arch-raven's picture
add readme
dae4dcc

Neural Audio Fingerprint Dataset

(c) 2021 by Sungkyun Chang https://github.com/mimbres/neural-audio-fp

This dataset includes all music sources, background noise and impulse-reponses (IR) samples that have been used in the work ["Neural Audio Fingerprint for High-specific Audio Retrieval based on Contrastive Learning"] (https://arxiv.org/abs/2010.11910).

Format:

16-bit PCM Mono WAV, Sampling rate 8000 Hz

Description:

/
fingerprint_dataset_icassp2021/
β”œβ”€β”€ aug
β”‚   β”œβ”€β”€ bg         <=== Pub/cafe etc. background noise mix
β”‚   β”œβ”€β”€ ir         <=== IR data for microphone and room reverb simulatio
β”‚   └── speech     <=== English conversation, NOT USED IN THE PAPER RESULT
β”œβ”€β”€ extras
β”‚   └── fma_info   <=== Meta data for music sources.
└── music
    β”œβ”€β”€ test-dummy-db-100k-full  <== 100K songs of full-lengths
    β”œβ”€β”€ test-query-db-500-30s    <== 500 songs (30s) and 2K synthesized queries
    β”œβ”€β”€ train-10k-30s            <== 10K songs (30s) for training
    └── val-query-db-500-30s     <== 500 songs (30s) for validation/mini-search

Data source:

β€’ Bacgkound noise from Audioset was retrieved using key words ['subway', 'metro', 'underground', 'not music'] β€’ Cochlear.ai pub-noise was recorded at the Strabucks branches in Seoul by Jeongsoo Park. β€’ Random noise was generated by Donmoon Lee. β€’ Room/space IR data was collected from Aachen IR and OpenAIR 1.4 dataset. β€’ Portions of MIC IRs were from Vintage MIC (http://recordinghacks.com/), and pre-processed with room/space IR data. β€’ Portions of MIC IRs were recorded by Donmoon Lee, Jeonsu Park and Hyungui Lim using mobile devices in the anechoic chamber at Seoul National University. β€’ All music sources were taken from the Free Music Archive (FMA) data set, and converted from stereo 44Khz to mono 8Khz. β€’ train-10k-30s contains all 8K songs from FMA_small. The remaining 2K songs were from FMA_medium. β€’ val- and test- data were isolated from train-, and taken from FMA_medium. β€’ test-query-db-500-30s/query consists of the pre-synthesized 2,000 queries of 10s each (SNR 0~10dB) and their corresponding 500 songs of 30s each. β€’ Additionally, query_fixed_SNR directory contains synthesized queries with fixed SNR of 0dB and -3dB. β€’ dummy-db-100k was taken from FMA_full, and duplicates with other sets were removed.

License:

This dataset is distributed under the CC BY-SA 2.0 license separately from the github source code, and licenses for composites from other datasets are attached to each sub-directory.