Wiki En Asr Adapt: Wikipedia Text for ASR Model Adaptation

English Wikipedia text and ASR error data presented in an ASRU-2023 paper. It contains 4.3 million unique words or phrases from Wikipedia titles occurring in 33.8 million paragraphs, plus 26 million phrase pairs representing ASR recognition errors. The dataset was created by bene-ges and last updated on Hugging Face in December 2023.

Use Cases

• Training ASR error correction models based on the 26 million phrase pairs of corrupted and correct text.
• Analyzing common ASR error patterns using the inventory of phrase corruptions.
• Building language models or text augmentation pipelines using the large-scale Wikipedia paragraph corpus.
• Benchmarking ASR adaptation techniques on a known, published corpus from ASRU-2023.

Strengths

• Large scale with 33.8 million English Wikipedia paragraphs.
• Includes a substantial inventory of 26 million phrase pairs for ASR error analysis.
• Associated with a peer-reviewed ASRU-2023 conference paper.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the main data files is unknown, which may limit suitability assessment.
• Freshness should be verified as the last update was 2023-12-14.

Provenance

Source

English Wikipedia

Collection Method

Likely extracted and processed from Wikipedia dumps for ASR research.

Freshness

Last updated 2023-12-14 10:59:19