RIR-Mega-Speech: 100K+ Reverberant Utterances with Acoustic Metadata

RIR-Mega-Speech is a synthetic corpus of 100,000 to 1,000,000 reverberant English speech utterances created by mandipgoswami and released in 2026. The dataset was generated by convolving LibriSpeech audio with simulated room impulse responses from the RIR-Mega collection and includes detailed acoustic metadata for each file.

Use Cases

• Training Automatic Speech Recognition (ASR) systems to improve accuracy in reverberant conditions
• Developing audio-to-audio speech enhancement models using the synthetic reverberant files
• Correlating acoustic metadata metrics with speech classification error rates

Strengths

• Scale of 100,000 to 1,000,000 utterances
• Includes per-file acoustic metadata for controlled analysis
• Reproducible generation process using RIR-Mega and LibriSpeech

Limitations

• Synthetic reverberation based on simulated room impulse responses rather than physical recordings
• Restricted to English language
• Non-commercial license (CC BY-NC 4.0) restricts usage

Provenance

Source

Derived from LibriSpeech and RIR-Mega collection

Collection Method

Synthetic convolution of speech utterances with simulated room impulse responses

Freshness

Last updated March 2026.

Geography

United States