Skip to content

Commit

Permalink
added fmin/max to melspec, fixed logamplitude
Browse files Browse the repository at this point in the history
  • Loading branch information
@bmcfee
bmcfee committed Mar 22, 2013
1 parent c3accf3 commit 14056d4
Show file tree
Hide file tree
Showing 4 changed files with 78 additions and 64 deletions.
35 changes: 2 additions & 33 deletions librosa/__init__.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -193,11 +193,10 @@ def logamplitude(S, amin=1e-10, gain_threshold=-80.0):
D = S in dBs
'''

SCALE = 20.0
D = SCALE * numpy.log10(numpy.maximum(amin, numpy.abs(S)))
D = 20.0 * numpy.log10(numpy.maximum(amin, numpy.abs(S)))

if gain_threshold is not None:
D[D < gain_threshold] = gain_threshold
D = numpy.maximum(D, D.max() + gain_threshold)
pass

return D
Expand All @@ -219,36 +218,6 @@ def frames_to_time(frames, sr=22050, hop_length=64):
return frames * float(hop_length) / float(sr)


def feature_sync(X, F, agg=numpy.mean):
'''
Synchronous aggregation of a feature matrix
Input:
X: d-by-T | feature matrix
F: t-vector | (ordered) array of frame numbers
agg: aggregator function | default: numpy.mean
Output:
Y: d-by-(<=t+1) vector
where
Y[:, i] = agg(X[:, F[i-1]:F[i]], axis=1)
In order to ensure total coverage, boundary points are added to F
'''

F = numpy.unique(numpy.concatenate( ([0], F, [X.shape[1]]) ))

Y = numpy.zeros( (X.shape[0], len(F)-1) )

lb = F[0]

for (i, ub) in enumerate(F[1:]):
Y[:, i] = agg(X[:, lb:ub], axis=1)
lb = ub
pass

return Y

def pad(w, d_pad, v=0.0, center=True):
'''
Pad a vector w out to d dimensions, using value v
Expand Down
48 changes: 19 additions & 29 deletions librosa/beat.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -185,16 +185,16 @@ def onset_estimate_bpm(onsets, start_bpm, sr, hop_length):

return start_bpm

def onset_strength_percussive(y, sr=22050, window_length=2048, hop_length=256, mel_channels=128, S=None):
def onset_strength_percussive(y, sr=22050, window_length=256, hop_length=32, mel_channels=40, S=None):
'''
Onset strength derived from harmonic-percussive source separation
Input:
y: time series signal
sr: sample rate of y | default: 22050
window_length: fourier analysis window length | default: 2048
hop_length: number of frames to hop | default: 256
mel_channels: number of mel bins to use | default: 128
window_length: fourier analysis window length | default: 256
hop_length: number of frames to hop | default: 32
mel_channels: number of mel bins to use | default: 40
'''

# Step 1: compute spectrogram
Expand Down Expand Up @@ -231,7 +231,7 @@ def onset_strength_percussive(y, sr=22050, window_length=2048, hop_length=256, m

return O / Onorm

def onset_strength(y, sr=22050, window_length=2048, hop_length=256, mel_channels=40, rising=True, htk=False, S=None):
def onset_strength(y, sr=22050, window_length=256, hop_length=32, mel_channels=40, htk=False, S=None):
'''
Adapted from McVicar, adapted from Ellis, etc...
Expand All @@ -240,10 +240,9 @@ def onset_strength(y, sr=22050, window_length=2048, hop_length=256, mel_channels
INPUT:
y = time-series waveform (t-by-1 vector)
sr = sampling rate of the input signal | default: 22050
window_length = number of samples per frame | default: 2048 | = 92.8ms @ 22KHz
hop_length = offset between frames | default: 256 | = 11.6ms @ 22KHz
window_length = number of samples per frame | default: 256
hop_length = offset between frames | default: 32
mel_channels = number of Mel bins to use | default: 40
rising = detect only rising edges of beats | default: True
htk = use HTK mels instead of Slaney | default: False
S = (optional) pre-computed spectrogram | default: None
Expand All @@ -252,51 +251,42 @@ def onset_strength(y, sr=22050, window_length=2048, hop_length=256, mel_channels
onset_envelope
'''

gain_threshold = 80.0

# First, compute mel spectrogram
if S is None:
S = librosa.feature.melspectrogram(y, sr=sr,
window_length=window_length,
hop_length=hop_length,
mel_channels=mel_channels,
S = librosa.feature.melspectrogram(y, sr = sr,
window_length = window_length,
hop_length = hop_length,
mel_channels = mel_channels,
htk=htk)
# Convert to dBs
S = librosa.logamplitude(S)

pass

### Only look at top 80 dB
onsets = numpy.maximum(S, S.max() - gain_threshold)

### Compute first difference
onsets = numpy.diff(onsets, n=1, axis=1)
onsets = numpy.diff(S, n=1, axis=1)

### Discard negatives (decreasing amplitude)
# falling edges could also be useful segmentation cues
# to catch falling edges, replace max(0,D) with abs(D)
if rising:
onsets = numpy.maximum(0.0, onsets)
else:
onsets = onsets**2
pass
onsets = numpy.maximum(0.0, onsets)

### Average over mel bands
onsets = numpy.mean(onsets, axis=0)
onsets = numpy.mean(onsets, axis=0)

### remove the DC component
onsets = scipy.signal.lfilter([1.0, -1.0], [1.0, -0.99], onsets)
onsets = scipy.signal.lfilter([1.0, -1.0], [1.0, -0.99], onsets)

### Threshold at zero
onsets = numpy.maximum(0.0, onsets)

### Normalize by the maximum onset strength
Onorm = numpy.max(onsets)
if Onorm == 0:
Onorm = 1.0
Onorm = onsets.max()
if Onorm > 0:
onsets = onsets / Onorm
pass

return onsets / Onorm
return onsets

def segment(X, k):
'''
Expand Down
37 changes: 35 additions & 2 deletions librosa/feature.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -297,7 +297,7 @@ def melfb(sr, nfft, nfilts=40, width=1.0, fmin=0.0, fmax=None, use_htk=False):

return wts

def melspectrogram(y, sr=22050, window_length=256, hop_length=128, mel_channels=40, htk=False, width=1):
def melspectrogram(y, sr=22050, window_length=256, hop_length=128, mel_channels=40, htk=False, width=1, fmin=0, fmax=None):
'''
Compute a mel spectrogram from a time series
Expand All @@ -309,6 +309,8 @@ def melspectrogram(y, sr=22050, window_length=256, hop_length=128, mel_channels=
mel_channels = number of Mel filters to use | default: 40
htk = use HTK mels instead of Slaney | default: False
width = width of mel bins | default: 1
fmin = minimum frequency | default: 0
fmax = maximum frequency | default: sr/2
Output:
S = Mel amplitude spectrogram
Expand All @@ -318,7 +320,7 @@ def melspectrogram(y, sr=22050, window_length=256, hop_length=128, mel_channels=
S = librosa.stft(y, sr=sr, n_fft=window_length, hann_w=window_length, hop_length=hop_length)

# Build a Mel filter
M = melfb(sr, window_length, nfilts=mel_channels, width=width, use_htk=htk)
M = melfb(sr, window_length, nfilts=mel_channels, width=width, use_htk=htk, fmin=fmin, fmax=fmax)

# Remove everything past the nyquist frequency
M = M[:, :(window_length / 2 + 1)]
Expand All @@ -327,3 +329,34 @@ def melspectrogram(y, sr=22050, window_length=256, hop_length=128, mel_channels=

return S

#-- miscellaneous utilities --#
def sync(X, F, agg=numpy.mean):
'''
Synchronous aggregation of a feature matrix
Input:
X: d-by-T | feature matrix
F: t-vector | (ordered) array of frame numbers
agg: aggregator function | default: numpy.mean
Output:
Y: d-by-(<=t+1) vector
where
Y[:, i] = agg(X[:, F[i-1]:F[i]], axis=1)
In order to ensure total coverage, boundary points are added to F
'''

F = numpy.unique(numpy.concatenate( ([0], F, [X.shape[1]]) ))

Y = numpy.zeros( (X.shape[0], len(F)-1) )

lb = F[0]

for (i, ub) in enumerate(F[1:]):
Y[:, i] = agg(X[:, lb:ub], axis=1)
lb = ub
pass

return Y

22 changes: 22 additions & 0 deletions tests/makeTestData.m
View file
Original file line number Diff line number Diff line change
Expand Up @@ -51,6 +51,9 @@ function testData(source_path, output_path)
display('resample');
testResample(output_path);

display('tempo');
testOnset(output_path);

%% Done!
display('Done.');
end
Expand Down Expand Up @@ -327,3 +330,22 @@ function testISTFT(output_path)
end
end
end

function testOnset(output_path)

wavfile = 'data/test2_8000.wav';

[y, sr] = wavread(wavfile);
y = mean(y, 2); % Convert to mono

% Generate the onset envelope first
[t, xcr, D, onsetenv, oesr] = tempo2(y, sr);

filename = sprintf('%s/beat-onset-000.mat', output_path);
display([' `-- saving ', filename]);
save(filename, 'wavfile', 'onsetenv');

filename = sprintf('%s/beat-tempo-000.mat', output_path);
display([' `-- saving ', filename]);
save(filename, 'wavfile', 't');
end

0 comments on commit 14056d4

Please sign in to comment.