InaGVAD : a Challenging French TV and Radio Corpus annotated for Speech Activity Detection and Speaker Gender Segmentation

# inaGVAD : a Challenging French TV and Radio Corpus annotated for Speech Activity Detection and Speaker Gender SegmentationLast update : 2024/05/14

## IMPORTANT ACRONYMSVAD : Voice Activity DetectionSGS : Speaker Gender SegmentationIER : Identification Error RateWSTP : Women Speaking Time PercentageUEM : NIST Unpartitioned Evaluation Map (UEM) format define audio regions that systems must process

## Detailed Corpus description

inaGVAD corpus has been fully described in the following paper :

@inproceedings{inagvad2024,title={InaGVAD : a Challenging French TV and Radio Corpus annotated for Speech Activity Detection and Speaker Gender Segmentation},author={David Doukhan and Christine Maertens and William {Le Personnic} and Ludovic Speroni and Reda Dehak},booktitle={Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING)},year={2024},}

## Download instructions* Evaluation code is hosted on github : https://github.com/ina-foss/InaGVAD* Corpus audio files can be donwloaded through INA's website : https://www.ina.fr/institut-national-audiovisuel/research/dataset-project

## Filetree

inaGVAD|- inaGVAD_devset.lst : list of 60 file identifiers to be used as dev set|- inaGVAD_testset.lst : list of 217 file identifiers to be used as test set|- METADATA.csv : meta data associated to each annotated file:    * fileid : file identifier with syntax --    * media : "tv" or "radio" (inferred from fileid)    * channel_code : 3 characters corresponding to INA's channel code (inferred from fileid)    * channel_name : full channel name (human friendly)    * channel_category : 4 options : "music_radio" (challenging), "generalist_radio" (easy), "news_tv" (easy), "generalist_tv" (challenging)    * broadcast_hour : broadcast time hour (inferred from fileid)    * set : "dev" if in the dev set, "test" if in the test set|- dev : audio files and annotations corresponding to dev set|- test : audio files and annotations corresponding to test set

Files in dev and test directories correspond to 1-minute long recordings, with syntax . .e.g. : tv-F24-042912.sgs.uem

5 extensions are provided : * .wav : 16 000 Hz mono wav audio file* .vad.csv : Voice Activity Detection annotations in csv with fields start time (seconds), stop time (seconds) and label in {"speech", "non speech"}* .sgs.csv : Speaker Gender Segmentation annotations in csv with fields start time (seconds), stop time (seconds) and label in {"male", "female"}. To be used for training speaker gender prediction systems or obtaining IER SGS evaluation metrics. Warning : several phenomena are discared from these annotation exports : non speech events, speakers with unknown gender, overlapped speech involving speakers with different perceived gender.* .sgs.uem : portions of annotated files to be used for obtaining IER SGS evaluation metrics : excluding speakers with unknown percieved gender or overlapped speech involving speakers with different genders.* detailed.csv : detailed CSV annotations used to obtain WSTP evaluation metrics and corpus description statistics with fields :    * label (full annotated label)    * start time (seconds)    * stop time (seconds)    * voice_activity (True if spoken voice, else False)    * overlap (True if speech overlap, else False)    * gender in {"male", "female", "undefgender" (no percieved gender or speech overlap between speakers of different gender)}    * age in {"adult", "senior", "child", "undefage" (no percieved age or speech overlap between speakers of different ages)}    * speech_quality in {"standard", "onomatopoeia", "atypical", "undefquality" (speech overlap between different speech qualities)}

## Citing

Please cite this work if using this corpus in a publication :

@inproceedings{inagvad2024,title={InaGVAD : a Challenging French TV and Radio Corpus annotated for Speech Activity Detection and Speaker Gender Segmentation},author={David Doukhan and Christine Maertens and William {Le Personnic} and Ludovic Speroni and Reda Dehak},booktitle={Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING)},year={2024},}

## Questions or issues

Questions or issues related to this corpus could be asked from the github repository (best): https://github.com/ina-foss/InaGVAD/issuesAnother option is to contact David Doukhan : [e-mail]

## AcknowledgementsThis work has been partially funded by the French National Research Agency (project Gender Equality Monitor - ANR-19-CE38-0012).