MUSAN

109-hour corpus of music, speech, and environmental noise designed for data augmentation in speech and speaker recognition experiments.

• Task: augmentation
• Languages: en
• Hours: 109
• Domain: music, speech, noise
• License: CC BY 4.0
• Homepage: https://www.openslr.org/17
• Paper: https://arxiv.org/abs/1510.08484

Recommendation

The standard noise/music augmentation bank for speaker verification and robust ASR training. Combine with RIR-based room simulation for a full augmentation pipeline. Small enough to keep fully in RAM during training. No VoxKitchen pipeline template — augmentation is embedded in task-specific pipelines such as examples/pipelines/noise-augment.yaml.

Getting the data

Downloadable via VoxKitchen (musan, source: openslr, size: 10.3 GB):

vkit docker download --tag slim musan --root ./data/musan

Subsets: musan.