Urgent2025 SQA: Noisy and Enhanced Speech with Quality Metrics

The URGENT Speech Enhancement Challenge dataset provides noisy and enhanced speech samples curated for SQA/SE research. Each entry includes audio/IDs with objective and model-predicted quality metrics, and human Mean Opinion Scores collected from 8 distinct subjects via Amazon Mechanical Turk. The dataset was created by urgent-challenge and last updated on December 10, 2025.

Use Cases

• Training speech enhancement models based on paired noisy and enhanced audio samples.
• Evaluating speech quality assessment algorithms using objective and model-predicted metrics.
• Benchmarking models against human Mean Opinion Scores (MOS) collected via crowdsourcing.
• Conducting research on the correlation between automated metrics and human perception of speech quality.

Strengths

• Includes human MOS labels collected from 8 distinct subjects per sample.
• Provides a suite of objective and model-predicted quality metrics for each audio sample.
• Specifically curated for the Interspeech 2025 URGENT Speech Enhancement Challenge.

Limitations

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

Provenance

Source

URGENT Speech Enhancement Challenge (Interspeech 2025)

Collection Method

Human MOS collected via Amazon Mechanical Turk (MTurk) platform following the P.808 standard.

Freshness

Last updated 2025-12-10 04:07:09; freshness should be verified.