Description

Methods

1) CadenzaWoodwind

This is a dataset of five woodwind instruments (flute, clarinet, oboe, alto saxophone and bassoon) created for two quartet orchestrations: (a) flute, clarinet, oboe, and bassoon and (b) flute, alto saxophone, oboe and bassoon. The stems for each solo instrument are presented along with the two mixtures.

The scores for the dataset came from the OpenScore String Quartet Corpus [6]. 21 of these were selected at random. The dice was weighted according to the length of the pieces so that longer scores with multiple movements were more likely to be chosen. This was done because there were manual steps in the subsequent rendering process, and having many small scores would make the process much longer. A maximum of two pieces per composer were chosen. Two of the scores would not render in the music notation software and so were excluded. The 19 scores selected are shown in CadenzaWoodwind_openscores_selected.txt.

The scores were interpretted using Dorico music notation software. The four string parts were allocated to flute, oboe, clarinet and bassoon and a professional sample library used to create the audio. Reverberation was added using a convolution reverb, using an impulse response from the Royal Tropical Institute, Amsterdam in Avid's Space Impulse Response IR Library.

2) EnsembleSet_Mix_1

A subset of the synthesised EnsembleSet [7] for Mix_1 stereo version provided to reduce download size. Please see the EnsembleSet page and paper for full details about methods.

3) Real Data for Tuning

Stereo_Reverb_Real_Data_For_Tuning.zip is a small sample of real recordings that are intended to help entrants deal with the mismatch between the synthetic training data and the real recorded evaluation set. It was generated from:

• Bach10 Dataset [2]
• URMP Dataset [3]

Both databases have mono recordings of isolated instruments in anechoic conditions. We have taken these and created stereo versions in small halls using convolution reverb. This used ambisonic b-format impulse responses from the Openair Database [4].

Four small to medium sized venue impulse responses were chosen: Arthur Sykes Rymer Auditorium, University of York; Central Hall, University of York; The Dixon Studio Theatre, University of York, and York Guildhall Council Chamber. Only impulse responses measured with a source-receiver distance greater than 5m were included.

The following procesure was used to convert the b-format to stereo. For each instrument.

• The b-format representation was rotated in azimuth to face the instrument being rendered. The instruments were spaced at 10 degrees apart.
• The b-format was convered to mid-side stereo using Eqn 1.7 from reference [5] (alpha=0.5).
• The mid-side stereo impulse responses were convolved with the anechoic mono music recordings.
• The mix was created by summing all the audio for the instruments in the ensemble.
• The mix was then normalised so the peak absolute sample value was 1. The scaling factor used to do this was then applied to all the solo instruments audio tracks.

The readme.txt files in each folder give the filename of the impulse response used.

code_to_create_urmp_bach10_stereo_with_reverb.zip contains the MATLAB code used to create the dataset.

Notes