README.md

Building the Single Note Database

Vienna Symphonic Library

Creating the VSL part is a manual process. For more information, please refer to VSL/README.md.

Remaining Databases

The SNDB consist of short signal excerpts, cut from longer signals in the originating databases. Each signal excerpt is defined in build/build.json:

ChosenChannel	CuttingPosition_end	CuttingPosition_start	FinalFilename	OriginalFilename	Stereo	original_db
0	304372	39777	01Pia1F013f_np___0.wav	01P001Ffnp	False	F
0	304383	39789	01Pia1F013m_np___0.wav	01P001Fmnp	False	F
0	1098379	833793	01Pia1F016m_np___0.wav	01P001Fmnp	False	F
0	9036241	8771650	01Pia1F046m_np___0.wav	01P001Fmnp	False	F
0	2290697	2198581	01Pia1R043p_tr___0.wav	011PFREP	False	R
...	...	...	...	...	...	...

After placing all input files from all databases in a directory structure like

SNDB/
├── F/  # VSL
│   ├── 01P001Ffnp.wav
│   └── ...
├── R/  # RWC
│   ├── 011PFREP.wav
│   └── ...
├── G/  # McGill
│   ├── McGill Master Samples - CD 1 - Solo Strings an....wav
│   └── ...
├── M/  # MIS
│   ├── Bass.arco.ff.sulE.E1B1.wav
│   └── ...
└── out/

You can process these input files using

import numpy as np
import pandas as pd
import soundfile as sf


df = pd.read_json('build/build.json')

for index, row in df.head().iterrows():
    data, fs = sf.read(
        "{}/{}.wav".format(row['original_db'], row['OriginalFilename']),
        always_2d=True
    )
    data = data[row['CuttingPosition_start']:row['CuttingPosition_end'], row['ChosenChannel']]

    sf.write(
        "out/{}".format(row['FinalFilename']),
        data,
        fs
    )