GigaSpeech 3: Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data
Dialect-specific models cut average word error by 16.6% relative to the strongest outside system covering all five Arabic dialects.
GigaSpeech 3 contains 39,000 hours and 10 million weakly transcribed segments spanning Algerian, Egyptian, Iraqi, Saudi and Syrian Arabic. The researchers started with 60 hours of manual transcription per dialect to train models that labeled dialect-matched web audio. Their final dialect-specific models reached a macro-average word error rate of 23.86% on GigaSpeechBench, beating the best third-party result on three dialects. The authors are releasing the corpus and training recipes.
HF Daily Papers' note
GigaSpeech 3 contains 39,000 hours and 10 million weakly transcribed segments spanning Algerian, Egyptian, Iraqi, Saudi and Syrian Arabic. The researchers started with 60 hours of manual transcription per dialect to train models that labeled dialect-matched web audio. Their final dialect-specific models reached a macro-average word error rate of 23.86% on GigaSpeechBench, beating the best third-party result on three dialects. The authors are releasing the corpus and training recipes.
HF Daily Papers' note
score 5