fix resampling scale for CQT

undid test relaxation for cqt position

updated image comparison test fixtures for cqt-based outputs

fixed resample-scaling, should be by sqrt(ratio)

Author: bmcfee

librosa/core/audio.py

@@ -203,7 +203,7 @@ def to_mono(y):
 
 
 @cache
-def resample(y, orig_sr, target_sr, res_type='sinc_best', fix=True, **kwargs):
+def resample(y, orig_sr, target_sr, res_type='sinc_best', fix=True, scale=False, **kwargs):
     """Resample a time series from orig_sr to target_sr
 
     Parameters
@@ -232,6 +232,10 @@ def resample(y, orig_sr, target_sr, res_type='sinc_best', fix=True, **kwargs):
         adjust the length of the resampled signal to be of size exactly
         `ceil(target_sr * len(y) / orig_sr)`
 
+    scale : bool
+        Scale the resampled signal so that `y` and `y_hat` have approximately
+        equal total energy.
+
     kwargs : additional keyword arguments
         If `fix==True`, additional keyword arguments to pass to
         `librosa.util.fix_length`.
@@ -271,20 +275,23 @@ def resample(y, orig_sr, target_sr, res_type='sinc_best', fix=True, **kwargs):
     if orig_sr == target_sr:
         return y
 
-    n_samples = int(np.ceil(y.shape[-1] * float(target_sr) / orig_sr))
+    ratio = float(target_sr) / orig_sr
+
+    n_samples = int(np.ceil(y.shape[-1] * ratio))
 
     scipy_resample = (res_type == 'scipy')
 
     if _HAS_SAMPLERATE and not scipy_resample:
-        y_hat = samplerate.resample(y.T,
-                                    float(target_sr) / orig_sr,
-                                    res_type).T
+        y_hat = samplerate.resample(y.T, ratio, res_type).T
     else:
         y_hat = scipy.signal.resample(y, n_samples, axis=-1)
 
     if fix:
         y_hat = util.fix_length(y_hat, n_samples, **kwargs)
 
+    if scale:
+        y_hat /= np.sqrt(ratio)
+
     return np.ascontiguousarray(y_hat, dtype=y.dtype)
 
 

librosa/core/constantq.py

@@ -248,7 +248,7 @@ def cqt(y, sr=22050, hop_length=512, fmin=None, n_bins=84,
 
         # Resample (except first time)
         if i > 0:
-            my_y = audio.resample(my_y, my_sr, my_sr/2.0, res_type=res_type)
+            my_y = audio.resample(my_y, my_sr, my_sr/2.0, res_type=res_type, scale=True)
             my_sr /= 2.0
             assert my_hop % 2 == 0
             my_hop //= 2
@@ -593,7 +593,7 @@ def __early_downsample(y, sr, hop_length, res_type, n_octaves,
         assert hop_length % downsample_factor == 0
         hop_length //= downsample_factor
 
-        y = audio.resample(y, sr, sr / downsample_factor, res_type=res_type)
+        y = audio.resample(y, sr, sr / downsample_factor, res_type=res_type, scale=True)
 
         sr /= downsample_factor
 

tests/baseline_images/test_display/cqt_hz.png

@@ -169,7 +169,6 @@ def __test(note_min):
         # Make sure that the max outside the peak is sufficiently small
         Cscale = Cbar / Cbar[idx]
         Cscale[idx] = np.nan
-
         assert np.nanmax(Cscale) < 1e-1
 
         Cscale[idx-1:idx+2] = np.nan

tests/baseline_images/test_display/cqt_note.png

@@ -81,11 +81,11 @@ def test_segment_load():
 
 def test_resample_mono():
 
-    def __test(y, sr_in, sr_out, res_type, fix):
+    def __test(y, sr_in, sr_out, res_type, fix, scale):
 
         y2 = librosa.resample(y, sr_in, sr_out,
                               res_type=res_type,
-                              fix=fix)
+                              fix=fix, scale=scale)
 
         # First, check that the audio is valid
         librosa.util.valid_audio(y2, mono=True)
@@ -109,7 +109,8 @@ def __test(y, sr_in, sr_out, res_type, fix):
         for sr_out in [8000, 22050]:
             for res_type in ['sinc_fastest', 'sinc_best', 'scipy']:
                 for fix in [False, True]:
-                    yield (__test, y, sr_in, sr_out, res_type, fix)
+                    for scale in [False, True]:
+                        yield (__test, y, sr_in, sr_out, res_type, fix, scale)
 
 
 def test_resample_stereo():