multi-dialect speech… · pp. 4–5

A. Dialect scope … D. Position relative to existing Arabic resources

GigaSpeech 3 scales dialect-matched weak supervision from a few dozen manually transcribed hours to a 39,028-hour corpus spanning five country-level Arabic varieties. Its web-derived data are separated from evaluation channels and complement, rather than replace, fine-grained gold resources.

A. Dialect scope

  • The study covers Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic as coarse country-level representatives, not complete taxonomies of their substantial regional and social variation.The varieties represent Maghrebi, Egyptian, Mesopotamian, Gulf, and Levantine groups, respectively.Source: A. Dialect scope · p. 5

B. Web collection and split integrity

  • Candidate public-web audio is screened using source identifiers, metadata, dates, and duration, then segmented and quality-checked before manual transcription or assignment to dialect-specific unlabeled pools.A small subset forms the professionally transcribed set; the remaining material forms the unlabeled pool.Source: B. Web collection and split integrity · p. 5
  • Manual-label, unlabeled, and GigaSpeechBench channels are mutually disjoint, reducing overlap from recurring speakers, programs, or channel acoustics and requiring labeling models to generalize to new channels.The split is enforced at channel level rather than only at segment level.Source: B. Web collection and split integrity · p. 5
  • The unlabeled pool preserves heterogeneous equipment, noise, speaking styles, topics, and MSA or code-switching, making the task real-world dialectal recognition rather than a single controlled domain.Dialect-matched labeling models and downstream selection are used to construct the final corpus from this diverse pool.Source: B. Web collection and split integrity · p. 5

C. Supervision inventory

  • 39,028 hours and ∼10 million segments make up the final weak-label inventories, with ∼4,361 hours for SYR and 11,461 hours for DZA.Source collections contain 12,183–20,135 hours per dialect, compared with 51.1–63.2 manually transcribed hours per dialect; released SAU and SYR partitions reflect four NST rounds.Source: C. Supervision inventory · p. 5, Table II · p. 5

D. Position relative to existing Arabic resources

  • Casablanca offers country-labeled multidialect data, CARDAMOM fine-grained regional and sociolinguistic annotation, and Hamsa Emirati ASR resources –, while YODAS v3 lacks reliable country-level labels.These resources occupy different design points: the first three prioritize transcription quality and linguistic detail at tens-of-hours scale, while YODAS v3 reports roughly 25,000 hours of Arabic.Source: D. Position relative to existing… · p. 5, Table III · p. 6
  • A few dozen manually transcribed hours train dialect-specific labeling models that turn larger web collections into country-level weak supervision, validated with independently trained recognizers.GigaSpeech 3 supplies scale for representation learning and robust ASR while retaining usable dialect identity; it is not a substitute for fine-grained gold corpora.Source: D. Position relative to existing… · pp. 5–6

IV. METHOD … C. Weak-label generation and selection

GigaSpeech 3 is built by generating and refining dialect-matched weak labels, then validating whether the resulting audio–text pairs provide useful supervision. Its final inventories cover 10 million segments and 39k hours across five dialects.

A. Overview

  • Corpus construction partitions dialect-matched web speech, generates labels with dialect-specific VietASR models, refines SAU and SYR through four noisy-student rounds, and validates the resulting audio–text pairs with new recognizers.The final refinement round produces the released SAU and SYR inventories and their r4 recognizers.Source: A. Overview · p. 6

B. VietASR dialect-labeling model construction

  • VietASR trains a labeled-only Iter0 Zipformer, then clusters its speech representations into ASR-biased targets that retain acoustic regularities shaped by the target dialect.The trained encoder learns speech–text alignment cues even when Iter0 is not yet competitive for every dialect.Source: B. VietASR dialect-labeling model… · p. 6
  • Stage 3 pretrains a fresh encoder on the unlabeled pool to predict cluster identities for masked spans of filter-bank features, steering its Zipformer representations toward ASR-relevant distinctions.The approach retains the computationally efficient Fbank front end and multi-rate Zipformer backbone.Source: B. VietASR dialect-labeling model… · p. 6
  • Stage 4 attaches a transducer predictor and joiner to the pretrained encoder and fine-tunes it on the same manual transcripts to produce the Iter1 labeling model.The Iter0–Iter1 comparison holds manual supervision fixed, and the reported labeling models use one SSL-plus-fine-tuning iteration.Source: B. VietASR dialect-labeling model… · p. 6

C. Weak-label generation and selection

  • Iter1 decodes only dialect-matched web channels excluded from labeled and evaluation sets; its hypotheses are normalized, paired with audio, and screened for duration, recognition score, duplication, and text quality.Each resulting example pairs an audio segment with a model-generated transcript rather than a manual reference.Source: C. Weak-label generation and selection · p. 6
  • Primary downstream Zipformers start independently and train only on weak-label pairs, so improvement over Iter1 indicates reusable training information beyond the labeling model’s pretrained parameters.The design distinguishes labeling-model accuracy from whether its generated transcripts are useful as a corpus.Source: C. Weak-label generation and selection · p. 6
  • 10 million segments and 39k hours make up the final corpus across five dialects, with SAU and SYR using r4-filtered inventories and the other dialects using selected inventories.Selection is dialect-specific rather than targeting a common number of hours, to account for differences in source composition and label reliability.Source: C. Weak-label generation and selection · p. 6

D. From-scratch training and iterative refinement · E. Multi-dialect training

The sections describe from-scratch weak-label Zipformer2 training, iterative noisy-student refinement for Saudi and Syrian Arabic, and a multi-dialect model that conditions recognition on dialect tokens. The multi-dialect experiment evaluates aggregate performance and MSA coverage, while differences from dialect-specific models are not interpreted as pure cross-dialect transfer.

D. From-scratch training and iterative refinement

  • The primary downstream Zipformer2 transducer is trained from random initialization on weak labels, with its encoder, predictor, and joiner learned together rather than inherited from Iter1.Dialect-specific models use one corpus partition and a dialect-specific output vocabulary, and can be reproduced without the labeling-model checkpoint.Source: D. From-scratch training and iterative… · pp. 6–7
  • SAU and SYR use noisy-student training, with r0 denoting the initial weak-label-trained recognizer; Iter0 and Iter1 instead identify labeling-model stages.The two numbering schemes refer to different stages of corpus construction.Source: D. From-scratch training and iterative… · p. 7
  • An NST round filters audio–text pairs, trains a recognizer on selected weak supervision, and regenerates hypotheses; r4 is the model after four complete rounds.The round index marks a refinement-and-training cycle, not an optimization epoch or a different architecture.Source: D. From-scratch training and iterative… · p. 7
  • The released SAU and SYR inventories are r4-filtered data containing 4,512 and 4,361 hours, respectively, and their NST r4 recognizers enter the main dialect-specific comparison.Refinement is limited to these dialects because the recipe assumes a sufficiently accurate labeling model; Table X traces improvement from r0 through r4.Source: D. From-scratch training and iterative… · p. 7

E. Multi-dialect training

  • The Multi-Dialect Zipformer pools five weak-label corpora with MGB-2, prepends the ground-truth dialect token to each training transcript, and predicts that token followed by the transcript.Dialect tokens extend a shared output vocabulary; training uses neither a separate LID classification head nor an LID loss.Source: E. Multi-dialect training · p. 7
  • At inference, the main evaluation conditions recognition on predicted dialect IDs, while ground-truth-ID decoding is a matched diagnostic for routing errors.The approach shares acoustic evidence across varieties while preserving dialect-conditioned output behavior; improving dialect identification itself is outside the study’s scope.Source: E. Multi-dialect training · p. 7
  • The multi-dialect experiment tests whether weak-label corpora support one recognizer with competitive aggregate performance and MSA coverage, but model-column differences are not a pure measure of cross-dialect transfer.The models also differ in data mixture, vocabulary, conditioning, and automatic-routing errors.Source: E. Multi-dialect training · p. 7

F. Cross-architecture transfer … C. Test sets and normalization

The experimental setup tests weak-label utility across model architectures and evaluation domains, using controlled comparison chains, fixed decoding and normalization procedures, and dialect-specific evaluation conditions.

F. Cross-architecture transfer

  • Qwen3-ASR-1.7B is fine-tuned on weak dialect transcripts plus manually labeled MSA, testing portability across architectures rather than serving as a controlled primary validation.The adapted model is evaluated with and without ground-truth LID guidance; the MSA condition does not test weak supervision alone.Source: F. Cross-architecture transfer · p. 7, B. Recognition systems · p. 7

A. Evaluation questions

  • Five research questions test ASR-biased SSL, weak-label corpus utility, multi-dialect recognition, iterative refinement and Qwen transfer, and generalization to external corpora.The questions distinguish validation on GigaSpeechBench from evaluation on independently released Arabic corpora.Source: A. Evaluation questions · p. 7
  • The comparison chain progresses from manual-only Iter0 to SSL-enhanced Iter1, then trains a weak-only system from scratch before broadening to multi-dialect and Qwen experiments.The weak-only system discards the Iter1 parameters, independently testing the generated labels.Source: A. Evaluation questions · p. 7

B. Recognition systems

  • Primary recognizers use the icefall Zipformer2 recipe, with dialect-specific 500-unit BPE tokenizers or a shared 1,500-unit vocabulary and automatically predicted LID for the multi-dialect model.Final Saudi and Syrian dialect-specific results use NST r4 models matched to their r4-filtered released inventories.Source: B. Recognition systems · p. 7

C. Test sets and normalization

  • GigaSpeechBench Arabic test sets provide the primary evaluation [7], with identical references for proposed and off-the-shelf systems and a shared Arabic normalizer for all hypotheses.Historical exploratory runs using different manifests or normalization paths are excluded from reported comparisons.Source: C. Test sets and normalization · pp. 7–8
  • External evaluation covers DZA and EGY in Casablanca, five varieties in MASC, and Saudi Arabic in SADA, with official development and test partitions pooled without model selection or tuning.The same final checkpoints, fixed decoding settings, and shared text normalizer are used; results remain separated by corpus–variety condition, including the three-minute MASC-DZA subset as limited evidence (Table IV).Source: C. Test sets and normalization · p. 8, Table IV · p. 8
  • WER is the primary metric, reported as a percentage; averages are unweighted means of listed WERs and are comparable only when evaluation coverage matches.The main comparison averages five dialects, LID and Qwen tables add MSA, and external evaluation averages eight corpus–variety conditions.Source: C. Test sets and normalization · p. 8

D. External systems and dialect-ID diagnostic · VI. RESULTS AND ANALYSIS · A. From manual labels to reusable weak supervision

The final corpus-trained systems improve over their labeling models across all five dialects, with four-round refinement for Saudi and Syrian Arabic. The section also describes the external-system comparison and matched-model diagnostics for dialect-ID guidance and duration-stratified LID accuracy.

D. External systems and dialect-ID diagnostic

  • The external comparison includes commercial APIs, multilingual recognizers, and public Arabic-capable models, using the strongest complete system as the principal reference rather than mixing systems across test sets.Best third-party results by dialect are italicized; bold and underlining mark the best and second-best overall values.Source: D. External systems and dialect-ID… · p. 8
  • The averaged Multi-Dialect Zipformer uses predicted variety tags in the main comparison, while a paired diagnostic measures WER with and without ground-truth LID guidance under the same model and test set.The classifier’s confusion matrix is reported, and its WERs characterize routing cost; dialect identification itself is not claimed as a contribution.Source: D. External systems and dialect-ID… · p. 8
  • Duration-stratified LID accuracy is evaluated with the averaged model, greedy decoding, and predicted IDs across utterances under 3 seconds, 3–less than 10 seconds, and at least 10 seconds.Training prepends the ground-truth dialect token to transcripts so LID and ASR are learned jointly; sample counts accompany accuracy by variety and duration bucket.Source: D. External systems and dialect-ID… · p. 8

A. From manual labels to reusable weak supervision

  • ASR-biased SSL followed by labeled fine-tuning improves every dialect over labeled-only Iter0, with relative WER reductions of 28.1% for DZA, 26.2% for IRQ, 27.6% for EGY, 74.3% for SYR, and 1.4% for SAU.SAU starts from a 13.17 WER; the passage attributes the smaller SSL gain to its stronger supervised baseline.Source: A. From manual labels to reusable weak… · p. 8
  • Final Dialect-Specific Zipformers improve on Iter1 across all five dialects; WERs are 33.64 for DZA, 29.99 for IRQ, 29.32 for EGY, 12.12 for SAU, and 14.25 for SYR.SAU and SYR include four NST rounds and use released r4-filtered inventories of 4,511.62 and 4,360.75 hours, respectively.Source: A. From manual labels to reusable weak… · p. 8, Table V · p. 9
  • VietASR uses limited manual transcripts to build labeling models, while downstream recognizers train on weakly labeled audio–text pairs without using those transcripts or inheriting Iter1’s SSL-adapted parameters.This supports the corpus as reusable supervision; SAU and SYR releases additionally use refinement chains.Source: A. From manual labels to reusable weak… · p. 8

B. Comparison with third-party systems · C. Cross-corpus validation · D. Dialect-specific versus multi-dialect training

GigaSpeech 3-trained Zipformers achieve a lower five-dialect macro-average WER than the strongest complete third-party system and the lowest WER among six evaluated systems across eight external conditions. Dialect-specific and multi-dialect models show complementary strengths, while the multi-dialect model also supports MSA.

B. Comparison with third-party systems

  • 23.9%, 26.8%, and 21.0% lower WER than the best third-party values on DZA, SAU, and EGY, respectively, is achieved by the final Dialect-Specific Zipformers.Under automatic routing, Qwen3.5-Omni-Plus remains stronger on IRQ and SYR (28.54 and 13.76 versus 28.85 and 16.33); ground-truth LID lowers these to 27.84 and 13.18, indicating that routing errors explain part of the gap. The compact Zipformers' macro-average advantage does not mean they lead on every dialect.Source: B. Comparison with third-party systems · p. 9

C. Cross-corpus validation

  • 32.63 eight-condition macro-average WER with ground-truth LID is worse than 31.95 with predicted-LID decoding; most degradations occur on MASC, whose dialect subsets are relatively small.The very small MASC-DZA subset warrants caution; limited evaluation support and domain-specific LID calibration may contribute, requiring further study on larger out-of-domain sets.Source: C. Cross-corpus validation · p. 9
  • 37.78 macro-average WER after GigaSpeech 3 fine-tuning is 19.8% below Qwen3-ASR-1.7B's 47.10, with relative reductions of 6.9%–31.0% across all eight conditions.The adapted model beats Whisper-CS in all eight conditions and QwenCleo in seven; these comparisons include only six models decoded under a fixed protocol, so they do not establish state of the art on the external corpora.Source: C. Cross-corpus validation · p. 9

D. Dialect-specific versus multi-dialect training

  • 25.51 five-dialect macro-average WER for the automatically routed Multi-Dialect Zipformer trades accuracy for unified coverage against 23.86 for Dialect-Specific Zipformers, while recognizing MSA at 10.83 WER.This is not a controlled cross-dialect-transfer comparison: the systems differ in vocabulary, sampling mixture, MSA coverage, dialect conditioning, and NST refinement for dialect-specific SAU/SYR models; the multi-dialect model also incurs automatic-routing errors.Source: D. Dialect-specific versus… · p. 9

B. Comparison with third-party systems

  • 23.86 macro-average WER for the final Dialect-Specific Zipformers is a 16.6% relative reduction from Qwen3.5-Omni-Plus at 28.62; automatic Multi-Dialect routing reaches 25.51, and ground-truth LID reaches 23.62.The dialect-specific aggregate includes NST r4 models for SAU and SYR.Source: B. Comparison with third-party systems · p. 9, Table VI · p. 10

C. Cross-corpus validation

  • 31.95 macro-average WER for Dialect-Specific Zipformers and 34.39 for Multi-Dialect Zipformer accompany the lowest WER among six evaluated systems in every external condition.Multi-Dialect is strongest on four of five MASC subsets and has a MASC macro-average of 27.21 versus 32.64 for Dialect-Specific; Dialect-Specific is stronger on MASC-SAU, both Casablanca subsets, and SADA, so the configurations' ordering depends on corpus and variety.Source: C. Cross-corpus validation · p. 9, Table VII · p. 10

E. Effect of ground-truth LID guidance · F. Iterative refinement

Four rounds of noisy-student refinement lower WER for the final Saudi and Syrian recognizers, while the Multi-Dialect Zipformer’s predicted-LID routing raises macro-average WER relative to ground-truth guidance. The analyses also show that automatic dialect identification remains a bottleneck, especially for short utterances.

E. Effect of ground-truth LID guidance

  • Predicted LID raises macro-average WER from 23.62 to 25.51 across the five target dialects compared with ground-truth LID guidance.EGY and SYR have the largest increases, 3.64 and 3.15 points, while SAU changes by 0.19 points. The “w/o LID” row uses predicted LID; the paired “w/ LID” row changes only the LID guidance.Source: E. Effect of ground-truth LID guidance · p. 10
  • Automatic LID classifies only 58–74% of dialectal utterances correctly, with DZA, EGY, and SAU often routed to MSA and SYR also confused with IRQ.MSA is identified almost perfectly, but the routing errors show that LID accuracy alone does not determine the size of the recognition penalty.Source: E. Effect of ground-truth LID guidance · p. 10, Table VIII · p. 11
  • For utterances below 3 seconds versus 3–10 seconds, LID accuracy rises from 39.8% to 85.0% for DZA, 51.5% to 84.6% for IRQ, and 8.4% to 76.0% for SAU.Short-utterance confusion matrices show 69.3% of short SAU, 62.2% of short SYR, and 52.4% of short DZA utterances labeled MSA; the authors note that the analysis does not establish the cause.Source: E. Effect of ground-truth LID guidance · pp. 10–11, Table IX · p. 11, Table VIII · p. 11
  • Longer speech does not uniformly resolve confusion: DZA and IRQ improve above 10 seconds, EGY falls from 65.6% to 54.1%, and SYR stays near 61% in middle and long buckets.SAU peaks in the middle duration bucket, so duration alone does not explain its or SYR’s aggregate recognition results.Source: E. Effect of ground-truth LID guidance · p. 11
  • Duration-stratified LID patterns are associations across different utterances, not a controlled test of longer speech, and they do not directly quantify ASR degradation.Paired WER measurements with and without ground-truth LID guidance within each duration bucket would be needed to localize recognition costs; notably, SAU’s poor short-utterance LID coexists with only a 0.19-point aggregate WER difference.Source: E. Effect of ground-truth LID guidance · pp. 10–11
  • Ground-truth LID improves average WER over predicted-LID decoding in this paired Zipformer diagnostic, but the experiment does not compare dialect conditioning with no dialect input.Improving LID is model- and taxonomy-dependent and outside the paper’s scope.Source: E. Effect of ground-truth LID guidance · p. 11

F. Iterative refinement

  • WER falls at every refinement round, from 13.13 to 12.12 for SAU and 14.95 to 14.25 for SYR after four rounds.These are absolute reductions of 1.01 and 0.70 WER points (7.7% and 4.7% relative to r0); the largest single-round gains occur at r1→r2 for SAU and r0→r1 for SYR. The r4 models are the final models used in the main comparison; released SAU and SYR weak-label inventories use r4-filtered data, unlike the initially selected DZA, IRQ, and EGY data. SAU’s r0 WER of 13.13 is slightly higher than Iter1’s 12.98, so the reductions are relative to r0, not Iter1.Source: F. Iterative refinement · p. 11

G. Cross-architecture transfer … B. Scale, diversity, and label noise

Weakly labeled dialect speech transfers beyond the Zipformer recipe, but its benefits depend on language-identification routing and vary across external domains. The results also show a tradeoff between transcription accuracy and coverage, with data selection and refinement shaping the available scale and diversity.

G. Cross-architecture transfer

  • Qwen adaptation improves all five target dialects and outperforms QwenCleo and Whisper-CS on every reported dialect, though those public fine-tunes are not data-matched controls.The MSA improvement uses manually labeled MGB-2 speech, so it does not establish weak-label-only MSA training.Source: G. Cross-architecture transfer · pp. 11–12, Table XI · p. 12
  • Predicted-LID Qwen adaptation lowers macro-average WER from 38.42 to 28.28 (26.4% relative); ground-truth LID lowers it to 24.10 (37.3% relative).Ground-truth LID improves every reported dialect over predicted-LID routing, with the largest absolute reductions on DZA and EGY.Source: G. Cross-architecture transfer · p. 11, Table XI · p. 12

A. What the experiments establish

  • Independently initialized recognizers trained on generated audio–text pairs outperform Iter1 on all five dialects, testing the corpus beyond VietASR’s model parameters.DZA, IRQ, and EGY use selected weak labels; SAU and SYR use r4-filtered data.Source: A. What the experiments establish · p. 12

B. Scale, diversity, and label noise

  • Weak labels trade transcription accuracy for coverage: selection controls noisy hypotheses, while thousands of dialect-matched hours provide diversity unavailable in the 51–63-hour manual sets.Volume alone does not explain performance: SAU starts with the largest pool but retains the smallest fraction after refinement, while SYR has the largest SSL gain despite similar retention. Labeling accuracy, source match, diversity, and residual noise remain intertwined, requiring controlled scaling and quality-stratification experiments to separate their effects.Source: B. Scale, diversity, and label noise · p. 12
  • Final Dialect-Specific Zipformers lead on DZA, EGY, and SAU, while Qwen3.5-Omni-Plus remains stronger than automatic routing on IRQ and SYR.Ground-truth LID guidance narrows the IRQ and SYR gaps, indicating routing explains part of the difference.Source: B. Scale, diversity, and label noise · p. 12
  • The Multi-Dialect Zipformer is stronger on four MASC subsets, while Dialect-Specific models lead on both Casablanca subsets and SADA.Given the six-model comparison and the three-minute MASC-DZA subset, the results support generalization rather than a state-of-the-art claim.Source: B. Scale, diversity, and label noise · p. 12

C. Dialect conditioning as an interface … X. ETHICS STATEMENT

GigaSpeech 3 turns web speech into reusable weak supervision for Arabic dialect ASR, with reported gains across dialects and external domains. The sections also describe routing and data limitations, release provisions, and research-only ethics conditions.

C. Dialect conditioning as an interface

  • Ground-truth dialect IDs reduce five-dialect macro-average WER from 25.51 to 23.62 for the same averaged Multi-Dialect checkpoint.Source metadata may provide IDs in curated settings, but open-audio routing remains unresolved and should be evaluated separately from ASR.Source: C. Dialect conditioning as an interface · p. 12
  • Short dialectal speech is often routed to MSA, and longer inputs do not eliminate confusion; country tags also collapse meaningful within-country variation.The passage recommends duration-specific LID interpretation and finer evaluation and adaptation using resources such as CARDAMOM, Casablanca, and Hamsa.Source: C. Dialect conditioning as an interface · p. 12

VIII. CORPUS RELEASE PLAN AND LIMITATIONS

  • The corpus, weak transcripts, metadata, construction and training recipes, and online demo are released for research and evaluation under a research-only license.The corpus includes selected audio, weak transcripts, dialect labels, and segment metadata subject to source-platform terms; code has a separate software license.Source: §VIII · p. 12
  • A community proofreading platform lets volunteers propose transcript corrections, but the current release has not been exhaustively human-verified.Corrections are intended for future corpus revisions.Source: §VIII · p. 12
  • Limitations include routing errors, WER conflation of recognition errors with spelling variation, uncontrolled weak-label factors, and potential web-data biases.Numerical expressions and code-switching remain limited; future work should add targeted data and evaluation that separates recognition from formatting errors and checks foreign-language content preservation.Source: §VIII · p. 13

IX. CONCLUSION

  • 23.86 macro-average WER on GigaSpeechBench is 16.6% below the strongest complete third-party system for the final Dialect-Specific Zipformers.The models improve over their labeling models on all five dialects; across eight external conditions, a GigaSpeech 3-trained Zipformer leads six evaluated models, and corpus adaptation improves Qwen3-ASR in all eight.Source: §IX · p. 13

X. ETHICS STATEMENT

  • Collected audio comes from Creative Commons materials; personally identifiable information is removed with rule-based scripts, and annotators are fairly compensated.The dataset is released only for non-commercial research and education, with ongoing maintenance for identified risks.Source: §X · p. 13
Loading 2611.02476v2…