# Vibravox: French speech recorded with body-conduction sensors

## Facts

- Provider: Julien Hauret, Malo Olivier, Thomas Joubaud, Christophe Langrenne, Sarah Poiree, Veronique Zimpfer, Eric Bavu (Cnam LMSSC, French-German Research Institute of Saint-Louis)
- License: CC-BY-4.0 (https://huggingface.co/datasets/Cnam-LMSSC/vibravox)
- Size: 360 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/Cnam-LMSSC/vibravox
- Access: public, readable without an account
- DOI: https://doi.org/10.57967/hf/2727
- Category: Speech & Audio
- Source: https://arxiv.org/abs/2407.11828
- Page: https://scigantic.com/archive/629b5463-e616-5ba6-9158-6bc83e2d4c32
- Keywords: Speech & Audio, speech recognition, Parquet, Hugging Face

## About

Vibravox is a general-purpose audio dataset of French speech captured simultaneously with body-conduction transducers and conventional microphones: a headset microphone near the mouth as reference, a forehead vibration sensor, a foam-embedded in-ear microphone, a rigid in-ear earpiece microphone, a temple vibration pickup and a laryngophone. Body-conducted speech is robust to ambient noise but band-limited and distorted, so the dataset is used for automatic speech recognition (speech-to-text and speech-to-phoneme), audio bandwidth extension, speech enhancement, speaker verification and identification, and voice cloning from non-conventional sensors. It has four subsets: speech_clean, speech_noisy (speech recorded in noise), speechless_clean and speechless_noisy (sensor responses without speech, for noise and physiological-sound studies). This repository holds them as four folders of Parquet shards, for example 255 files in speech_clean (measured 2026-11-14; the Hub reports about 564 GB). Load with datasets.load_dataset('Cnam-LMSSC/vibravox', 'speech_clean', streaming=True); a pip package 'vibravox' provides training code. Released under CC BY 4.0; cite Hauret et al., Speech Communication 2025.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("Cnam-LMSSC/vibravox")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `speech_clean/test-00000-of-00030.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/629b5463-e616-5ba6-9158-6bc83e2d4c32
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/629b5463-e616-5ba6-9158-6bc83e2d4c32
- Over MCP: `get_schema_card` with id `629b5463-e616-5ba6-9158-6bc83e2d4c32` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
- [Speech recognition datasets](https://scigantic.com/topics/speech-recognition)
