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

GigaSpeech 3. Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data. By Junzhe Liu and fifteen others, from Shanghai Jiao Tong University, the Shanghai Innovation Institute, and eight other institutions. Published on the third of November.

Web audio in dialectal Arabic is abundant. Transcribed speech resolved by dialect is not, and that scarcity is what holds back dialectal Arabic speech recognition. GigaSpeech 3 is a corpus built to close that gap, released together with recipes as a foundation for dialect-aware Arabic speech processing.

It holds thirty-nine thousand hours of audio, cut into ten million speech segments. The transcriptions are weak labels: they come from models rather than from people transcribing each segment by hand. Five dialects are covered: Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Each is handled as its own dialect, rather than being folded into a single undifferentiated Arabic.

The labels are bootstrapped from a small seed. For each dialect, the work starts from sixty hours of manual transcription. Labeling models are trained from that seed using self-supervised learning biased toward speech recognition, and are then applied to web audio matched to the dialect. Data selection follows, and for Saudi and Syrian Arabic, four rounds of noisy-student refinement, which together yield the final corpus.

The corpus is aimed at anyone building speech recognition for dialectal Arabic, whether one model per dialect or a single model across all of them. Its usefulness is checked through dialect-specific and multi-dialect training and through evaluation across domains. On GigaSpeechBench, the final dialect-specific models beat their own labeling models in all five dialects, reaching a macro-average word error rate of twenty-three point eight six percent. That is a sixteen point six percent relative reduction from the strongest third-party system covering all five dialects, and these models also beat the best third-party result on three dialects.

A single multi-dialect model, using dialect identities that it predicts itself, reaches a macro-average word error rate of twenty-five point five one percent, and also supports Modern Standard Arabic. On three external corpora, spanning eight conditions, a Zipformer trained on GigaSpeech 3 has the lowest word error rate among six evaluated models in every condition. Adapting the Qwen3 speech recognizer on the corpus improves it in all eight conditions as well.

The arc of the project runs from a small amount of human labeling to a far larger pool of web audio. Sixty hours per dialect become thirty-nine thousand hours overall, and the gains hold up when the models leave home and face other corpora.