earthspecies / library Goto Github PK

NIPS4Bplus, is the first richly annotated birdsong audio dataset, that is comprised of recordings containing bird vocalizations along with their active species tags plus the temporal annotations acquired for them. NIPS4Bplus contains temporal annotations for the recordings that comprised the training set of the 2013 Neural Information Processing Scaled for Bioacoustics (NIPS4B) challenge for bird song classification (http://sabiod.univ-tln.fr/nips4b/challenge1.html).

Approximately 30 h of field recordings were collected. Any recording longer than 5 s was split into multiple 5 s files. SonoChiro, a chirp detection tool used for bat vocalization detection, was used on each file to identify recordings with bird vocalizations (http://www.leclub-biotope.com/fr/72-sonochiro).

Stratified random sampling was then applied to all acquired recordings, based on locations and clustering of features, to maximize the diversity in the labeled dataset, resulting in nearly 5,000 files being chosen. Following the first stage of selection, manual annotations were produced for the classes active in these 5,000 files and any recordings that contained unidentified species’ vocalizations were discarded.

Furthermore, the training set and testing set recordings were allocated so that the same species were active in both. Finally, for training purposes, only species that could be covered by at least seven recordings in the training set were included in the final dataset, the rest were considered rare species’ occurrences that would make it hard to train any classifier; hence, they were discarded. The final training and testing set consists of 687 files of the total duration of less than an hour, and 1,000 files of the total duration of nearly two hours, respectively.

The data is accessible at Figshare (https://doi.org/10.6084/m9.figshare.6798548)

Citation: Morfi V, Bas Y, Pamuła H, Glotin H, Stowell D. 2019. NIPS4Bplus: a richly annotated birdsong audio dataset. PeerJ Computer Science 5:e223 https://doi.org/10.7717/peerj-cs.223

The ReadMe should:

• Give attribution and acknowledgement to the Yossi Yovel lab for the data set (creating this data set was a time consuming and intense undertaking!)
• Describe the data set (can probably quote description in paper) ... how was it created, the purpose for which it was created, etc.

For this project in particular, where everything is collaboration/interdependence, attribution to/centering biologists and field researchers is important.

Tagging @oliver-adams-b (I can’t assign this ticket to you, so assigned it to @radekosmulski)

┆Issue is synchronized with this Asana task by Unito

@kzacarian commented on Thu Oct 29 2020

We're grateful to receive the first of several datasets from Brenda McCowan, PhD, Professor, Population Health & Reproduction, Animal Behavior Laboratory for Welfare & Conservation, School of Veterinary Medicine and Unit Leader & Core Scientist, Neuroscience & Behavior Unit, California National Primate Research Center at UC Davis.

Data is available at Box link and includes the raw audio, annotations, and paper which was published with this data.

Abstract
In this study we sought to determine whether dog barks could be divided into subtypes based on context. We recorded barking from 10 adult dogs, Canis familiaris, of six breeds in three different test situations: (1) a disturbance situation in which a stranger rang the doorbell, (2) an isolation situation in which the dog was locked outside or in a room isolated from its owner and (3) a play situation in which either two dogs or a human and a dog played together. We analysed spectrograms of 4672 barks using macros that took 60 sequential frequency measurements and 60 sequential amplitude measurements along the length of the call. Statistical analyses revealed that barks are graded vocalizations that range from harsh, low-frequency, unmodulated calls to harmonically rich, higher-frequency, modulated calls. The harsh, low-frequency, unmodulated barks were more commonly given in the disturbance situation, and the more tonal, higher-pitch, modulated barks were more commonly given in the isolation and play situations. Disturbance barks were also longer in duration with more rapid repetition than the barks given in other contexts. Discriminant analysis revealed that dog barks can be divided into different subtypes based on context even within individual dogs, and that dogs can be identified by their bark spectrograms despite the context of the bark.

Barking in domestic dogs: Context specificity and individual identification. Yin S., McCowan B. (2004) Animal Behaviour, 68 (2) , pp. 343-355.

@kzacarian commented on Wed Nov 11 2020

Issue moved to earthspecies/library #29 via ZenHub

This dataset contains audio recordings of 71 different bird species gathered from three different sources (The Macaulay Library, the Art & Science Centre (UCLA), and the Great Himalayan national Park dataset). The full dataset is only ~450MB, and is split up by folder into testing, training, and validation sets. The original recording lengths vary from 0.5 seconds to 320 seconds, and are all sampled at 44.1kHz. However, in the provided dataset only the Mel-spectrograms of the bird calls are provided, and are stored in .npy format.

The Birdcalls71 dataset was found in the paper (?) by Anshul Thakur et al: "Multiscale CNN based Deep Metric Learning for Bioacoustic Classification: Overcoming Training Data Scarcity Using Dynamic Triplet Loss" (still under review) (I suspect the researchers are the people who curated this dataset, though I'm not completely sure).

┆Issue is synchronized with this Asana task by Unito

Birdvox-Full-Night

Description of the Data:

35402 flight calls in total, among ~25 different species of passerines. The annotation process took 102 hours.

Data Download location:

https://wp.nyu.edu/birdvox/birdvox-full-night/

Article Abstract:

This article addresses the automatic detection of vocal, nocturnally migrating birds from a network of acoustic sensors. Thus far, owing to the lack of annotated continuous recordings, existing methods had been benchmarked in a binary classification setting (presence vs. absence). Instead, with the aim of comparing them in event detection, we release BirdVox-full-night, a dataset of 62 hours of audio comprising 35402 flight calls of nocturnally migrating birds, as recorded from 6 sensors. We find a large performance gap between energy-based detection functions and data-driven machine listening. The best model is a deep convolutional neural network trained with data augmentation. We correlate recall with the density of flight calls over time and frequency and identify the main causes of false alarm.

Citation

V. Lostanlen, J. Salamon, A. Farnsworth, S. Kelling and J. P. Bello, "Birdvox-Full-Night: A Dataset and Benchmark for Avian Flight Call Detection," 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Calgary, AB, 2018, pp. 266-270, doi: 10.1109/ICASSP.2018.8461410.

I would like to let you know that I have kicked off the work on adding a document on a adding dataset to the library!

It is a rough draft for now. You can find it here

@bs @aza @kzacarian @pcbermant @oliver-adams-b would really love your thoughts. I think we can work on the markdown directly?

Is there anything high level that I am missing? Also - please feel free to put more substance around these steps.

All thoughts / edits / contributions to the doc are much welcome! 🙂

┆Issue is synchronized with this Asana task by Unito

As well as an overview of what kind of data is in there (hopefully with some figures and sounds).

We recently had a great call with Carrie Wall Bell, who is leading the NOAA passive acoustic archive, available on GCP to support immediate access and facilitate 'in place' processing. According to their website, Data are free and without copyright restriction.

_Data Collection and Applications
Passive acoustic monitoring of the ocean sound field is a critical aspect of NOAA’s mandate for ocean and coastal data stewardship. Sound can travel vast distances underwater (e.g., across ocean basins) making passive acoustic monitoring a powerful observational tool that is used across NOAA to detect and characterize:

• sounds produced and used by living marine resources (e.g., endangered marine mammals, commercially important fish species)
• natural sources of noise from physical oceanographic processes
• anthropogenic noise sources that contribute to the overall ocean noise environment
• Passive acoustic data are used broadly across NOAA’s National Marine Fisheries Service (NMFS), National Ocean Service (NOS), and Oceanic and Atmospheric Research (OAR) line offices for a wide range of activities central to NOAA’s mission including marine mammal stock assessments, monitoring of earthquake and geological activity, and assessing impacts of anthropogenic noise on marine life.

Data Stewardship
NCEI and NOAA line offices are collaborating to create an archive for passive acoustic data from a number of sources throughout government and academia._

Some kind of table with meta data like:

• years of capture
• size/hours/number of calls
• lead author

┆Issue is synchronized with this Asana task by Unito

This dataset contains Mel-spectrogram representations of audio recordings labeled by species type across 10 species. Anura is an order of amphibians found mostly in South America, and an Anuran is just a frog/toad of that order. So this dataset is just a bunch of recordings of cute/colorful frogs! The data are stored in .npy format, 441 kHz and have durations varying from 3 to 360 seconds. There are nearly 2.4k samples across the testing, validation, and training sets. Some analysis and explorations into the Anuran dataset can be found in the paper here.

┆Issue is synchronized with this Asana task by Unito

@markjohngraham, any luck on ingesting your test hard drive to archive.org?

Hey,

I just spot-checked your classification data set and it seems like a valuable addition to our annotation efforts. I've added a link to your efforts within Orcasound's orcadata wiki where we are trying to track the open model development process for orca signals.

I did find one true negative that contains a faint SRKW signal. 587.wav contains a faint SRKW call. This may have been due to an error in my initial labeling effort, or a bug in your code for generating the no/call clips and annotations.

For modelers interested in orca echolocation, please note that many of the true negative samples contain prominent (high SNR) echolocation clicks (e.g. 584.wav)

Also, in the orca README, the intro ends with what appears to be an errant phrase:
s are the most widespread primate genus.

Finally, you could add in the metadata/description that these orca signals are from the Southern Resident Killer Whale ecotype.

Orcasound is a cooperative effort of many dedicated individuals and organizations. Their “AI for Orcas” project represents a 2-decade effort by the Orcasound and open-source communities to develop machine learning models to classify signals made by killer whales (calls, clicks, and whistles), in real-time, starting with the calls of the endangered orcas known as the Southern Resident Killer Whales (SRKWs). An important part of the process is generating high-quality and open-access training and testing data sets. On the Orcasound Github, they share all of their data publicly, via AWS. 

The repository also has links to Orca Machine Learning resources.

Right now they live in:

• Earth Species Project Collection
  • Macaques, Otters, and a reference to author of macaques
• with creator metadata 'Earth Species'
  • II.IO workshop talk
  • This seems to also fuzzy match on 'Earth Species Project'
• [Yossi Yovel's Egyptian fruit bat dataset, uploaded by @oliver-adams-b]

A dataset includes:

• binary compressed file of all data, to be downloaded for training in a jupyter notebook
• Extracted contents to be browsed and played online
• Associated files: papers, videos, photos

@markjohngraham, how does map to Archive.org taxonomy?

Before I got the sweet lead on the Egyptian fruit bat datasets, I was planning on getting my noises from the Watkins Marine Mammal Sound Database. Recordings from many different marine mammal vocalizations can be downloaded directly from the site, along with the metadata unique to each recording. I created this webscraper a few weeks ago so that I could download and organize the audio/metadata for the 'Best Of' selection of recordings from the website. You can find the archive of the datasets I downloaded from the site here. I originally just downloaded recordings from whales, but there are a lot more species to be studied!

The recordings are sampled at varying rates and quality, with some dating back to around 1960! There is a varying number of samples for each of the 12 species of whales whose vocalizations I grabbed. All vocalizations, however, are stored in .wav format, with the metadata stored in .csv. The vocalizations can be labeled by species through their parent folder, but to predict other fields you'll need to go through the metadata. The information required to decode the metadata can be found here. On the website it is stated that the data is free to download and use for personal or academic (non commercial) use.

┆Issue is synchronized with this Asana task by Unito

@kzacarian commented on Thu Oct 29 2020

We're grateful to receive the first of several datasets from Brenda McCowan, PhD, Professor, Population Health & Reproduction, Animal Behavior Laboratory for Welfare & Conservation, School of Veterinary Medicine and Unit Leader & Core Scientist, Neuroscience & Behavior Unit, California National Primate Research Center at UC Davis.

Data is available at Box link and includes the raw audio, annotations, and paper which was published with this data.

Abstract
In this study we sought to determine whether dog barks could be divided into subtypes based on context. We recorded barking from 10 adult dogs, Canis familiaris, of six breeds in three different test situations: (1) a disturbance situation in which a stranger rang the doorbell, (2) an isolation situation in which the dog was locked outside or in a room isolated from its owner and (3) a play situation in which either two dogs or a human and a dog played together. We analysed spectrograms of 4672 barks using macros that took 60 sequential frequency measurements and 60 sequential amplitude measurements along the length of the call. Statistical analyses revealed that barks are graded vocalizations that range from harsh, low-frequency, unmodulated calls to harmonically rich, higher-frequency, modulated calls. The harsh, low-frequency, unmodulated barks were more commonly given in the disturbance situation, and the more tonal, higher-pitch, modulated barks were more commonly given in the isolation and play situations. Disturbance barks were also longer in duration with more rapid repetition than the barks given in other contexts. Discriminant analysis revealed that dog barks can be divided into different subtypes based on context even within individual dogs, and that dogs can be identified by their bark spectrograms despite the context of the bark.

Barking in domestic dogs: Context specificity and individual identification. Yin S., McCowan B. (2004) Animal Behaviour, 68 (2) , pp. 343-355.

I tried both a more traditional remote merge, as well as using subtree.

Both preserve history completely (as seen by using a recent version of git, or the --follow flag to git-log), but Github does not yet reflect the history in their default log browser. Contributor list is correctly merged.

See this issue for more details.

@PomoML @AdPostma @dienhoa @radekosmulski