LibriMix: An Open-Source Dataset for Generalizable Speech Separation

By Joris Cosentino, Manuel Pariente, Samuele Cornell, Antoine Deleforge, Emmanuel Vincent. 2020

428 citations recorded by the source corpus.

In recent years, wsj0-2mix has become the reference dataset for single-channel speech separation. Most deep learning-based speech separation models today are benchmarked on it. However, recent studies have shown important performance drops when models trained on wsj0-2mix are evaluated on other, similar datasets. To address this generalization issue, we created LibriMix, an open-source alternative to wsj0-2mix, and to its noisy extension, WHAM!. Based on LibriSpeech, LibriMix consists of two- or three-speaker mixtures combined with ambient noise samples from WHAM!. Using Conv-TasNet, we achieve competitive performance on all LibriMix versions. In order to fairly evaluate across datasets, we introduce a third test set based on VCTK for speech and WHAM! for noise. Our experiments show that the generalization error is smaller for models trained with LibriMix than with WHAM!, in both clean and noisy conditions. Aiming towards evaluation in more realistic, conversation-like scenarios, we also release a sparsely overlapping version of LibriMix's test set.

Subfields: Artificial Intelligence, Signal Processing

View Paper: https://www.semanticscholar.org/paper/8793b3e9b1d2b2b448b343edb9e9a8a8c3553a92

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Hung-yi Lee
2. Shinji Watanabe
3. Wei-Ning Hsu
4. Jinyu Li
5. James R. Glass
6. Abdel-rahman Mohamed
7. Dong Yu
8. Shang-Wen Li

In Signal Processing

1. Shinji Watanabe
2. Jinyu Li
3. Hung-yi Lee
4. Wei-Ning Hsu
5. Dong Yu
6. Zhuo Chen
7. Deliang Wang
8. James R. Glass

Citations in the corpus

• AISHELL-4: An Open Source Dataset for Speech Enhancement, Separation, Recognition and Speaker Diarization in Conference Scenario (2021, methodology): “LibriMix [16], derived from LibriSpeech [17], is an alternative to datasets mentioned above for noisy speech separation task.”
• Separate Anything You Describe (2023, positive): “Specifically, following [64], we first uniformly sampled the loudness of the two audio samples between - 35 dB and - 25 dB LUFS (Loudness Units Full Scale); then we mixed the signals together.”
• Generative Pre-training for Speech with Flow Matching (2023, methodology): “We tested the fine-tuned model on LibriMix (Cosentino et al., 2020) 16khz min.”
• Wavesplit: End-to-End Speech Separation by Speaker Clustering (2020, neutral): “We also perform experiments on the recently released LibriMix dataset [60].”
• SUPERB: Speech processing Universal PERformance Benchmark (2021, methodology): “LibriMix [27] is adopted where LibriSpeech train-clean-100/dev-clean/test-clean are used to generate mixtures for training/validation/testing.”
• SUPERB-SG: Enhanced Speech processing Universal PERformance Benchmark for Semantic and Generative Capabilities (2022, neutral): “We investigate the speech separation problem on a dataset simulated from LibriSpeech (Cosentino et al., 2020) (CC BY 4.”