English Speech Recognition Dataset With Expert Annotations

SPGISpeech is a monolingual English dataset for automatic speech recognition tasks. The dataset is categorized as containing between 1 million and 10 million data instances. It was created by the author 'kensho' and was last updated in January 2026.

Use Cases

• Train an automatic speech recognition model on a large-scale English audio dataset.
• Fine-tune speech-to-text models using expert-generated text annotations.
• Analyze speech patterns and transcription accuracy within a monolingual (English) context.
• Benchmark ASR model performance on data sourced from a US region.

Strengths

• Dataset size is categorized between 1 million and 10 million instances, indicating substantial scale.
• Annotations are expert-generated, suggesting high-quality labels for the speech recognition task.
• The dataset is specifically designed for automatic speech recognition, providing a clear task focus.

Limitations

• Specific details on data instances, columns, splits, and file formats are unavailable, limiting preliminary assessment.
• The dataset is monolingual (English) and regionally focused (US), which may limit generalizability to other languages or accents.
• The raw description lacks concrete numbers for rows, columns, and audio/text specifics, making precise evaluation difficult.

Provenance

Source

kensho on Hugging Face.

Collection Method

Source data is original, with expert-generated annotations.

Freshness

Last updated on 2026-01-20.

Geography

Region is specified as 'us'.