Welcome to the first 

   HiFi Stereo Bird Cocktail Party Challenge 

                     *** 
          
            [SBCP Bird Clef 2021]
 
Herve' Glotin, Coraline Mounier, Noreen Blaukat

============

The TESTSET : 

It is composed of 512 * 20 sec chorus files (~ 30 min), in STEREO 192 kHz SR 16 bits.
It has been recorded during the lockdown spring 2020 into Domaine du Rayol, Côte d'Azur, France.
Microphones are super Cardioids, connected at the a Scarlet sound card. 
Micorphone aperture is 17 cm, 120 degres.
Recordings by H. Glotin

============

The TASK : 

You have to predict in each 5 second chunk of each test file, the presence or absence of each of the 85 species of the training set.
All the run must be fully automatic with depicted script to rerun for validation of the best run.

============

Your SUBMISSION : 

You will email your runs in .cvs format.
Each raw per file. One species per column x per chunck 0-5, 5-10, 10-15, 15-20 
Each cell = probability that the species is present into the chunk. 
You will also submit a decision file with the thresholded probability, same format, but 0 for absence, 1 for presence.

e.g. for the probability run : 
HDBirdCocktailParty-001.wav, p(S1|C1),..., p(S1|C4), p(S2|C1),..., p(S85|C4)
...
HDBirdCocktailParty-527.wav, p(S1|C1),..., p(S1|C4), p(S2|C1),..., p(S85|C4)


============

The SCHEDULE :
 
You can submit up to 5 runs by the 30th of may 2021 to glotin@univ-tln.fr

============

The TRAINSET :

The training set is at your choice (except a hand labelled version of SBCP-TEST dataset). You can use the training set of BirdCLEF 2021 main task as well as any other training set from previous LifeCLEF editions. You are also allowed to process self-supervised training, source separation or BSS on the SBCP-TEST dataset itself.

We suggest the challengers to use the ICML2103 ML4B training data set (H. Glotin et J. Sueur 2013, available in this filder (suggestedTRAINSET_fromICML2013.zip) :
30 sec of samples for each of these 35 species, total duration 18min, monophone 16bits, 44.1kHz SR, extracted from F. Deroussen naturophonia.fr (Deroussen, 2001. Oiseaux des jardins de France. Nashvert Prod. ; Deroussen et Jiguet, 2006. La sonotheque du Museum: Oiseaux de France, les passereaux. Nashvert prod.). This training data includes the phylogenetic tree (Thuillier et al., Nature, 2011) of these 35 species to maybe regulate the Chorus analysis.

Another interesting training set could be the NIPS4Bplus training data set (Morfi, V., Bas, Y., Pamuła, H., Glotin, H., & Stowell, D. (2019). NIPS4Bplus: a richly annotated birdsong audio dataset. PeerJ Computer Science, 5, e223). NIPS4Bplus is the richly annotated birdsong audio dataset, that is comprised of recordings containing bird vocalisations along with their active species tags plus the temporal annotations acquired for them. This training set (687 + 1000 files, nearly 3 hours) has been designed for task such as training models for bird population monitoring, species classification, birdsong vocalisation detection and classification. It was collected in seven regions in France and Spain.
============

CONTACT : 

glotin@univ-tln.fr