Dataset Card for UK Choral AI Dataset (uk-choral-ai)

Name: uk-choral-ai-preview-ogg
Creator: Serpentine Arts Technologies
License: https://choosealicense.com/licenses/other/

The UK Choral AI Dataset (uk-choral-ai) is an audio dataset curated by the Serpentine Arts Technologies team for The Call exhibition by Holly Herndon and Mat Dryhurst.

Dataset Details

Dataset Description

The dataset was created to train and finetune generative audio models, like those produced by IRCAM with Herndon and Dryhurst for the exhibition. The data collection methods were designed to create a rich dataset which maybe useful for a wide range of audio tasks.

The dataset contains:

483 Recordings of performances of vocal exercises and compositions devised by Holly Herndon and Matt Dryhust
Recordings are licenced for ML research and development and are GDPR compliant
The complete dataset of 48000hz 24bit WAVs is approx. 240 GB
15 UK-based choirs are represented
Each choir was recorded according to the same schema with a multi-microphone array capturing: close miked 8 soloists, 4 room mics, and a first-order ambisonic microphone.
The dataset includes the stereo mixes from the different microphone sources, a main mix and the isolated mono recordings from each microphone.
The dataset is available in 3 segments: Complete WAV, Complete Ogg, and Preview Ogg.
Curated by: Serpentine Arts Technologies with Holly Herndon and Matt Dryhurst
Trusted Data Intermediary: SG COMMERCE LIMITED
Data Steward: Jennifer Ding
License: LINK TO LICENCE
Code of Conduct: LINK TO CODE OF CONDUCT

Data Community

The UK Choral AI Dataset includes the performance recording data from 15 UK choirs, whose members form the Data Community. The contributing choirs are as follows:

Blackburn People’s Choir, Blackburn
Carnoustie Choir, Carnoustie
Cunninghame Choir, Beith
The Fitzhardinge Consort, Bristol
The Fourth Choir, London
HIVE Choir, Belfast
Leeds Vocal Movement, Leeds
London Contemporary Voices, London
Musarc, London
New London Chamber Choir, London
Open Arts Community Choir, Belfast
Ordsall Acappella Singers, Salford
The Call Finissage Public Choir Performance at Stone's Nest
The Ravenswood Singers, Newcastle
South Lakes A Cappella, Windermere
Spectrum Singers, Penarth

Uses

uk-choral-ai is intended for audio machine learning research and development but may be suitable for a wide range of music information retrieval (MIR) and natural language processing (NLP) tasks.

All usage must comply with the Code of Conduct which outlines acceptable and unacceptable use. LINK TO CODE OF CONDUCT

Annotations

Fields:

Session
Song Name
Take Number
Stem
Stem_simplified
BPM
Annotations (manual)
Silent?
Selected?

Note: Tracks marked as "Selected?" are the subset used for training the IRCAM model, this identifies the better performances/takes.

Example:

[
 {
   "Session": "10_180624",
   "Song Name": "264 We Swarm",
   "Take Number": 1,
   "Stem": "Master",
   "Stem_simplified": "Master",
   "BPM": 115,
   "Annotations (manual)": "",
   "Silent?": "",
   "Selected?": "1"
 },
 {
   "Session": "10_180624",
   "Song Name": "328 Endless",
   "Take Number": 1,
   "Stem": "Master",
   "Stem_simplified": "Master",
   "BPM": 90,
   "Annotations (manual)": "spoken word section",
   "Silent?": "",
   "Selected?": "1"
 }
]

Dataset Creation

Curation Rationale

The dataset was created to train and finetune generative audio models, like those produced by IRCAM with Herndon and Dryhurst for the exhibition.

Source Data

Data Collection and Processing

Microphone Array

The choir was recorded in the round, with 8 soloists standing in an inner circle surrounded by the rest of the choir. The recordings were taken in different venues but with efforts to use similar size rooms.

Microphones used include: 4 x MKH50 designed to capture the entire choir 8 x DPA 6066 Headset to capture the soloists 1 x Soundfield ST450 Ambisonic mic placed in the centre of the choir.

Signal Path Diagram

Signal Path Diagram Index

MKH50 – Sennheiser MKH50 hypercardiod mic
MTB40 – Wisycom MTB40s boom transmitter
HS – DPA 6066 Headset with microdot to XLRM adapter 8 x 5m XLRF to XLRM connecting headsets to multi-core inputs
ST450 – Soundfield ST450 Ambisonic mic
ST450 Ctrl – Soundfield Control, power, input and output unit for ST450 mic 2 x Soundfield specific 5 pin XLRF to 2 x XLRM output cables
12 Ch Multi-Core – 10-metre 12 channel XLR multi-core stage box
13a out – 240-volt 13-amp wall outlet with 20-metre extension reel to power ST450 Ctrl
G4 Tx – Sennheiser G4 ew100 belt pack transmitter with (mute switch active) for note taking.
ME2 – Sennheiser ME2 lavalliere mic with locking ring 3.5mm TRS termination for G4 Tx
MCR54 Rx – Wisycom MCR54 quad-channel receiver with custom dipole antennae centred for UK TV channel 38
SL2 – Sound Devices SL2 dual receiver sled, mounted on 833 containing MCR54 Rx
833 – Sound Devices 833 mixer/recorder with timecode. (MKH50’s panned 1-L, 2-R, 3-L, 4-R and monitored in Left and Right mix busses. All other ISOs faded down and monitored by PFL)
Mix Pre 10 II – Sound Device Mix Pre 10 II mixer/recorder with timecode
HP – Sennheiser HD25 headphones
TC – Tentacle Sync-E Bluetooth timecode generator providing 25fps sync to Mix Pre, TC out of Mix Pre cabled to timecode input of 833. 833 in External Timecode mode.
ST Mix Return from HP out - 3.5mm TRS to 3.5mm TRS connecting headphone output of Mix Pre to the return input of the 833. Mix Pre headphone monitoring left and right stereo busses, all ISO channels panned centre

Post-processing

The dataset comprises three supplementary subsets, each processed differently:

NoFX: This subset features raw, unmixed audio recordings.
Mix: This subset provides the stereo mixed versions of all recordings as well as the individual soloist recordings and the MKH50 sources. The audio was post-processed with EQ and compression. This subset is not noise reduced.
NR: This subset presents an automatically noise-reduced version of the dataset.

Denoising

An automated processing pipeline was created using iZotope RX Voice De-noise, De-click, De-plosive, and Deconstruct. Additionally, several lavalier recordings were manually edited using iZotope RX's spectral editor to minimise mic bleed and insert silence during non-singing sections.

Ambisonic mic limitations

Note: Due to a recording error only three channels of the first-order ambisonic microphone were recorded. This issue is not present in session 1. From sessions 2-16, there is recording data on each channel of the ambisonics file, but two channels (W and Y) are the same waveform, possibly due to a routing issue.

Ogg Vorbis Compression

Ogg Vorbis is a free and open-source audio format for lossy audio compression. The sound quality is superior to MP3 and Ogg Vorbis files are generally smaller than MP3 files. The Ogg subset of the uk-choral-ai dataset was encoded using ffmpeg, vorbis codec, quality=9.

❗️ If you require futher detail about the recording or post-processing process, please contact us

Dataset Version and Maintenance

Latest Version	Name	Size (approx)	SHA-256 Checksum	Date	Status
Complete WAV	uk-choral-ai-complete-ogg	240 GB	N/A	N/A	Unavailable
Complete Ogg	uk-choral-ai-complete-wav	25.1 GB	N/A	N/A	Unavailable
Preview Ogg	uk-choral-ai-preview-ogg	5.81 GB	N/A	N/A	Available

Roadmap

Add songbook lyrics and midi files to dataset

Changelog

Personal and Sensitive Information

This dataset includes information about individuals that can be used to identify them directly or indirectly. This means it falls under the scope of the General Data Protection Regulation (GDPR), a European Union law designed to protect personal data. When working with this dataset, you must adhere to both the specified license and code of conduct.

License: LINK TO LICENCE
Code of Conduct: LINK TO CODE OF CONDUCT

Bias, Risks, and Limitations

No demographic information was collected during the creation of this dataset. Although efforts were taken to record choirs from around the UK - including Scotland, Northern Ireland, Wales and England - the range of voices in this dataset should not be considered as representative of the diverse population of the UK or further afield.

Whilst the same recording schema was used in each sessions, the choirs were recorded in different rooms and so there will be considerable variation in the acoustics captured in each session.

Citation

When using the dataset, please use the following citation:

Serpentine Arts Technologies with Holly Herndon, Matt Dryhurst and Blackburn People’s Choir, Carnoustie Choir, Cunninghame Choir, The Fitzhardinge Consort, The Fourth Choir, HIVE Choir, Leeds Vocal Movement, London Contemporary Voices, Musarc, New London Chamber Choir, Open Arts Community Choir, Ordsall Acappella Singers, The Ravenswood Singers, South Lakes A Cappella, Spectrum Singers. UK Choral AI Dataset, 2025.

BibTeX:

@techreport{uk_choral_ai_dataset_2025, author = {{Serpentine Arts Technologies} and {Holly Herndon} and {Mat Dryhurst} and {Blackburn People's Choir} and {Carnoustie Choir} and {Cunninghame Choir} and {The Fitzhardinge Consort} and {The Fourth Choir} and {HIVE Choir} and {Leeds Vocal Movement} and {London Contemporary Voices} and {Musarc} and {New London Chamber Choir} and {Open Arts Community Choir} and {Ordsall Acappella Singers} and {The Ravenswood Singers} and {South Lakes A Cappella} and {Spectrum Singers}}, year = {2025}, title = {UK Choral AI Dataset}, institution = {Serpentine Arts Technologies},number = {Version: Complete WAV 1}}

More Information

Dataset Card Contact

[email protected]

Datasets:

serpentineartstech
/

uk-choral-ai-preview-ogg

You need to agree to share your contact information to access this dataset