Merge pull request #5 from meinardmueller/sjs_dependencies

Version 1.1.2

libtsm/pitchshift.py

@@ -11,7 +11,7 @@
 import scipy.interpolate
 from fractions import Fraction as frac
 from .tsm import hps_tsm
-from .utils import normalize_length
+from .utils import ensure_validity, normalize_length
 
 
 def pitch_shift_original(x, n, Fs=22050) -> np.ndarray:
@@ -136,6 +136,8 @@ def pitch_shift(x, p, t_p=None, Fs=22050, order="res-tsm", **kwargs) -> np.ndarr
         # only keep unique indices
         anchor_points = anchor_points[np.unique(anchor_points[:, 1], return_index=True)[1], :]
 
+        anchor_points = ensure_validity(anchor_points)
+
         # Time-Scale Modification
         y_ps = hps_tsm(y_ps, anchor_points, Fs=Fs, **kwargs)
 
@@ -145,6 +147,8 @@ def pitch_shift(x, p, t_p=None, Fs=22050, order="res-tsm", **kwargs) -> np.ndarr
         # only keep unique indices
         anchor_points = anchor_points[np.unique(anchor_points[:, 1], return_index=True)[1], :]
 
+        anchor_points = ensure_validity(anchor_points)
+
         # Time-Scale Modification
         y_tsm = hps_tsm(x, anchor_points, Fs=Fs, **kwargs)
 

libtsm/utils.py

@@ -30,7 +30,14 @@ def ensure_validity(alpha, syn_hop=128):
     """
     d_ana = np.diff(alpha[:,0])
     d_syn = np.diff(alpha[:,1])
-    not_too_steep = np.pad(np.round(d_syn / d_ana) <= syn_hop, (1,0), constant_values=True)
+    # divide considering d_ana = 0
+    ratio = np.divide(d_syn, d_ana, out=np.ones_like(d_syn, dtype=float) * syn_hop + 1, where=(d_ana != 0))
+    not_too_steep = np.pad(np.round(ratio) <= syn_hop, (1,0), constant_values=True)
+
+    # make sure that the last point is retained (and try to remove the second-to-last instead)
+    if not_too_steep[-1] == False:
+        not_too_steep[-2] = False
+        not_too_steep[-1] = True
 
     # call recursively to remove consecutive points that are too steep
     if not np.all(not_too_steep):

setup.py

@@ -6,12 +6,12 @@
 
 setup(
     name='libtsm',
-    version='1.1.1',
+    version='1.1.2',
     description='Python Package for Time-Scale Modification and Pitch-Shifting',
     author='Sebastian Rosenzweig, Simon Schwär, Jonathan Driedger and Meinard Müller',
     author_email='[email protected]',
     url='https://www.audiolabs-erlangen.de/resources/MIR/2021-DAFX-AdaptivePitchShifting',
-    download_url='https://github.com/meinardmueller/libtsm/archive/refs/tags/v1.1.0.tar.gz',
+    download_url='https://github.com/meinardmueller/libtsm/archive/refs/tags/v1.1.2.tar.gz',
     packages=find_packages(),
     long_description=long_description,
     long_description_content_type='text/markdown',
@@ -24,19 +24,24 @@
     ],
     keywords=['audio', 'music', 'tsm', 'pitch-shifting'],
     license='MIT',
-    install_requires=['ipython >= 7.8.0',
-                      'librosa >= 0.8.0',
-                      'matplotlib >= 3.1.0',
-                      'numba >= 0.51.0',
-                      'numpy >= 1.17.0',
-                      'pysoundfile >= 0.9.0',
+    install_requires=['numpy >= 1.17.0',
                       'scipy >= 1.3.0'],
     python_requires='>=3.6',
     extras_require={
-        'dev': ['jupyter == 1.0.*',
-                'nbstripout == 0.4.*'],
-        'tests': ['pytest == 6.2.*'],
-        'docs': ['sphinx == 4.0.*',
-                 'sphinx-rtd-theme == 0.5.*'],
+        'dev': [ # required for running the Jupyter notebook `demo_libtsm.ipynb
+            'ipython >= 7.8.0',
+            'jupyter == 1.0.*',
+            'librosa >= 0.8.0',
+            'nbstripout == 0.4.*',
+            'matplotlib >= 3.1.0',
+        ],
+        'test': [
+            'pytest == 6.2.*',
+            'soundfile >= 0.9.0',
+        ],
+        'docs': [
+            'sphinx == 4.0.*',
+            'sphinx-rtd-theme == 0.5.*',
+        ],
     }
 )

test/test_pitch_shifting.py

@@ -7,12 +7,11 @@
 
 import numpy as np
 import libtsm
-import librosa
+import soundfile as sf
 
 
 def test_fixed_pitch_shift_even():
-    fn_in = 'data/three_sinusoidals.wav'
-    x, Fs = librosa.load(fn_in, sr=22050)
+    x, Fs = sf.read('data/three_sinusoidals.wav')
 
     p = 1200  # cents
     y_psf1 = libtsm.pitch_shift(x, p, order='res-tsm')[:, 0]
@@ -25,8 +24,7 @@ def test_fixed_pitch_shift_even():
 
 
 def test_fixed_pitch_shift_odd():
-    fn_in = 'data/three_sinusoidals.wav'
-    x, Fs = librosa.load(fn_in, sr=22050)
+    x, Fs = sf.read('data/three_sinusoidals.wav')
 
     p = 1231  # cents
     y_psf1 = libtsm.pitch_shift(x, p, order='res-tsm')[:, 0]
@@ -59,9 +57,7 @@ def test_adaptive_pitch_shift1():
 
 
 def test_adaptive_pitch_shift2():
-    # Load Audio File
-    fn_in = 'data/three_sinusoidals.wav'
-    x, Fs = librosa.load(fn_in, sr=22050)
+    x, Fs = sf.read('data/three_sinusoidals.wav')
 
     # Adaptive Pitch-Shifting (Figure 3)
     t = np.arange(0, len(x) / Fs, 1 / Fs)  # sec
@@ -81,7 +77,7 @@ def test_adaptive_pitch_shift2():
 
 def test_adaptive_pitch_shift3():
     Fs = 48000
-    t = np.linspace(0, 5, int(5*Fs))
+    t = np.arange(int(5*Fs)) / Fs
     x = 0.5 * np.sin(2 * np.pi * 220 * t)
 
     p = np.array([0, 100, 50, -100, 0])
@@ -92,3 +88,19 @@ def test_adaptive_pitch_shift3():
 
     assert len(x) == len(y1)
     assert len(x) == len(y2)
+
+def test_adaptive_pitch_shift_random():
+    Fs = 44100
+    x = np.random.rand(int(5*Fs)) - 0.5
+
+    # try 20 different random trajectories
+    for _ in range(20):
+        N = np.random.randint(1, 250)
+        p = np.random.rand(N) * 1200 - 600
+        t_p = np.linspace(0, 5, N)
+
+        y1 = libtsm.pitch_shift(x, p, t_p, Fs=Fs, order='res-tsm')[:, 0]
+        y2 = libtsm.pitch_shift(x, p, t_p, Fs=Fs, order='tsm-res')[:, 0]
+
+        assert len(x) == len(y1)
+        assert len(x) == len(y2)