GigaSpeech 3 Scales Arabic ASR Across Five Dialects with Weak Labels

Introduction

Arabic speech recognition faces a persistent resource imbalance. Modern Standard Arabic has benefited from more research and training data, while regional dialects remain difficult to model because dialect-resolved transcriptions are expensive and scarce. This gap matters in practice: conversational audio, media, and user-generated recordings are often dialectal rather than formal.

GigaSpeech 3 presents a data-centric response. Instead of relying on large-scale manual transcription, it starts with a small amount of trusted supervision and uses automatic labeling to turn dialect-matched web audio into a much larger training resource.

What the study contributes

• Five dialects in one corpus. The dataset covers Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. The study corpus contains approximately 39,000 hours and 10 million weakly transcribed speech segments.
• A small supervised seed. For each dialect, the pipeline begins with roughly 51–63 hours of manually transcribed speech. These samples are used to train labeling models before scaling to matched web audio.
• Speech-recognition-oriented representation learning. The approach uses self-supervised learning designed with ASR in mind, rather than treating generic speech representations as sufficient for dialectal transcription.
• Iterative noise control. Data selection and four rounds of noisy-student refinement are applied for Saudi and Syrian Arabic. The goal is to improve pseudo-label quality while expanding the training pool.
• Strong benchmark results. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects. Their macro-average word error rate is 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects. They also beat the best third-party result on three dialects.
• A multilingual alternative. A multi-dialect model using self-predicted dialect identities reaches a 25.51% macro-average word error rate and also supports Modern Standard Arabic.

Why it matters

The main contribution is not simply the headline number of hours. GigaSpeech 3 illustrates a repeatable strategy for low-resource speech: use a limited amount of reliable human supervision to train an initial labeler, retrieve or select relevant unlabeled audio, and then improve the system through carefully managed self-training. This helps break the cycle in which a dialect needs a strong recognizer to obtain labels, but needs labels to build that recognizer in the first place.

The reported cross-domain results strengthen the case. Across eight conditions from three external corpora, a Zipformer trained on GigaSpeech 3 achieves the lowest word error rate among six evaluated models in every condition. Adapting the Qwen3 speech recognizer with the corpus also improves results across all eight conditions.

These findings suggest practical value for dialect-aware transcription, voice assistants, and multilingual speech models. At the same time, weak labels are not equivalent to perfect human annotations: corpus selection, pseudo-label errors, and evaluation design still influence the outcome. The work is therefore best understood as a blueprint for scaling dialectal supervision, not as a replacement for careful data quality controls.

Source: Hugging Face Daily Papers