[Paper Review] Highly-Reverberant Real Environment database: HRRE

Juan Pablo Escudero, Víctor Poblete+6 more|arXiv (Cornell University)|Jan 29, 2018

Speech and Audio Processing5 references3 citations

TL;DR

The HRRE database provides 13.4 hours of speech data recorded in real, highly reverberant environments, simulating 20 distinct conditions across varying reverberation times and speaker-to-microphone distances. Generated by re-recording the Aurora-4 clean test set, it enables robust evaluation of speech recognition systems in realistic, challenging acoustic conditions.

ABSTRACT

Speech recognition in highly-reverberant real environments remains a major challenge. An evaluation dataset for this task is needed. This report describes the generation of the Highly-Reverberant Real Environment database (HRRE). This database contains 13.4 hours of data recorded in real reverberant environments and consists of 20 different testing conditions which consider a wide range of reverberation times and speaker-to-microphone distances. These evaluation sets were generated by re-recording the clean test set of the Aurora-4 database which corresponds to five loudspeaker-microphone distances in four reverberant conditions.

Motivation & Objective

• Address the lack of standardized evaluation datasets for speech recognition in highly reverberant real environments.
• Create a realistic benchmark to test the robustness of automatic speech recognition (ASR) systems under extreme acoustic conditions.
• Provide diverse testing conditions with controlled variations in reverberation time and distance between speaker and microphone.
• Support reproducible evaluation of ASR models in environments that closely mimic real-world reverberant spaces.
• Enable systematic comparison of signal processing and ASR techniques under extreme acoustic degradation.

Proposed method

• Recorded clean speech from the Aurora-4 test set in 13 real, highly reverberant rooms to simulate diverse acoustic conditions.
• Varied speaker-to-microphone distances across five levels and four distinct reverberation conditions to generate 20 unique testing scenarios.
• Preserved the original linguistic content and phonetic diversity of the Aurora-4 database to ensure consistency and validity.
• Used real environmental acoustics rather than simulated or synthetic reverberation to ensure authenticity and practical relevance.
• Produced a standardized, publicly available dataset to support benchmarking and reproducible research in reverberant ASR.
• Ensured data fidelity through careful calibration and recording procedures to maintain signal-to-noise ratio and intelligibility.

Experimental results

Research questions

• RQ1 How does speech recognition performance degrade under highly reverberant real-world conditions compared to anechoic or simulated environments?
• RQ2 To what extent do varying speaker-to-microphone distances and reverberation times affect ASR accuracy in real rooms?
• RQ3 Can a standardized, real-world database improve the reliability and reproducibility of ASR evaluation in reverberant settings?
• RQ4 How do existing ASR systems perform when tested on real, highly reverberant recordings versus synthetic or simulated data?
• RQ5 What are the key acoustic variables that most significantly impact recognition accuracy in real reverberant environments?

Key findings

• The HRRE database contains 13.4 hours of speech data collected in 13 real, highly reverberant rooms, providing a realistic benchmark for ASR evaluation.
• The dataset comprises 20 distinct testing conditions, combining five speaker-to-microphone distances and four reverberation levels.
• The database is derived from the Aurora-4 clean test set, ensuring linguistic and phonetic consistency with a widely used benchmark.
• Reverberation times in the dataset span a wide range, enabling evaluation across both moderate and extreme reverberant conditions.
• The use of real recordings, as opposed to synthetic reverb, provides greater acoustic fidelity and realism for evaluating robust ASR systems.
• The dataset is publicly available and designed for reproducible evaluation, supporting systematic comparison of ASR models under realistic degradation.