LibriReplay-DOA (Anonymous Submission)

Overview

LibriReplay-DOA is a multi-channel multi-speaker replay dataset designed for evaluating robust speech processing systems under realistic room playback conditions.

The dataset contains replay recordings captured in real rooms under multiple playback configurations (DOA settings). Each session includes multiple overlapping speakers.

This dataset is released for peer-review purposes.

---

Dataset Structure

The dataset contains a single evaluation split:

• test

Each sample corresponds to:

• one playback configuration (location),
• one session,
• one speaker within the session.

Data Fields (test.csv)

Each row in test.csv contains:

• id: unique sample identifier
• location: playback configuration
• session_key: session identifier
• speaker_id: speaker identity within session
• recorded_path: path to replayed mixture recording
• clean_path: path to clean source signal of the speaker
• asr_label_path: transcription file
• speaker_activity_path: speech activity time segments file

---

Recording Conditions

• Real room playback recordings
• Multiple playback positions (DOA configurations)
• Multi-speaker overlapping speech
• Clean source audio available per speaker

Further technical details will be provided after peer review.

---

Usage Agreement

This dataset is released for research and peer-review purposes only.

By downloading or using this dataset, you agree:

• Not to redistribute the dataset.
• Not to use the dataset for commercial purposes.

The dataset will be officially released after peer review.

---

License

License will be specified upon official release.

---

Citation

Anonymous submission.