Bioacoustics Datasets

A collection of tools and 1,000,000+ unified annotations for bioacoustics datasets.

Dataset	Species	# of annotated calls	Dataset size (GB)	Duration (hh:mm:ss)	License
Animal Sounds	Birds, cats, chickens, cows, dogs, donkeys, frogs, lions, monkeys, sheep	809	0.13	0:57:47	-
AnuraSet	Anurans	16089	18.6	27:00:00	cc-by-4.0
BIRDeep	38 avian species	3749	13.41	8:50:00	MIT
BirdVox	25 avian species	35402	0.79	7:23:00	cc-by-4.0
Domestic Canary	Canary	23308	0.856	3:00:00	cc-by-4.0
Columbia/Costa Rica Coffee Farms	89 avian species	6952	3.8	34:00:00	cc-by-4.0
Darpa	Humans	1718	0.67	4:00:00	No license specified, the work may be protected by copyright
Avain Dawn	58 avian species, 1 amphibian species	41183	20.3	131:15:00	cc-by-4.0
DCASE	Birds	7206	32.00	17:25:00	cc-by-4.0
Egyptian fruit bat	Egyptian fruit bat	90000	91.00	37:45:00	cc-by-4.0
ENABirds	Birds	16052	1.40	6:20:00	cc-by-1.0
Female Rook	Rook birds	3417	54.37	10:45:36	cc-by-nc-nd-4.0
The Vocal Repertoire of Adult and Neonate Otters	Otter	441	0.57	0:06:23	cc
Hainan Gibbons	Hainan Gibbons	1233	13.39	104:00:00	cc-by-4.0
Hawaii Birds	27 avian species	59583	5.8	51:00:00	cc-by-4.0
HICEAS	Whales, Dolphins	796	3.10	12:40:00	"Public dataset hosted in Google Cloud Storage"
Distributed acoustic cues for caller identity in macaque vocalization	Macaques monkeys	7285	0.15	0:45:00	cc-by-1.0
InfantMarmosetVox	Marmosets monkeys	169318	21.2	58:20:00	cc-by-4.0
Northeast US Sounds	81 avian species	50760	27.8	285:00:00	cc-by-4.0
Orcas Classifications	Orca whales	398	0.26	0:26:30	-
Pigs	Pig	6887	0.2	0:40:26	cc-by-4.0
Rainforest	Birds, frogs	1216	13.05	20:16:00	"Free for personal or academic purposes"
Rodents	Rodents: mouse, gerbil	4576	1.36	0:48:34	cc-by-4.0
Rook	Rook birds	17662	23.49	17:21:17	cc-by-4.0
Sierra Nevada	21 avian species	10976	3.57	16:40:00	cc-by-4.0
Southwest Amazon	132 avian species	16482	4.51	21:00:00	cc-by-4.0
Watkins Marine Animal Sounds	21 dolphin, 13 seal, 32 whale species	15152	9.61	29:10:15	"Sound files are free to download for personal or academic use"
Western US	56 avian species	20147	7.08	33:00:00	cc-by-4.0

To install all of the data, you need about TOTAL_GB of free space. But you can also pick and choose which datasets you'd like to download.

Installation Instructions

1. Run ./scripts/download_data.sh

After running the download script, your datasets folder should look like this:

└── datasets/
    ├── annotations.pkl
    ├── dataset1/
    │   ├── audio/
    │   │   ├── audio1.wav
    │   │   └── audio2.wav
    │   ├── annotations.pkl
    │   └── stats.txt
    └── dataset2/
        ├── audio/
        ├── annotations.pkl
        └── stats.txt

There are individual annotation files for each dataset and one master annotations file located directly in the datasets folder. annotations.pkl is a Python dictionary structured as the following

{
   wav_file_path: [
                   {'start_time': 0, 'end_time': 1.7, 'species': 'bird', 'sub-species': 'serinus canaria'},
                   {'start_time': 2.3, 'end_time': 2.48, ...},
                   ...
                  ]
   ...
}

Tools

1. ./scripts/generate_spectrograms.sh - Running this will generate 100 high quality mel spectrograms for each dataset and place them in the visualizations folder

2. scripts/moving_spectrogram.py example_audio.wav output.wav - This script takes in a wav file and generates a moving spectogram with audio called example_audio.wav and saves it in visualizations/output.wav

3. training/data_engine.py - This a very helpful file to take in datasets and easily produce a PyTorch Dataset. The __ getitem __() method has the output [audio, is_vocalization, species, speaker]. Audio is a tensor of numbers, is_vocalization is a boolean, species is the species of the vocalization, and speaker is the speaker of the vocalization. species and speaker will both be "Noise" if it is a non-vocalization event and speaker will be 'no-speaker' if there is no speaker data. Dataset has three required parameters: datasets_path which should just be datsets folder. Save_path which is where train and val splits will be stored. And datasets which is a list of all the datasets you would like to utilize. After the data is loaded into the data_engine, you can call data_engine.get_annotated_dataset(dataset_names=[]) which returns the above stated PyTorch dataset.