Oriel — AI & technology worth noticing, every day.

speechcolab/gigaspeech3

Dataset · https://huggingface.co/datasets/speechcolab/gigaspeech3

GigaSpeech 3 Scaling Multi-Dialect Arabic Speech Recognition with Weakly Labeled Data GigaSpeech 3 is a large-scale Arabic speech recognition corpus containing approximately 10 million weakly transcribed speech segments across five dialects: Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. It supports research on dialect-specific and multi-dialect ASR, weak supervision, and low-resource speech recognition. Dataset: GigaSpeech 3 Paper: arXiv:2611.02476 Evaluation benchmark:… See the full description on the dataset page: https://huggingface.co/datasets/speechcolab/gigaspeech3.

First seen by Oriel on .

AI summary

GigaSpeech 3 is a large-scale corpus for dialect-specific and multi-dialect Arabic speech recognition, built from weakly transcribed web audio across five country-level dialects.

  • It contains approximately 10 million segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic, while human-transcribed evaluation data are provided separately in GigaSpeechBench.
  • Dialect-specific labeling models use VietASR-style ASR-biased self-supervised learning and 51–63 manually transcribed hours per dialect; SAU and SYR receive four noisy-student refinement rounds.
  • The manuscript reports 23.86% macro-average WER for final dialect-specific Zipformers on GigaSpeechBench and evaluates transfer to MASC, Casablanca, and SADA.
  • Each dialect provides Parquet metadata and multi-gigabyte tar shards; metadata streaming does not download or decode the referenced audio.
  • Machine-generated transcripts may contain errors and biases, dialect labels are coarse, and the corpus is restricted to scientific research and evaluation without commercial-use permission.

#arabic asr #speech recognition #weak supervision #speech datasets

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Datasets: 32, latest observation ; +22 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Datasets: 15,595, latest observation ; +15,246 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Datasetsspeechcolab/gigaspeech3https://huggingface.co/datasets/speechcolab/gigaspeech3 Read discussion