README.md

metadata

license: cc-by-4.0
size_categories:
  - 100K<n<1M
task_categories:
  - audio-classification
pretty_name: VocalSimilarity
dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: label
      dtype: string
    - name: speaker
      dtype: string
    - name: subset
      dtype: string
  splits:
    - name: train
      num_bytes: 21758664803.439
      num_examples: 269523
  download_size: 0
  dataset_size: 21758664803.439
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Description

This multi-species dataset was customized to benchmark k-NN retrieval and cluster separation tecniques on Human and Songbird vocalizations.

Download Dataset

from huggingface_hub import snapshot_download
snapshot_download('anonymous-submission000/vocsim', local_dir = "data/vocsim", repo_type="dataset" )

For more usage details, please refer to the GitHub repository: https://anonymous.4open.science/anonymize/neural_embeddings-6EE5

Data Fields

1. Subset: Specifies the subset/category of the dataset. It can indicate whether the sample is from humans or songbirds, and possibly more detailed categorization.
2. Audio: Contains the audio sample.
3. Label: Represents the label or class of the audio clip, indicating the type of vocalization or sound.
4. Speaker: Identifies the speaker or source of the vocalization in the case of human datasets, or the individual bird in the case of songbird datasets.

Human Datasets

1. AMI: The AMI Meeting Corpus comprises 100 hours of multi-modal meeting recordings, including audio data for utterances, words, and vocal sounds, alongside detailed speaker metadata.
2. TIMIT: The TIMIT dataset contains manual phonetic transcriptions of utterances read by 630 English speakers with various dialects.
3. VocImSet: The Vocal Imitation Set contains recordings of 236 unique sound sources being imitated by 248 speakers.
4. VocalSketch: The Vocal Sketch Dataset contains two sets of 10'705 and 5'700 imitations respectively of 240 sounds.

Songbird Datasets

1. DAS: The Deep Audio Segmenter Dataset features single male Bengalese finch songs, including 473 vocalizations of 6 vocalization types.
2. Tomka: The Gold-Standard Zebrafinch dataset contains 48,059 vocalizations of 36 vocalization types from 4 zebra finches.
3. Nicholson: The Bengalese finch song repository includes songs of four Bengalese finches recorded in the Sober lab at Emory University and manually clustered by two authors.
4. Elie: Vocal repertoires from zebra finches, collected between 2011 and 2014 at the University of California Berkeley by Julie E Elie. This dataset contains 3,500 vocalizations from 50 individuals and 65 vocalization types.

Contact