def __init__(self, **kwargs):
from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
from madmom.audio.stft import ShortTimeFourierTransformProcessor
from madmom.audio.spectrogram import (FilteredSpectrogramProcessor,
LogarithmicSpectrogramProcessor,
SpectrogramDifferenceProcessor)
from madmom.processors import SequentialProcessor, ParallelProcessor
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
for frame_size in [4096]:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100)
stft = ShortTimeFourierTransformProcessor(
window=np.hamming(frame_size)) # caching FFT window
filt = FilteredSpectrogramProcessor(num_bands=12,
fmin=30,
fmax=16000,
norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=5, add=1)
#diff = SpectrogramDifferenceProcessor(diff_ratio=0.5, positive_diffs=True, stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, stft, filt, spec)))
#multi.append(SequentialProcessor((frames, stft, filt)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
super(PianoNoteProcessor, self).__init__(pre_processor)
def __init__(self, fs, hopsize_t):
from madmom.audio.signal import SignalProcessor, FramedSignalProcessor
from madmom.audio.stft import ShortTimeFourierTransformProcessor
from madmom.audio.filters import MelFilterbank
from madmom.audio.spectrogram import (FilteredSpectrogramProcessor,
LogarithmicSpectrogramProcessor)
# from madmom.features.onsets import _cnn_onset_processor_pad
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=fs)
# process the multi-resolution spec in parallel
multi = ParallelProcessor([])
for frame_size in [2048, 1024, 4096]:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100)
stft = ShortTimeFourierTransformProcessor() # caching FFT window
filt = FilteredSpectrogramProcessor(
filterbank=MelFilterbank, num_bands=80, fmin=27.5, fmax=16000,
norm_filters=True, unique_filters=False)
spec = LogarithmicSpectrogramProcessor(log=np.log, add=EPSILON)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor([frames, stft, filt, spec]))
# stack the features (in depth) and pad at beginning and end
stack = np.dstack
# pad = _cnn_onset_processor_pad
# pre-processes everything sequentially
pre_processor = SequentialProcessor([sig, multi, stack])
# instantiate a SequentialProcessor
super(MadmomMelbank3ChannelsProcessor, self).__init__([pre_processor])
def CreateProcesser(fps=100):
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
frame_sizes = [1024, 2048, 4096]
num_bands = [3, 6, 12]
for frame_size, num_bands in zip(frame_sizes, num_bands):
frames = FramedSignalProcessor(frame_size=frame_size, fps=fps)
stft = ShortTimeFourierTransformProcessor() # caching FFT window
filt = FilteredSpectrogramProcessor(num_bands=num_bands,
fmin=30,
fmax=17000,
norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=1, add=1)
diff = SpectrogramDifferenceProcessor(diff_ratio=0.5,
positive_diffs=True,
stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, stft, filt, spec, diff)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
return pre_processor
def create_feature_extraction_pipeline(sr=44100,
frame_sizes=[1024, 2048, 4096],
fps_hz=100.):
audio_loading = Pipeline([
("load_audio", FeatureExtractor(librosa.load, sr=sr, mono=True)),
("normalize", FeatureExtractor(librosa.util.normalize, norm=np.inf))
])
sig = SignalProcessor(num_channels=1, sample_rate=sr)
multi = ParallelProcessor([])
for frame_size in frame_sizes:
frames = FramedSignalProcessor(frame_size=frame_size, fps=fps_hz)
stft = ShortTimeFourierTransformProcessor() # caching FFT window
filt = FilteredSpectrogramProcessor(filterbank=LogarithmicFilterbank,
num_bands=12,
fmin=30,
fmax=17000,
norm_filters=True,
unique_filters=True)
spec = LogarithmicSpectrogramProcessor(log=np.log10, mul=5, add=1)
diff = SpectrogramDifferenceProcessor(diff_ratio=0.5,
positive_diffs=True,
stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor([frames, stft, filt, spec, diff]))
feature_extractor = FeatureExtractor(
SequentialProcessor([sig, multi, np.hstack]))
feature_extraction_pipeline = Pipeline([("audio_loading", audio_loading),
("feature_extractor",
feature_extractor)])
return feature_extraction_pipeline
def __init__(self, **kwargs):
# pylint: disable=unused-argument
from ..audio.signal import SignalProcessor, FramedSignalProcessor
from ..audio.spectrogram import (
FilteredSpectrogramProcessor, LogarithmicSpectrogramProcessor,
SpectrogramDifferenceProcessor)
from ..models import NOTES_BRNN
from ..ml.nn import NeuralNetwork
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
for frame_size in [1024, 2048, 4096]:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100)
filt = FilteredSpectrogramProcessor(
num_bands=12, fmin=30, fmax=17000, norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=5, add=1)
diff = SpectrogramDifferenceProcessor(
diff_ratio=0.5, positive_diffs=True, stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, filt, spec, diff)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
# process the pre-processed signal with a NN
nn = NeuralNetwork.load(NOTES_BRNN[0])
# instantiate a SequentialProcessor
super(RNNPianoNoteProcessor, self).__init__((pre_processor, nn))
def __init__(self, **kwargs):
# pylint: disable=unused-argument
from ..audio.signal import SignalProcessor, FramedSignalProcessor
from ..audio.stft import ShortTimeFourierTransformProcessor
from ..audio.spectrogram import (FilteredSpectrogramProcessor,
LogarithmicSpectrogramProcessor,
SpectrogramDifferenceProcessor)
from ..models import NOTES_BRNN
from ..ml.nn import NeuralNetwork
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
for frame_size in [1024, 2048, 4096]:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100)
stft = ShortTimeFourierTransformProcessor() # caching FFT window
filt = FilteredSpectrogramProcessor(num_bands=12,
fmin=30,
fmax=17000,
norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=5, add=1)
diff = SpectrogramDifferenceProcessor(diff_ratio=0.5,
positive_diffs=True,
stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, stft, filt, spec, diff)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
# process the pre-processed signal with a NN
nn = NeuralNetwork.load(NOTES_BRNN[0])
# instantiate a SequentialProcessor
super(RNNPianoNoteProcessor, self).__init__((pre_processor, nn))
def __init__(self, **kwargs):
# pylint: disable=unused-argument
from ..audio.signal import SignalProcessor, FramedSignalProcessor
from ..audio.filters import MelFilterbank
from ..audio.spectrogram import (FilteredSpectrogramProcessor,
LogarithmicSpectrogramProcessor)
from ..models import ONSETS_CNN
from ..ml.nn import NeuralNetwork
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec in parallel
multi = ParallelProcessor([])
for frame_size in [2048, 1024, 4096]:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100)
filt = FilteredSpectrogramProcessor(
filterbank=MelFilterbank, num_bands=80, fmin=27.5, fmax=16000,
norm_filters=True, unique_filters=False)
spec = LogarithmicSpectrogramProcessor(log=np.log, add=EPSILON)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, filt, spec)))
# stack the features (in depth) and pad at beginning and end
stack = np.dstack
pad = _cnn_onset_processor_pad
# pre-processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, stack, pad))
# process the pre-processed signal with a NN ensemble
nn = NeuralNetwork.load(ONSETS_CNN[0])
# instantiate a SequentialProcessor
super(CNNOnsetProcessor, self).__init__((pre_processor, nn))
def __init__(self, online=False, **kwargs):
# pylint: disable=unused-argument
from ..audio.signal import SignalProcessor, FramedSignalProcessor
from ..audio.stft import ShortTimeFourierTransformProcessor
from ..audio.spectrogram import (FilteredSpectrogramProcessor,
LogarithmicSpectrogramProcessor,
SpectrogramDifferenceProcessor)
from ..models import ONSETS_RNN, ONSETS_BRNN
from ..ml.nn import NeuralNetworkEnsemble
# choose the appropriate models and set frame sizes accordingly
if online:
origin = 'online'
nn_files = ONSETS_RNN
frame_sizes = [512, 1024, 2048]
else:
origin = 'offline'
nn_files = ONSETS_BRNN
frame_sizes = [1024, 2048, 4096]
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
for frame_size in frame_sizes:
frames = FramedSignalProcessor(frame_size=frame_size,
fps=100,
origin=origin)
stft = ShortTimeFourierTransformProcessor() # caching FFT window
filt = FilteredSpectrogramProcessor(num_bands=6,
fmin=30,
fmax=17000,
norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=5, add=1)
diff = SpectrogramDifferenceProcessor(diff_ratio=0.25,
positive_diffs=True,
stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, stft, filt, spec, diff)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
# process the pre-processed signal with a NN ensemble
nn = NeuralNetworkEnsemble.load(nn_files, **kwargs)
# instantiate a SequentialProcessor
super(RNNOnsetProcessor, self).__init__((pre_processor, nn))
def process_beats_and_chords(youtube_id):
print('START PROCESS >> ', str(datetime.now()))
from madmom.processors import ParallelProcessor
processMulti = ParallelProcessor([], num_threads=2)
processMulti.append(get_beat_processor())
processMulti.append(get_chords_processor())
print('BEAT PROCESS >> ', str(datetime.now()))
return processMulti.process(OUT_FILE_PATH + youtube_id + '.' +
OUT_FILE_EXT)
def __init__(self, online=False, **kwargs):
# pylint: disable=unused-argument
from ..audio.signal import SignalProcessor, FramedSignalProcessor
from ..audio.spectrogram import (
FilteredSpectrogramProcessor, LogarithmicSpectrogramProcessor,
SpectrogramDifferenceProcessor)
from ..models import ONSETS_RNN, ONSETS_BRNN
from ..ml.nn import NeuralNetworkEnsemble
# choose the appropriate models and set frame sizes accordingly
if online:
origin = 'online'
nn_files = ONSETS_RNN
frame_sizes = [512, 1024, 2048]
else:
origin = 'offline'
nn_files = ONSETS_BRNN
frame_sizes = [1024, 2048, 4096]
# define pre-processing chain
sig = SignalProcessor(num_channels=1, sample_rate=44100)
# process the multi-resolution spec & diff in parallel
multi = ParallelProcessor([])
for frame_size in frame_sizes:
frames = FramedSignalProcessor(frame_size=frame_size, fps=100,
origin=origin)
filt = FilteredSpectrogramProcessor(
num_bands=6, fmin=30, fmax=17000, norm_filters=True)
spec = LogarithmicSpectrogramProcessor(mul=5, add=1)
diff = SpectrogramDifferenceProcessor(
diff_ratio=0.25, positive_diffs=True, stack_diffs=np.hstack)
# process each frame size with spec and diff sequentially
multi.append(SequentialProcessor((frames, filt, spec, diff)))
# stack the features and processes everything sequentially
pre_processor = SequentialProcessor((sig, multi, np.hstack))
# process the pre-processed signal with a NN ensemble
nn = NeuralNetworkEnsemble.load(nn_files, **kwargs)
# instantiate a SequentialProcessor
super(RNNOnsetProcessor, self).__init__((pre_processor, nn))
def extract(yt_id):
beats = SequentialProcessor(
[RNNBeatProcessor(),
DBNBeatTrackingProcessor(fps=100)])
chordrec = SequentialProcessor(
[CNNChordFeatureProcessor(),
CRFChordRecognitionProcessor()])
processMulti = ParallelProcessor([])
processMulti.append(beats)
processMulti.append(chordrec)
beatSync = SequentialProcessor(
[printTime, processMulti, printTime, arrange, printTime])
return beatSync('tmp/' + yt_id + '.wav')
