def analyse(self, workable):
nmf = fluid.nmf(
workable,
iterations=self.iterations,
components=self.components,
fftsettings=self.fftsettings,
)
bases = get_buffer(nmf.bases, "numpy")
bases_smoothed = np.zeros_like(bases)
for i, x in enumerate(bases):
bases_smoothed[i] = savgol_filter(x, self.smoothing,
self.polynomial)
clusterer = hdbscan.HDBSCAN(
min_cluster_size=self.min_cluster_size,
min_samples=self.min_samples,
cluster_selection_method=self.cluster_selection_method,
)
cluster_labels = clusterer.fit_predict(bases_smoothed)
unique_clusters = list(dict.fromkeys(cluster_labels))
sound = get_buffer(nmf.resynth, "numpy")
for x in unique_clusters:
summed = np.zeros_like(
sound[0]) # make an empty numpy array of same size
base = Path(workable).name
output = self.output / f"{base}_{x}.wav"
for idx, cluster in enumerate(cluster_labels):
if cluster == x:
summed += sound[idx]
wavfile.write(output, 44100, summed)
def analyse(self, workable):
noveltyslice = fluid.noveltyslice(
workable,
feature=self.feature,
fftsettings=self.fftsettings,
filtersize=self.filtersize,
minslicelength=self.minslicelength,
threshold=self.threshold,
)
self.buffer[workable] = [int(x) for x in get_buffer(noveltyslice)]
def analyse(self, workable):
noveltyslice = fluid.noveltyslice(
workable,
feature=self.feature,
fftsettings=self.fftsettings,
filtersize=self.filtersize,
kernelsize=self.kernelsize,
minslicelength=self.minslicelength,
threshold=self.threshold,
)
self.buffer[str(workable)] = get_buffer(noveltyslice)
def analyse(self, workable):
hsh = create_hash(workable, self.identity)
cache = self.process.cache / f"{hsh}.wav"
if not cache.exists():
slice_output = get_buffer(
fluid.onsetslice(
workable,
indices=cache,
fftsettings=self.fftsettings,
filtersize=self.filtersize,
framedelta=self.framedelta,
metric=self.metric,
minslicelength=self.minslicelength,
threshold=self.threshold,
),
"numpy",
)
else:
slice_output = get_buffer(cache)
self.buffer[str(workable)] = slice_output.tolist()
def analyse(self, workable):
hsh = create_hash(workable, self.identity)
cache = self.process.cache / f"{hsh}.npy"
if cache.exists():
mfcc = np.load(cache, allow_pickle=True)
else:
mfcc = get_buffer(
fluid.mfcc(
str(workable),
fftsettings=self.fftsettings,
numbands=self.numbands,
numcoeffs=self.numcoeffs,
minfreq=self.minfreq,
maxfreq=self.maxfreq,
),
"numpy",
)
np.save(cache, mfcc)
self.buffer[str(workable)] = mfcc.tolist()
def analyse(self, workable):
slices = self.input[workable]
if len(slices) == 1:
self.buffer[workable] = slices
return
count = 0
standardise = StandardScaler()
model = AgglomerativeClustering(n_clusters=self.numclusters)
while (count + self.windowsize) <= len(slices):
indices = slices[
count:count +
self.windowsize] # create a section of the indices in question
data = []
for _, (start, end) in enumerate(zip(indices, indices[1:])):
mfccs = fluid.mfcc(
workable,
fftsettings=self.fftsettings,
startframe=int(start),
numframes=int(end - start),
)
stats = get_buffer(
fluid.stats(mfccs, numderivs=self.numderivs), "numpy")
data.append(stats.flatten())
data = standardise.fit_transform(data)
cluster = model.fit(data)
cur = -2
for j, c in enumerate(cluster.labels_):
prev = cur
cur = c
if cur == prev:
try:
slices.pop(j + count)
except IndexError:
pass # FIXME why are some indices erroring?
count += 1
self.buffer[workable] = slices
def analyse(self, workable):
hsh = create_hash(workable, self.identity)
cache = self.process.cache / f"{hsh}.npy"
if cache.exists():
loudness = np.load(cache, allow_pickle=True)
else:
loudness = get_buffer(
fluid.loudness(workable["file"],
windowsize=self.windowsize,
hopsize=self.hopsize,
kweighting=self.kweighting,
truepeak=self.truepeak,
numframes=workable["numframes"],
startframe=workable["startframe"]),
"numpy",
)
np.save(cache, loudness)
workable["feature"] = loudness.tolist()
self.buffer[workable["id"]] = workable
def make_image(audio_file):
data = get_buffer(audio_file, 'numpy')
p = image_dir / audio_file.name
length = data.shape[0]
if length < w * h:
diff = (w * h) - length
data = np.pad(data, (diff - 1, 1), 'wrap')
container = np.zeros(shape=(w, h, 3))
for i, point in enumerate(data):
x = i % w - 1
y = round((i + 1) / h)
try:
scale = ((point + 1) * 0.5) * 256
container[x][y] = [scale, scale, scale]
except:
pass
img = Image.fromarray(container.astype('uint8'), 'RGB')
img.save(p.with_suffix('.png'))
def analyse(self, workable):
hsh = create_hash(workable, self.identity)
cache = self.process.cache / f"{hsh}.npy"
if cache.exists():
pitch = np.load(cache, allow_pickle=True)
else:
pitch = get_buffer(
fluid.pitch(workable["file"],
algorithm=self.algorithm,
minfreq=self.minfreq,
maxfreq=self.maxfreq,
unit=self.unit,
fftsettings=self.fftsettings,
numframes=workable["numframes"],
startframe=workable["startframe"]),
"numpy",
)
np.save(cache, pitch)
workable["features"] = pitch.tolist()
self.buffer[workable["id"]] = workable
def analyse_items(self):
# TODO import this function from a global place that can be used in CORPUS too
median_loudness = {}
for x in self.input:
hsh = create_hash(x, self.min_loudness, self.max_loudness)
cache = self.process.cache / f"{hsh}.npy"
if not cache.exists():
med_loudness = get_buffer(
stats(loudness(x, hopsize=4410, windowsize=17640)),
"numpy")
np.save(cache, med_loudness)
else:
med_loudness = np.load(cache, allow_pickle=True)
median_loudness[str(x)] = med_loudness[0][5]
vals = np.array([x for x in median_loudness.values()])
min_perc = np.percentile(vals, self.min_loudness)
max_perc = np.percentile(vals, self.max_loudness)
self.output = [
k for k, v in median_loudness.items()
if v <= max_perc and v >= min_perc
]
def loudness(self, min_loudness: int = 0, max_loudness: int = 100):
hopsize = 4410
windowsize = 17640
with Progress() as progress:
task = progress.add_task("[cyan]Corpus Filtering: Loudness",
total=len(self.items))
median_loudness = {}
for x in self.items:
hsh = create_hash(x, hopsize, windowsize)
# Make sure a sane temporary path exists
tmp = Path("/tmp") / "ftis_cache"
tmp.mkdir(exist_ok=True)
cache = tmp / f"{hsh}.npy"
if not cache.exists():
med_loudness = get_buffer(
stats(
loudness(x, hopsize=hopsize,
windowsize=windowsize)), "numpy")
np.save(cache, med_loudness)
else:
med_loudness = np.load(cache, allow_pickle=True)
median_loudness[str(x)] = med_loudness[0][5]
progress.update(task, advance=1)
# Get percentiles and filter
vals = np.array([x for x in median_loudness.values()])
min_perc = np.percentile(vals, min_loudness)
max_perc = np.percentile(vals, max_loudness)
self.items = [
k for k, v in median_loudness.items()
if v <= max_perc and v >= min_perc
]
return self
from sklearn.preprocessing import StandardScaler
from sklearn.cluster import AgglomerativeClustering
from datetime import datetime
THRESHOLD = 0.47
WINDOWSIZE = 25
HOPSIZE = 1
media = Path("../reaper/source/media/")
source = media / "02-200420_0928.wav"
source = source.resolve()
output = Path("slices").resolve()
print('Slicing')
slices = get_buffer(
fluid.noveltyslice(source, threshold=THRESHOLD, fftsettings=[2048, -1,
-1]))
slices = [int(x) for x in slices]
# clustering
standardise = StandardScaler()
original_slices = list(slices) # make a templated copy
tracks = {}
pos = 0
for i, (start, end) in enumerate(zip(original_slices, original_slices[1:])):
start = (start / 44100)
end = (end / 44100)
item = {
return [y - x for x, y in zip(arr, arr[1:])]
# files = [x for x in Path("outputs/micro_segmentation/2_ExplodeAudio").iterdir()]
# clusters = read_json("outputs/micro_clustering/5_HDBSCluster.json")
# files = [x for x in clusters["37"]]
files = [x for x in Path("outputs/concat").iterdir() if x.suffix == ".wav"]
print(files)
d = {}
for i, f in enumerate(files):
print(f)
print(i / len(files))
ts = get_buffer(fluid.transientslice(f))
if ts[0] != 0:
ts.insert(0, 0)
if len(ts) <= 2 and ts[0] == 0.0:
d[str(f)] = -1
else:
# Let's grab the orderedness of the onsets
norm = normalise(ts)
average = mean(norm)
robustified = [x / average for x in norm]
first_deriv = deriv(robustified)
d[str(f)] = stdev(first_deriv)
mi = 99999
MINDIST = 0.05 # UMAP minimum distance
PLOT = True
#HDBSCAN
HDBCLUSTSIZE = 3
HDBSAMPS = 1
media = Path("../reaper/source/media/")
source = media / "twovoice.wav"# "06-xbox controller-200518_1319.wav"# "02-200420_0928.wav"
data, labels = [], []
if not resynth.exists(): # let's not redo long NMF's each time
nmf = fluid.nmf(source, resynth=resynth, iterations=50, components=)
if not features.exists():
mfcc = fluid.mfcc(resynth, features=features, minfreq=500, maxfreq=15000, numcoeffs=13)
stats = get_buffer(fluid.stats(features, numderivs=1))
# flatten statistics because its channels...channels...channels
flatstats = []
for i in range(NMFCOMPONENTS):
offset = i * NUMCOEFFS
temp = []
for j in range(NUMCOEFFS):
for x in stats[j+offset]:
temp.append(x)
flatstats.append(temp)
# standardise data
print('Standardising Data')
standardise = StandardScaler()
data = np.array(flatstats)
# media = Path("../reaper/source/media/")
media = Path("../reaper/highgain_source/bounces/")
source = media / "highgain_source-002.wav" #"twovoice.wav"# "06-xbox controller-200518_1319.wav"# "02-200420_0928.wav"
bases = Path("bases.wav")
resynth = Path("resynth.wav")
data = []
if not bases.exists() or not resynth.exists():
nmf = fluid.nmf(source,
resynth=resynth,
bases=bases,
iterations=100,
components=NMFCOMPONENTS)
bases = get_buffer(bases, "numpy")
bases_smoothed = np.zeros_like(bases)
# lets smooth the bases a bit
for i, x in enumerate(bases):
bases_smoothed[i] = savgol_filter(x, 11, 2)
# clustering
print(f'Clustering data')
clusterer = hdbscan.HDBSCAN(min_cluster_size=HDBCLUSTSIZE,
min_samples=HDBSAMPS)
cluster_labels = clusterer.fit_predict(bases)
unique_clusters = list(dict.fromkeys(cluster_labels))
sound = get_buffer(resynth, "numpy")
HDBCLUSTSIZE = 2
HDBSAMPS = 1
media = Path("../reaper/source/media/")
source = media / "twovoice.wav" # "06-xbox controller-200518_1319.wav"# "02-200420_0928.wav"
output = Path("slices").resolve()
activation_pickle = Path("activations.wav")
data, labels = [], []
nmf = fluid.nmf(source,
activations=activation_pickle.resolve(),
iterations=50,
components=NMFCOMPONENTS)
activations = get_buffer(nmf.activations, "numpy")
stats = get_buffer(fluid.stats(nmf.activations, numderivs=1), "numpy")
# clustering
print(f'Clustering data')
clusterer = hdbscan.HDBSCAN(min_cluster_size=HDBCLUSTSIZE,
min_samples=HDBSAMPS)
cluster_labels = clusterer.fit_predict(stats)
unique_clusters = list(dict.fromkeys(cluster_labels))
sound = get_buffer(nmf.resynth, "numpy")
# Format everything into a lovely little reaper project
# initiate jinja2 business
env = jinja2.Environment(loader=jinja2.FileSystemLoader(['../RPRTemplates']))