GigaSpeech 3 releases 39,000-hour weakly-labeled Arabic corpus

GigaSpeech 3 publishes 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi and Syrian Arabic, built against a gap that, as the paper on arXiv states, leaves "Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio."

The pipeline starts from 60 hours of manual transcription per dialect, trains labeling models with "speech-recognition-biased self-supervised learning," and applies them to dialect-matched web audio. Data selection plus "four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus."

On GigaSpeechBench, the dialect-specific models "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 beating the best third-party result on three dialects. A multi-dialect variant using self-predicted dialect identities lands at 25.51% macro WER and also handles Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer posts the lowest WER among six evaluated models in every condition, and adapting the Qwen3 speech recognizer on the corpus improves performance on all eight.

The authors say they "release the corpus and recipes as a foundation for dialect-aware Arabic speech processing." The abstract does not name the third-party systems compared against, nor does it state the license under which the corpus is distributed.