MiniLibriMix: Clean and Noisy Two-Speaker Speech Mixtures

MiniLibriMix is a small demonstration version of the LibriMix dataset created by Cosentino Joris of the Centre National de la Recherche Scientifique. It contains 800 training mixtures and 200 validation mixtures. Each set includes folders for clean two-speaker mixtures, mixtures with noise, and the raw source signals for each speaker and noise.

Use Cases

• Training speech source separation models based on the provided clean and noisy two-speaker mixtures.
• Benchmarking model performance on a small-scale dataset based on the described train/validation split.
• Demonstrating audio preprocessing pipelines using the raw signal folders for speakers and noise.
• Developing data loaders and utilities for audio machine learning tasks based on the included metadata.

Strengths

• Contains a defined split of 800 training and 200 validation mixtures.
• Provides organized folders for clean mixtures, noisy mixtures, and raw source signals.
• Includes metadata to facilitate data loading.

Limitations

• Row count and total file size are unknown, which may limit suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.

Provenance

Source

Centre National de la Recherche Scientifique

Collection Method

Derived from the LibriMix dataset for demonstration purposes.