cometbeat-assets / assets /instruments /vcsl /Scripts /createInstruments.py
cstr's picture
instruments (part 13)
db9932d verified
Raw
History Blame Contribute Delete
13.8 kB
"""Create SFZ instruments from samples.
Written by Peter Eastman. This file is in the public domain. For usage instructions, run this script as
python createInstruments.py --help
This script use the PySoundFile library. See https://pypi.org/project/PySoundFile/ for installation instructions.
You can still use it without PySoundFile, but only wav files will be supported.
"""
from __future__ import print_function
from __future__ import division
import argparse
import collections
import os
import sys
import numpy as np
import scipy.signal
try:
# Use PySoundFile if possible.
import soundfile as sf
formats = ['.wav', '.ogg', '.flac']
def loadFile(path):
data, rate = sf.read(path)
if len(data.shape) == 1:
data = np.expand_dims(data, 1)
return data, rate
except:
# It isn't installed, so use wavio instead.
import wavio
formats = ['.wav']
def loadFile(path):
wav = wavio.read(path)
return wav.data, wav.rate
noteNames = ['C', 'C#', 'D', 'D#', 'E', 'F', 'F#', 'G', 'G#', 'A', 'A#', 'B']
dynamicNames = ['pppp', 'ppp', 'pp', 'p', 'mp', 'mf', 'f', 'ff', 'fff', 'ffff',
'v11', 'vl2', 'vl3', 'vl4', 'vl5',
'v1', 'v2', 'v3', 'v4', 'v5', 'v6', 'v7', 'v8', 'v9',
'soft', 'quiet', 'med', 'medium', 'loud']
# Parse command line options.
parser = argparse.ArgumentParser(description='Create SFZ instruments from samples.')
parser.add_argument('--attack', type=float, default=0.004, help='Attack time in seconds (default=0.004)')
parser.add_argument('--release', type=float, default=0.3, help='Release time in seconds (default=0.3)')
parser.add_argument('--velocityexponent', type=float, default=0.6, help='Exponent shaping the curve assigning layers to velocities (default=0.6)')
parser.add_argument('--transpose', type=int, default=0, help='Transposition that has been applied to the samples, in semitones (default=0)')
parser.add_argument('--volume', type=float, default=0.0, help='Amplification to apply to all samples, in dB (default=0.0)')
parser.add_argument('--releasevolume', type=float, default=0.0, help='Amplification to apply to release samples, in dB (default=0.0)')
parser.add_argument('--crossfade', action='store_true', help='Perform crossfading between velocity layers')
parser.add_argument('--unpitched', action='store_true', help='Treat this as an unpitched instrument')
parser.add_argument('--noreleases', action='store_true', help='Ignore release samples')
parser.add_argument('--notuning', action='store_true', help='Do not apply tuning to correct pitches')
parser.add_argument('--articulation', type=str, default=None, help='Name of the articulation to create (default is to create all articulations)')
parser.add_argument('instrumentdir', help="Top level directory containing the instrument's samples")
args = parser.parse_args()
outputDir = os.path.abspath(os.path.join(args.instrumentdir, os.pardir))
# Create objects to represent all the notes in the scale.
Note = collections.namedtuple('Note', ['index', 'name', 'frequency'])
noteWithName = {}
for octave in range(-1, 10):
for i, name in enumerate(noteNames):
index = i+(octave+2)*12-args.transpose
name = name+str(octave)
frequency = 440.0 * 2**((index-69)/12.0)
noteWithName[name] = Note(index, name, frequency)
# Define a class to represent a single sample file.
class Sample(object):
def __init__(self, filename):
self.filename = filename
self.note = None
self.layer = None
self.rr = None
self.offset = 0
self.tuning = 0
self.loudness = 0
noteFields = []
fields = filename[:filename.index('.')].split('_')
for field in fields:
if field in noteWithName and not args.unpitched:
self.note = noteWithName[field]
elif field in dynamicNames:
self.layer = dynamicNames.index(field)
elif field.startswith('rr'):
self.rr = int(field[2:])
elif field.startswith('vl'):
self.layer = int(field[2:])
elif field[-1].isdigit() and field[:-1] in dynamicNames:
self.layer = dynamicNames.index(field[:-1])
self.rr = int(field[-1])
else:
noteFields.append(field)
if args.unpitched:
# For unpitched instrument, two samples are considered the same "note" if all parts of the filename
# match except layer and round robin.
self.note = tuple(noteFields)
# Define a class to represent an articulation.
class Articulation(object):
def __init__(self, instrument, directory, isRelease):
self.instrument = instrument
self.directory = directory
self.isRelease = isRelease
self.samples = []
self.name = os.path.split(directory)[-1]
if isRelease and self.name == 'Releases':
# This happens when there's a single articulation, with the samples directly in the "Sustains" and "Releases" folders.
self.name = 'Sustains'
def addSample(self, sample):
self.samples.append(sample);
# Define a class to represent an instrument.
class Instrument(object):
def __init__(self, directory):
self.directory = directory
self.articulations = []
head, tail = os.path.split(directory)
if len(tail) > 0:
self.name = tail
else:
self.name = os.path.split(head)[-1]
# This function computes the low end of the velocity range for a layer.
def layer_lowvel(layerIndex, layers):
numLayers = len(layers)
if args.crossfade:
numLayers -= 1
return int((layerIndex/numLayers)**args.velocityexponent*128)
# The following functions are adapted from https://gist.github.com/endolith/255291. They
# are used to estimate the pitch of a sample.
def parabolic(f, x):
xv = 1/2. * (f[x-1] - f[x+1]) / (f[x-1] - 2 * f[x] + f[x+1]) + x
yv = f[x] - 1/4. * (f[x-1] - f[x+1]) * (xv - x)
return (xv, yv)
def freq_from_autocorr(sig, fs):
corr = scipy.signal.fftconvolve(sig, sig[::-1], mode='full')
corr = corr[len(corr)//2:]
d = np.diff(corr)
start = np.where(d > 0)[0][0]
peak = np.argmax(corr[start:]) + start
px, py = parabolic(corr, peak)
return fs / px
# Scan the directory to create a list of articulations and samples for the instrument.
def scanDirectory(directory, instrument, isRelease):
if os.path.split(directory)[-1].lower() == 'releases':
if args.noreleases:
return
isRelease = True
articulation = Articulation(instrument, directory, isRelease)
for filename in os.listdir(directory):
filepath = os.path.join(directory, filename)
if os.path.isdir(filepath):
scanDirectory(filepath, instrument, isRelease)
elif any(filename.endswith(e) for e in formats):
articulation.addSample(Sample(filename))
if len(articulation.samples) > 0:
if args.articulation is None or articulation.name == args.articulation:
instrument.articulations.append(articulation)
instrument = Instrument(args.instrumentdir)
scanDirectory(args.instrumentdir, instrument, False)
# Loop over all samples and analyze them.
for articulation in instrument.articulations:
for sample in articulation.samples:
data, rate = loadFile(os.path.join(articulation.directory, sample.filename))
# Identify silence at the start of the sample.
amplitude = np.abs(np.max(data, axis=1))
smoothedAmplitude = np.convolve(amplitude, np.ones((100,))/100, mode='valid')
cutoff = np.max(smoothedAmplitude)/50
offset = np.min(np.where(smoothedAmplitude > cutoff))
if offset > 100:
sample.offset = offset-100
# Identify the tuning correction.
if not args.unpitched and not args.notuning:
frequency = freq_from_autocorr(data[:,0], rate)
for mult in [3, 2, 1, 0.5, 1/3.0]:
f = frequency*mult
ratio = f/sample.note.frequency
if 0.95 < ratio < 1.05:
sample.tuning = int(-100*np.log(ratio)/np.log(2**(1/12.0)))
# Estimate the loudness as the 90th percentile of amplitude during the first second.
sample.loudness = np.percentile(smoothedAmplitude[offset:offset+rate], 90)
# Write an articulation to a SFZ file.
def writeArticulation(articulation, outfile, isSustain, isRelease):
print('<group>', file=outfile)
attack = args.attack
release = args.release
if isSustain:
print('trigger=attack', file=outfile)
release = 0.1
elif isRelease:
print('trigger=release', file=outfile)
attack = 0.1
print('ampeg_attack=%f' % attack, file=outfile)
if not isRelease:
print('ampeg_release=%f' % release, file=outfile)
print(file=outfile)
# Find the range of keys to use each note for.
notes = sorted(set([s.note for s in articulation.samples]))
if args.unpitched:
lowkey = np.arange(60, 60+len(notes))
highkey = lowkey
else:
lowkey = np.zeros(len(notes))
highkey = np.zeros(len(notes))
lowkey[0] = notes[0].index
for i in range(len(notes)-1):
highkey[i] = (notes[i].index+notes[i+1].index)//2
lowkey[i+1] = highkey[i]+1
highkey[-1] = notes[-1].index+1
# Compute a target loudness for every note.
averageLoudness = np.zeros(len(notes))
for noteIndex, note in enumerate(notes):
samples = [s for s in articulation.samples if s.note == note]
averageLoudness[noteIndex] = np.mean([s.loudness for s in samples])
if args.unpitched:
targetLoudness = averageLoudness
else:
targetLoudness = np.zeros(len(notes))
for i in range(len(notes)):
lower = np.max([0, i-2])
targetLoudness[i] = np.mean(averageLoudness[lower:i+3])
# Loop over notes.
for noteIndex, note in enumerate(notes):
samples = [s for s in articulation.samples if s.note == note]
# Loop over velocity layers for each note.
layers = sorted(set([s.layer for s in samples]))
for layerIndex, layer in enumerate(layers):
# Loop over round robins for each layer.
rrs = sorted([s for s in samples if s.layer == layer], key=lambda x: x.rr)
for rrIndex, sample in enumerate(rrs):
print('<region>', file=outfile)
if len(rrs) > 1:
print('seq_length=%d' % len(rrs), file=outfile)
print('seq_position=%d' % (rrIndex+1), file=outfile)
if len(layers) > 1:
if args.crossfade:
if layerIndex > 0:
print('xfin_lovel=%d' % layer_lowvel(layerIndex-1, layers), file=outfile)
print('xfin_hivel=%d' % (layer_lowvel(layerIndex, layers)-1), file=outfile)
if layerIndex < len(layers)-1:
print('xfout_lovel=%d' % layer_lowvel(layerIndex, layers), file=outfile)
print('xfout_hivel=%d' % (layer_lowvel(layerIndex+1, layers)-1), file=outfile)
else:
print('lovel=%d' % layer_lowvel(layerIndex, layers), file=outfile)
print('hivel=%d' % (layer_lowvel(layerIndex+1, layers)-1), file=outfile)
print('sample=%s' % os.path.relpath(os.path.join(articulation.directory, sample.filename), outputDir), file=outfile)
if args.unpitched:
keycenter = lowkey[noteIndex]
else:
keycenter = note.index
print('pitch_keycenter=%s' % keycenter, file=outfile)
print('lokey=%d' % lowkey[noteIndex], file=outfile)
print('hikey=%d' % highkey[noteIndex], file=outfile)
if sample.offset > 0:
print('offset=%d' % sample.offset, file=outfile)
if sample.tuning != 0 and not args.notuning:
print('tune=%d' % sample.tuning, file=outfile)
if articulation.isRelease:
volume = args.releasevolume
else:
volume = args.volume
amplification = targetLoudness[noteIndex]/sample.loudness
db = 20*np.log10(amplification) + volume
print('volume=%f' % db, file=outfile)
print(file=outfile)
# Create SFZ files for all the articulations.
articulations = [a for a in instrument.articulations if not a.isRelease]
for articulation in articulations:
if not articulation.isRelease:
if len(articulations) > 1:
filename = '%s - %s.sfz' % (instrument.name, articulation.name)
elif args.articulation is not None:
filename = '%s - %s.sfz' % (instrument.name, args.articulation)
else:
filename = '%s.sfz' % instrument.name
# See if we have release samples for this articulation.
release = None
for a in instrument.articulations:
if a.isRelease and a.name == articulation.name:
release = a
with open(os.path.join(outputDir, filename), 'w') as outfile:
print('// Generation Options:', ' '.join(sys.argv[1:-1]), file=outfile)
print(file=outfile)
writeArticulation(articulation, outfile, release is not None, False)
if release is not None:
writeArticulation(release, outfile, False, True)