LoquaciousSet: 25,000 Hours of Heterogeneous English Speech Recognition Data

25,000 hours of transcribed English speech form the core of this dataset for automatic speech recognition research. The collection includes read and spontaneous speech in both clean and noisy acoustic conditions, organized into subsets of varying size. SpeechBrain authored the dataset, which was last updated on February 11, 2026.

Use Cases

• Training large-scale English speech recognition models based on the 25,000-hour 'large' subset.
• Benchmarking ASR model robustness based on the described mix of clean and noisy audio conditions.
• Developing models for specific speech types based on the described read and spontaneous speech categories.
• Creating smaller, curated training sets based on the 2,500-hour 'medium' or 250-hour 'small' subsets.

Strengths

• The 'large' subset contains 25,000 hours of transcribed speech, providing substantial scale for training.
• The dataset explicitly includes heterogeneous conditions: both read and spontaneous speech, and both clean and noisy audio.
• It offers tiered subsets ('large', 'medium', 'small') with different total durations, likely facilitating different research needs.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file formats are unknown, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

SpeechBrain

Freshness

Last updated 2026-02-11 14:31:57; freshness should be verified.