ebellob/voxpopuli_spanish_enhanced

VoxPopuli Spanish Enhanced (CleanUNet + FlashSR)

Dataset Summary

This dataset is a processed and enhanced version of the Spanish subset of: facebook/voxpopuli. Furthermore, as this is a personal project, we give no guarantees that the audio is completely clean from any artifacts or noise the CleanUNet model could not remove. However, we have personally tested the corpus via the fine-tuning of some SOTA speech models and the results have been satisfactory.

In addition to ebellob/voxforge_spanish_enhanced, it has been created to support high-quality speech research, particularly for:

• —Text-to-Speech (TTS)
• —Automatic Speech Recognition (ASR)
• —Speech enhancement and robustness studies

due to the lack of big and clean speech datasets in Castilian (Peninsular) Spanish. The original VoxPopuli Spanish recordings are sourced exclusively from the European Parliament and therefore represent Peninsular Spanish only; however, they are extremely noisy, which motivated this enhanced release.

Find demo samples in: https://erikupv.github.io/CleanUNet-FlashSR-samples/

Main processing steps

1. 1.Speech denoising (CleanUNet) All audio samples are processed using CleanUNet, a neural speech enhancement model, to reduce background noise and artifacts while preserving speech content.

1. 1.Audio super-resolution (FlashSR) Enhanced audio is further processed using FlashSR to improve temporal and spectral resolution, resulting in cleaner and more detailed waveforms suitable for modern TTS and ASR pipelines.

The metadata structure is preserved from the original dataset unless stated otherwise.

Dataset Structure

Configurations

This dataset currently provides a single `train` split.

Note on splits: This dataset contains only the original `train` split from the VoxPopuli Spanish subset. The original dataset also provides separate validation and test splits, which are not included here. Users who require the original official validation or test sets for benchmarking should refer to the upstream facebook/voxpopuli dataset.

Dataset Features

• —audio_id: Unique audio segment identifier
• —language: Language code (es for Spanish)
• —audio: Enhanced audio waveform
• —raw_text: Original raw transcription
• —normalized_text: Normalized transcription (lowercased, punctuation normalized)
• —gender: Speaker gender (male / female)
• —speaker_id: Speaker identifier
• —is_gold_transcript: Whether the transcript is considered gold quality
• —accent: Accent label (if available; None for most Spanish recordings in the original corpus)

Dataset Size

• —Number of examples: 50,922
• —Approximate size: ~105 GB

Language

• —Spanish (es) — Castilian / Peninsular Spanish only

This dataset contains exclusively Peninsular Spanish audio from European Parliament plenary sessions. It does not include Latin American or other regional Spanish variants, making it particularly suitable for research requiring a clean, homogeneous Peninsular Spanish corpus.

Intended Uses

This dataset is suitable for:

• —High-quality Peninsular Spanish TTS training
• —ASR model training and evaluation
• —Speech enhancement and super-resolution research
• —Robustness studies on enhanced speech data

Out-of-Scope Uses

• —Speaker identification or biometric profiling
• —Any attempt to deanonymize speakers
• —Uses violating the original VoxPopuli license or European Parliament terms

Processing Notes

• —Audio waveforms are not bit-identical to the original VoxPopuli release.
• —Speech content and textual annotations are unchanged.
• —Processing may introduce minor spectral or temporal artifacts inherent to neural enhancement models.
• —The original dataset does not provide quality scores (e.g., mean_quality); therefore, no quality-based filtering was applied prior to enhancement.

License

This dataset inherits the license of the original VoxPopuli corpus:

• —Creative Commons Zero v1.0 Universal (CC0-1.0)

Please refer to the European Parliament's legal notice and the original dataset page for exact terms.

Citation

If you use this dataset, please cite both the original VoxPopuli paper and this processed version.

Original dataset

@inproceedings{wang-etal-2021-voxpopuli,
  title = {VoxPopuli: A Large-Scale Multilingual Speech Corpus for Representation Learning, Semi-Supervised Learning and Interpretation},
  author = {Wang, Changhan and Riviere, Morgane and Lee, Ann and Wu, Anne and Talnikar, Chaitanya and Haziza, Daniel and Williamson, Mary and Pino, Juan and Dupoux, Emmanuel},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
  month = {aug},
  year = {2021},
  address = {Online},
  publisher = {Association for Computational Linguistics},
  url = {https://aclanthology.org/2021.acl-long.80},
  pages = {993--1003}
}

This dataset

@dataset{voxpopuli_spanish_enhanced,
  title = {VoxPopuli Spanish Enhanced (CleanUNet + FlashSR)},
  author = {Erik Beltrán},
  year = {2026},
  note = {Enhanced with CleanUNet denoising and FlashSR super-resolution}
}

Tools used

• —CleanUNet — Kong et al., "Speech Denoising in the Waveform Domain with Self-Attention", ICASSP 2022. Paper: arXiv:2202.07790 · Code: NVIDIA/CleanUNet

• —FlashSR — Im & Nam, "FlashSR: One-step Versatile Audio Super-resolution via Diffusion Distillation", 2025. Paper: arXiv:2501.10807 · Code: ysharma3501/FlashSR

Acknowledgements

• —European Parliament for creating and sharing the original event recordings
• —Meta AI (Facebook) for the VoxPopuli curation and release
• —CleanUNet authors (Kong, Ping, Dantrey & Catanzaro)
• —FlashSR authors (Jaekwon Im & Juhan Nam)