-
Notifications
You must be signed in to change notification settings - Fork 931
/
spectral.py
2149 lines (1860 loc) · 74.7 KB
/
spectral.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Spectral feature extraction"""
import numpy as np
import scipy
import scipy.signal
import scipy.fftpack
from .. import util
from .. import filters
from ..util.exceptions import ParameterError
from ..core.convert import fft_frequencies
from ..core.audio import zero_crossings
from ..core.spectrum import power_to_db, _spectrogram
from ..core.constantq import cqt, hybrid_cqt, vqt
from ..core.pitch import estimate_tuning
from typing import Any, Optional, Union, Collection
from typing_extensions import Literal
from numpy.typing import DTypeLike
from .._typing import _FloatLike_co, _WindowSpec, _PadMode, _PadModeSTFT
__all__ = [
"spectral_centroid",
"spectral_bandwidth",
"spectral_contrast",
"spectral_rolloff",
"spectral_flatness",
"poly_features",
"rms",
"zero_crossing_rate",
"chroma_stft",
"chroma_cqt",
"chroma_cens",
"chroma_vqt",
"melspectrogram",
"mfcc",
"tonnetz",
]
# -- Spectral features -- #
def spectral_centroid(
*,
y: Optional[np.ndarray] = None,
sr: float = 22050,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
freq: Optional[np.ndarray] = None,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
) -> np.ndarray:
"""Compute the spectral centroid.
Each frame of a magnitude spectrogram is normalized and treated as a
distribution over frequency bins, from which the mean (centroid) is
extracted per frame.
More precisely, the centroid at frame ``t`` is defined as [#]_::
centroid[t] = sum_k S[k, t] * freq[k] / (sum_j S[j, t])
where ``S`` is a magnitude spectrogram, and ``freq`` is the array of
frequencies (e.g., FFT frequencies in Hz) of the rows of ``S``.
.. [#] Klapuri, A., & Davy, M. (Eds.). (2007). Signal processing
methods for music transcription, chapter 5.
Springer Science & Business Media.
Parameters
----------
y : np.ndarray [shape=(..., n,)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(..., d, t)] or None
(optional) spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
freq : None or np.ndarray [shape=(d,) or shape=(d, t)]
Center frequencies for spectrogram bins.
If `None`, then FFT bin center frequencies are used.
Otherwise, it can be a single array of ``d`` center frequencies,
or a matrix of center frequencies as constructed by
`librosa.reassigned_spectrogram`
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length ``win_length`` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
`t` is centered at ``y[t * hop_length]``.
- If `False`, then frame ``t`` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
Returns
-------
centroid : np.ndarray [shape=(..., 1, t)]
centroid frequencies
See Also
--------
librosa.stft : Short-time Fourier Transform
librosa.reassigned_spectrogram : Time-frequency reassigned spectrogram
Examples
--------
From time-series input:
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> cent = librosa.feature.spectral_centroid(y=y, sr=sr)
>>> cent
array([[1768.888, 1921.774, ..., 5663.477, 5813.683]])
From spectrogram input:
>>> S, phase = librosa.magphase(librosa.stft(y=y))
>>> librosa.feature.spectral_centroid(S=S)
array([[1768.888, 1921.774, ..., 5663.477, 5813.683]])
Using variable bin center frequencies:
>>> freqs, times, D = librosa.reassigned_spectrogram(y, fill_nan=True)
>>> librosa.feature.spectral_centroid(S=np.abs(D), freq=freqs)
array([[1768.838, 1921.801, ..., 5663.513, 5813.747]])
Plot the result
>>> import matplotlib.pyplot as plt
>>> times = librosa.times_like(cent)
>>> fig, ax = plt.subplots()
>>> librosa.display.specshow(librosa.amplitude_to_db(S, ref=np.max),
... y_axis='log', x_axis='time', ax=ax)
>>> ax.plot(times, cent.T, label='Spectral centroid', color='w')
>>> ax.legend(loc='upper right')
>>> ax.set(title='log Power spectrogram')
"""
# input is time domain:y or spectrogram:s
#
S, n_fft = _spectrogram(
y=y,
S=S,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
)
if not np.isrealobj(S):
raise ParameterError(
"Spectral centroid is only defined " "with real-valued input"
)
elif np.any(S < 0):
raise ParameterError(
"Spectral centroid is only defined " "with non-negative energies"
)
# Compute the center frequencies of each bin
if freq is None:
freq = fft_frequencies(sr=sr, n_fft=n_fft)
if freq.ndim == 1:
# reshape for broadcasting
freq = util.expand_to(freq, ndim=S.ndim, axes=-2)
# Column-normalize S
centroid: np.ndarray = np.sum(
freq * util.normalize(S, norm=1, axis=-2), axis=-2, keepdims=True
)
return centroid
def spectral_bandwidth(
*,
y: Optional[np.ndarray] = None,
sr: float = 22050,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
freq: Optional[np.ndarray] = None,
centroid: Optional[np.ndarray] = None,
norm: bool = True,
p: float = 2,
) -> np.ndarray:
"""Compute p'th-order spectral bandwidth.
The spectral bandwidth [#]_ at frame ``t`` is computed by::
(sum_k S[k, t] * (freq[k, t] - centroid[t])**p)**(1/p)
.. [#] Klapuri, A., & Davy, M. (Eds.). (2007). Signal processing
methods for music transcription, chapter 5.
Springer Science & Business Media.
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(..., d, t)] or None
(optional) spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length ``win_length`` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
``t`` is centered at ``y[t * hop_length]``.
- If ``False``, then frame ``t`` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
freq : None or np.ndarray [shape=(d,) or shape=(..., d, t)]
Center frequencies for spectrogram bins.
If `None`, then FFT bin center frequencies are used.
Otherwise, it can be a single array of ``d`` center frequencies,
or a matrix of center frequencies as constructed by
`librosa.reassigned_spectrogram`
centroid : None or np.ndarray [shape=(..., 1, t)]
pre-computed centroid frequencies
norm : bool
Normalize per-frame spectral energy (sum to one)
p : float > 0
Power to raise deviation from spectral centroid.
Returns
-------
bandwidth : np.ndarray [shape=(..., 1, t)]
frequency bandwidth for each frame
Examples
--------
From time-series input
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> spec_bw = librosa.feature.spectral_bandwidth(y=y, sr=sr)
>>> spec_bw
array([[1273.836, 1228.873, ..., 2952.357, 3013.68 ]])
From spectrogram input
>>> S, phase = librosa.magphase(librosa.stft(y=y))
>>> librosa.feature.spectral_bandwidth(S=S)
array([[1273.836, 1228.873, ..., 2952.357, 3013.68 ]])
Using variable bin center frequencies
>>> freqs, times, D = librosa.reassigned_spectrogram(y, fill_nan=True)
>>> librosa.feature.spectral_bandwidth(S=np.abs(D), freq=freqs)
array([[1274.637, 1228.786, ..., 2952.4 , 3013.735]])
Plot the result
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=2, sharex=True)
>>> times = librosa.times_like(spec_bw)
>>> centroid = librosa.feature.spectral_centroid(S=S)
>>> ax[0].semilogy(times, spec_bw[0], label='Spectral bandwidth')
>>> ax[0].set(ylabel='Hz', xticks=[], xlim=[times.min(), times.max()])
>>> ax[0].legend()
>>> ax[0].label_outer()
>>> librosa.display.specshow(librosa.amplitude_to_db(S, ref=np.max),
... y_axis='log', x_axis='time', ax=ax[1])
>>> ax[1].set(title='log Power spectrogram')
>>> ax[1].fill_between(times, np.maximum(0, centroid[0] - spec_bw[0]),
... np.minimum(centroid[0] + spec_bw[0], sr/2),
... alpha=0.5, label='Centroid +- bandwidth')
>>> ax[1].plot(times, centroid[0], label='Spectral centroid', color='w')
>>> ax[1].legend(loc='lower right')
"""
S, n_fft = _spectrogram(
y=y,
S=S,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
)
if not np.isrealobj(S):
raise ParameterError(
"Spectral bandwidth is only defined " "with real-valued input"
)
elif np.any(S < 0):
raise ParameterError(
"Spectral bandwidth is only defined " "with non-negative energies"
)
# centroid or center?
if centroid is None:
centroid = spectral_centroid(
y=y, sr=sr, S=S, n_fft=n_fft, hop_length=hop_length, freq=freq
)
# Compute the center frequencies of each bin
if freq is None:
freq = fft_frequencies(sr=sr, n_fft=n_fft)
if freq.ndim == 1:
deviation = np.abs(
np.subtract.outer(centroid[..., 0, :], freq).swapaxes(-2, -1)
)
else:
deviation = np.abs(freq - centroid)
# Column-normalize S
if norm:
S = util.normalize(S, norm=1, axis=-2)
bw: np.ndarray = np.sum(S * deviation**p, axis=-2, keepdims=True) ** (1.0 / p)
return bw
def spectral_contrast(
*,
y: Optional[np.ndarray] = None,
sr: float = 22050,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
freq: Optional[np.ndarray] = None,
fmin: float = 200.0,
n_bands: int = 6,
quantile: float = 0.02,
linear: bool = False,
) -> np.ndarray:
"""Compute spectral contrast
Each frame of a spectrogram ``S`` is divided into sub-bands.
For each sub-band, the energy contrast is estimated by comparing
the mean energy in the top quantile (peak energy) to that of the
bottom quantile (valley energy). High contrast values generally
correspond to clear, narrow-band signals, while low contrast values
correspond to broad-band noise. [#]_
.. [#] Jiang, Dan-Ning, Lie Lu, Hong-Jiang Zhang, Jian-Hua Tao,
and Lian-Hong Cai.
"Music type classification by spectral contrast feature."
In Multimedia and Expo, 2002. ICME'02. Proceedings.
2002 IEEE International Conference on, vol. 1, pp. 113-116.
IEEE, 2002.
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(..., d, t)] or None
(optional) spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length `win_length` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
``t`` is centered at ``y[t * hop_length]``.
- If `False`, then frame ``t`` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
freq : None or np.ndarray [shape=(d,)]
Center frequencies for spectrogram bins.
If `None`, then FFT bin center frequencies are used.
Otherwise, it can be a single array of ``d`` center frequencies.
fmin : float > 0
Frequency cutoff for the first bin ``[0, fmin]``
Subsequent bins will cover ``[fmin, 2*fmin]`, `[2*fmin, 4*fmin]``, etc.
n_bands : int > 1
number of frequency bands
quantile : float in (0, 1)
quantile for determining peaks and valleys
linear : bool
If `True`, return the linear difference of magnitudes:
``peaks - valleys``.
If `False`, return the logarithmic difference:
``log(peaks) - log(valleys)``.
Returns
-------
contrast : np.ndarray [shape=(..., n_bands + 1, t)]
each row of spectral contrast values corresponds to a given
octave-based frequency
Examples
--------
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> S = np.abs(librosa.stft(y))
>>> contrast = librosa.feature.spectral_contrast(S=S, sr=sr)
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=2, sharex=True)
>>> img1 = librosa.display.specshow(librosa.amplitude_to_db(S,
... ref=np.max),
... y_axis='log', x_axis='time', ax=ax[0])
>>> fig.colorbar(img1, ax=[ax[0]], format='%+2.0f dB')
>>> ax[0].set(title='Power spectrogram')
>>> ax[0].label_outer()
>>> img2 = librosa.display.specshow(contrast, x_axis='time', ax=ax[1])
>>> fig.colorbar(img2, ax=[ax[1]])
>>> ax[1].set(ylabel='Frequency bands', title='Spectral contrast')
"""
S, n_fft = _spectrogram(
y=y,
S=S,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
)
# Compute the center frequencies of each bin
if freq is None:
freq = fft_frequencies(sr=sr, n_fft=n_fft)
freq = np.atleast_1d(freq)
if freq.ndim != 1 or len(freq) != S.shape[-2]:
raise ParameterError(f"freq.shape mismatch: expected ({S.shape[-2]:d},)")
if n_bands < 1 or not isinstance(n_bands, (int, np.integer)):
raise ParameterError("n_bands must be a positive integer")
if not 0.0 < quantile < 1.0:
raise ParameterError("quantile must lie in the range (0, 1)")
if fmin <= 0:
raise ParameterError("fmin must be a positive number")
octa = np.zeros(n_bands + 2)
octa[1:] = fmin * (2.0 ** np.arange(0, n_bands + 1))
if np.any(octa[:-1] >= 0.5 * sr):
raise ParameterError(
"Frequency band exceeds Nyquist. " "Reduce either fmin or n_bands."
)
# shape of valleys and peaks based on spectrogram
shape = list(S.shape)
shape[-2] = n_bands + 1
valley = np.zeros(shape)
peak = np.zeros_like(valley)
for k, (f_low, f_high) in enumerate(zip(octa[:-1], octa[1:])):
current_band = np.logical_and(freq >= f_low, freq <= f_high)
idx = np.flatnonzero(current_band)
if k > 0:
current_band[idx[0] - 1] = True
if k == n_bands:
current_band[idx[-1] + 1 :] = True
sub_band = S[..., current_band, :]
if k < n_bands:
sub_band = sub_band[..., :-1, :]
# Always take at least one bin from each side
idx = np.rint(quantile * np.sum(current_band))
idx = int(np.maximum(idx, 1))
sortedr = np.sort(sub_band, axis=-2)
valley[..., k, :] = np.mean(sortedr[..., :idx, :], axis=-2)
peak[..., k, :] = np.mean(sortedr[..., -idx:, :], axis=-2)
contrast: np.ndarray
if linear:
contrast = peak - valley
else:
contrast = power_to_db(peak) - power_to_db(valley)
return contrast
def spectral_rolloff(
*,
y: Optional[np.ndarray] = None,
sr: float = 22050,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
freq: Optional[np.ndarray] = None,
roll_percent: float = 0.85,
) -> np.ndarray:
"""Compute roll-off frequency.
The roll-off frequency is defined for each frame as the center frequency
for a spectrogram bin such that at least roll_percent (0.85 by default)
of the energy of the spectrum in this frame is contained in this bin and
the bins below. This can be used to, e.g., approximate the maximum (or
minimum) frequency by setting roll_percent to a value close to 1 (or 0).
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(d, t)] or None
(optional) spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length `win_length` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
``t`` is centered at ``y[t * hop_length]``.
- If `False`, then frame ``t`` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
freq : None or np.ndarray [shape=(d,) or shape=(..., d, t)]
Center frequencies for spectrogram bins.
If `None`, then FFT bin center frequencies are used.
Otherwise, it can be a single array of ``d`` center frequencies,
.. note:: ``freq`` is assumed to be sorted in increasing order
roll_percent : float [0 < roll_percent < 1]
Roll-off percentage.
Returns
-------
rolloff : np.ndarray [shape=(..., 1, t)]
roll-off frequency for each frame
Examples
--------
From time-series input
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> # Approximate maximum frequencies with roll_percent=0.85 (default)
>>> librosa.feature.spectral_rolloff(y=y, sr=sr)
array([[2583.984, 3036.182, ..., 9173.145, 9248.511]])
>>> # Approximate maximum frequencies with roll_percent=0.99
>>> rolloff = librosa.feature.spectral_rolloff(y=y, sr=sr, roll_percent=0.99)
>>> rolloff
array([[ 7192.09 , 6739.893, ..., 10960.4 , 10992.7 ]])
>>> # Approximate minimum frequencies with roll_percent=0.01
>>> rolloff_min = librosa.feature.spectral_rolloff(y=y, sr=sr, roll_percent=0.01)
>>> rolloff_min
array([[516.797, 538.33 , ..., 764.429, 764.429]])
From spectrogram input
>>> S, phase = librosa.magphase(librosa.stft(y))
>>> librosa.feature.spectral_rolloff(S=S, sr=sr)
array([[2583.984, 3036.182, ..., 9173.145, 9248.511]])
>>> # With a higher roll percentage:
>>> librosa.feature.spectral_rolloff(y=y, sr=sr, roll_percent=0.95)
array([[ 3919.043, 3994.409, ..., 10443.604, 10594.336]])
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots()
>>> librosa.display.specshow(librosa.amplitude_to_db(S, ref=np.max),
... y_axis='log', x_axis='time', ax=ax)
>>> ax.plot(librosa.times_like(rolloff), rolloff[0], label='Roll-off frequency (0.99)')
>>> ax.plot(librosa.times_like(rolloff), rolloff_min[0], color='w',
... label='Roll-off frequency (0.01)')
>>> ax.legend(loc='lower right')
>>> ax.set(title='log Power spectrogram')
"""
if not 0.0 < roll_percent < 1.0:
raise ParameterError("roll_percent must lie in the range (0, 1)")
S, n_fft = _spectrogram(
y=y,
S=S,
n_fft=n_fft,
hop_length=hop_length,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
)
if not np.isrealobj(S):
raise ParameterError(
"Spectral rolloff is only defined " "with real-valued input"
)
elif np.any(S < 0):
raise ParameterError(
"Spectral rolloff is only defined " "with non-negative energies"
)
# Compute the center frequencies of each bin
if freq is None:
freq = fft_frequencies(sr=sr, n_fft=n_fft)
# Make sure that frequency can be broadcast
if freq.ndim == 1:
# reshape for broadcasting
freq = util.expand_to(freq, ndim=S.ndim, axes=-2)
total_energy = np.cumsum(S, axis=-2)
# (channels,freq,frames)
threshold = roll_percent * total_energy[..., -1, :]
# reshape threshold for broadcasting
threshold = np.expand_dims(threshold, axis=-2)
ind = np.where(total_energy < threshold, np.nan, 1)
rolloff: np.ndarray = np.nanmin(ind * freq, axis=-2, keepdims=True)
return rolloff
def spectral_flatness(
*,
y: Optional[np.ndarray] = None,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
amin: float = 1e-10,
power: float = 2.0,
) -> np.ndarray:
"""Compute spectral flatness
Spectral flatness (or tonality coefficient) is a measure to
quantify how much noise-like a sound is, as opposed to being
tone-like [#]_. A high spectral flatness (closer to 1.0)
indicates the spectrum is similar to white noise.
It is often converted to decibel.
.. [#] Dubnov, Shlomo "Generalization of spectral flatness
measure for non-gaussian linear processes"
IEEE Signal Processing Letters, 2004, Vol. 11.
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
S : np.ndarray [shape=(..., d, t)] or None
(optional) pre-computed spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length `win_length` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
``t`` is centered at ``y[t * hop_length]``.
- If `False`, then frame `t` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
amin : float > 0 [scalar]
minimum threshold for ``S`` (=added noise floor for numerical stability)
power : float > 0 [scalar]
Exponent for the magnitude spectrogram.
e.g., 1 for energy, 2 for power, etc.
Power spectrogram is usually used for computing spectral flatness.
Returns
-------
flatness : np.ndarray [shape=(..., 1, t)]
spectral flatness for each frame.
The returned value is in [0, 1] and often converted to dB scale.
Examples
--------
From time-series input
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> flatness = librosa.feature.spectral_flatness(y=y)
>>> flatness
array([[0.001, 0. , ..., 0.218, 0.184]], dtype=float32)
From spectrogram input
>>> S, phase = librosa.magphase(librosa.stft(y))
>>> librosa.feature.spectral_flatness(S=S)
array([[0.001, 0. , ..., 0.218, 0.184]], dtype=float32)
From power spectrogram input
>>> S, phase = librosa.magphase(librosa.stft(y))
>>> S_power = S ** 2
>>> librosa.feature.spectral_flatness(S=S_power, power=1.0)
array([[0.001, 0. , ..., 0.218, 0.184]], dtype=float32)
"""
if amin <= 0:
raise ParameterError("amin must be strictly positive")
S, n_fft = _spectrogram(
y=y,
S=S,
n_fft=n_fft,
hop_length=hop_length,
power=1.0,
win_length=win_length,
window=window,
center=center,
pad_mode=pad_mode,
)
if not np.isrealobj(S):
raise ParameterError(
"Spectral flatness is only defined " "with real-valued input"
)
elif np.any(S < 0):
raise ParameterError(
"Spectral flatness is only defined " "with non-negative energies"
)
S_thresh = np.maximum(amin, S**power)
gmean = np.exp(np.mean(np.log(S_thresh), axis=-2, keepdims=True))
amean = np.mean(S_thresh, axis=-2, keepdims=True)
flatness: np.ndarray = gmean / amean
return flatness
def rms(
*,
y: Optional[np.ndarray] = None,
S: Optional[np.ndarray] = None,
frame_length: int = 2048,
hop_length: int = 512,
center: bool = True,
pad_mode: _PadMode = "constant",
dtype: DTypeLike = np.float32,
) -> np.ndarray:
"""Compute root-mean-square (RMS) value for each frame, either from the
audio samples ``y`` or from a spectrogram ``S``.
Computing the RMS value from audio samples is faster as it doesn't require
a STFT calculation. However, using a spectrogram will give a more accurate
representation of energy over time because its frames can be windowed,
thus prefer using ``S`` if it's already available.
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
(optional) audio time series. Required if ``S`` is not input.
Multi-channel is supported.
S : np.ndarray [shape=(..., d, t)] or None
(optional) spectrogram magnitude. Required if ``y`` is not input.
frame_length : int > 0 [scalar]
length of analysis frame (in samples) for energy calculation
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
center : bool
If `True` and operating on time-domain input (``y``), pad the signal
by ``frame_length//2`` on either side.
If operating on spectrogram input, this has no effect.
pad_mode : str
Padding mode for centered analysis. See `numpy.pad` for valid
values.
dtype : np.dtype, optional
Data type of the output array. Defaults to float32.
Returns
-------
rms : np.ndarray [shape=(..., 1, t)]
RMS value for each frame
Examples
--------
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> librosa.feature.rms(y=y)
array([[1.248e-01, 1.259e-01, ..., 1.845e-05, 1.796e-05]],
dtype=float32)
Or from spectrogram input
>>> S, phase = librosa.magphase(librosa.stft(y))
>>> rms = librosa.feature.rms(S=S)
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=2, sharex=True)
>>> times = librosa.times_like(rms)
>>> ax[0].semilogy(times, rms[0], label='RMS Energy')
>>> ax[0].set(xticks=[])
>>> ax[0].legend()
>>> ax[0].label_outer()
>>> librosa.display.specshow(librosa.amplitude_to_db(S, ref=np.max),
... y_axis='log', x_axis='time', ax=ax[1])
>>> ax[1].set(title='log Power spectrogram')
Use a STFT window of constant ones and no frame centering to get consistent
results with the RMS computed from the audio samples ``y``
>>> S = librosa.magphase(librosa.stft(y, window=np.ones, center=False))[0]
>>> librosa.feature.rms(S=S)
>>> plt.show()
"""
if y is not None:
if center:
padding = [(0, 0) for _ in range(y.ndim)]
padding[-1] = (int(frame_length // 2), int(frame_length // 2))
y = np.pad(y, padding, mode=pad_mode)
x = util.frame(y, frame_length=frame_length, hop_length=hop_length)
# Calculate power
power = np.mean(util.abs2(x, dtype=dtype), axis=-2, keepdims=True)
elif S is not None:
# Check the frame length
if S.shape[-2] != frame_length // 2 + 1:
raise ParameterError(
"Since S.shape[-2] is {}, "
"frame_length is expected to be {} or {}; "
"found {}".format(
S.shape[-2], S.shape[-2] * 2 - 2, S.shape[-2] * 2 - 1, frame_length
)
)
# power spectrogram
x = util.abs2(S, dtype=dtype)
# Adjust the DC and sr/2 component
x[..., 0, :] *= 0.5
if frame_length % 2 == 0:
x[..., -1, :] *= 0.5
# Calculate power
power = 2 * np.sum(x, axis=-2, keepdims=True) / frame_length**2
else:
raise ParameterError("Either `y` or `S` must be input.")
rms_result: np.ndarray = np.sqrt(power)
return rms_result
def poly_features(
*,
y: Optional[np.ndarray] = None,
sr: float = 22050,
S: Optional[np.ndarray] = None,
n_fft: int = 2048,
hop_length: int = 512,
win_length: Optional[int] = None,
window: _WindowSpec = "hann",
center: bool = True,
pad_mode: _PadModeSTFT = "constant",
order: int = 1,
freq: Optional[np.ndarray] = None,
) -> np.ndarray:
"""Get coefficients of fitting an nth-order polynomial to the columns
of a spectrogram.
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(..., d, t)] or None
(optional) spectrogram magnitude
n_fft : int > 0 [scalar]
FFT window size
hop_length : int > 0 [scalar]
hop length for STFT. See `librosa.stft` for details.
win_length : int <= n_fft [scalar]
Each frame of audio is windowed by `window()`.
The window will be of length `win_length` and then padded
with zeros to match ``n_fft``.
If unspecified, defaults to ``win_length = n_fft``.
window : string, tuple, number, function, or np.ndarray [shape=(n_fft,)]
- a window specification (string, tuple, or number);
see `scipy.signal.get_window`
- a window function, such as `scipy.signal.windows.hann`
- a vector or array of length ``n_fft``
.. see also:: `librosa.filters.get_window`
center : boolean
- If `True`, the signal ``y`` is padded so that frame
`t` is centered at ``y[t * hop_length]``.
- If `False`, then frame ``t`` begins at ``y[t * hop_length]``
pad_mode : string
If ``center=True``, the padding mode to use at the edges of the signal.
By default, STFT uses zero padding.
order : int > 0
order of the polynomial to fit
freq : None or np.ndarray [shape=(d,) or shape=(..., d, t)]
Center frequencies for spectrogram bins.
If `None`, then FFT bin center frequencies are used.
Otherwise, it can be a single array of ``d`` center frequencies,
or a matrix of center frequencies as constructed by
`librosa.reassigned_spectrogram`
Returns
-------
coefficients : np.ndarray [shape=(..., order+1, t)]
polynomial coefficients for each frame.
``coefficients[..., 0, :]`` corresponds to the highest degree (``order``),
``coefficients[..., 1, :]`` corresponds to the next highest degree (``order-1``),
down to the constant term ``coefficients[..., order, :]``.
Examples
--------
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> S = np.abs(librosa.stft(y))
Fit a degree-0 polynomial (constant) to each frame
>>> p0 = librosa.feature.poly_features(S=S, order=0)
Fit a linear polynomial to each frame
>>> p1 = librosa.feature.poly_features(S=S, order=1)
Fit a quadratic to each frame
>>> p2 = librosa.feature.poly_features(S=S, order=2)
Plot the results for comparison
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=4, sharex=True, figsize=(8, 8))