Back home

GigaSpeech 3

hf 0.82
Visit source

Intro

GigaSpeech 3 is SpeechColab's large-scale multi-dialect Arabic ASR corpus: ~10.3M weakly transcribed segments / ~39k hours across five dialects — DZA (Algerian), EGY (Egyptian), IRQ (Iraqi), SAU (Saudi), SYR (Syrian). HF: https://huggingface.co/datasets/speechcolab/gigaspeech3 ; paper arXiv 2611.02476 (Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data); eval suite GigaSpeechBench; proofreading portal VoxLabel.

Data & method (official card / manuscript):

  • ~51–63 hours of manual transcripts per dialect train VietASR-style ASR-biased SSL labelers that then weakly label dialect-matched web audio;
  • SAU/SYR go through selection plus four noisy-student rounds;
  • Audio in per-dialect tar shards + Parquet metadata; HF footprint ~4.82 TB / ~10,271,066 rows;
  • Manuscript reports 23.86% macro-average WER for final dialect-specific Zipformers on GigaSpeechBench's five dialects;
  • Continues the numbered GigaSpeech / GigaSpeech 2 brand line.

Explain

Why it matters

Dialectal Arabic ASR has long lacked large public training resources. GigaSpeech 3 uses weak supervision to push five dialects to the ~10M-segment scale, with a public bench and proofreading portal.

License / use

Official terms: scientific research and evaluation only — no commercial-use grant; underlying recording rights stay with rights holders. Evaluate on independently transcribed data (GigaSpeechBench), not weak train labels as gold.

Limits

  • Weak labels include recognition/normalization errors; country-level dialect tags are coarse and may mix MSA/code-switching;
  • Web sourcing adds topic and demographic bias;
  • Individual shards can be multi-GB — download on demand via metadata.

Usage

Who it's for

  • Teams working on multi-dialect Arabic ASR, weak supervision and low-resource speech;
  • People benchmarking on GigaSpeechBench or transferring to MASC / Casablanca / SADA;
  • Annotation projects using VoxLabel for proofreading.

Getting started

  1. Read arXiv 2611.02476 and the HF card;
  2. Stream metadata first: load_dataset("speechcolab/gigaspeech3", "SAU", split="train", streaming=True);
  3. Pull matching tars with hf_hub_download on dialect/audio/{shard};
  4. Evaluate on GigaSpeechBench; do not use this release for commercial products under the stated terms.

Advertisement