Merge branch 'develop' into issue_792_no_default_shape

opensoundscape/spectrogram.py

@@ -733,7 +733,7 @@ def from_audio(
             raise TypeError("Class method expects Audio class as input")
 
         # Generate a linear-frequency spectrogram
-        # with raw stft values rather than decibels
+        # with linear stft values rather than dB values (dB_scale=False)
         linear_spec = Spectrogram.from_audio(
             audio,
             window_type=window_type,
@@ -743,7 +743,7 @@ def from_audio(
             overlap_fraction=overlap_fraction,
             fft_size=fft_size,
             decibel_limits=decibel_limits,
-            dB_scale=dB_scale,
+            dB_scale=False,
             scaling=scaling,
         )
 
@@ -753,7 +753,7 @@ def from_audio(
         filter_bank = librosa.filters.mel(
             sr=audio.sample_rate, n_fft=n_fft, n_mels=n_mels, norm=norm, htk=htk
         )
-        # normalize filter bank: rows should sum to 1 #TODO: is this correct?
+        # normalize filter bank: rows should sum to 1
         fb_constant = np.sum(filter_bank, 1).mean()
         filter_bank = filter_bank / fb_constant
 

tests/test_spectrogram.py

@@ -13,6 +13,11 @@ def veryshort_wav_str():
     return "tests/audio/veryshort.wav"
 
 
+@pytest.fixture()
+def cswa_str():
+    return "tests/audio/aru_1.wav"
+
+
 @pytest.fixture()
 def spec():
     return Spectrogram(
@@ -202,6 +207,16 @@ def test_to_image_with_bandpass():
     )
 
 
+def test_melspectrogram_underflow(cswa_str):
+    """
+    Fixed a bug where log transform was applied twice.
+    Added a test to check the max value of dB scaled spec is as expected
+    """
+    audio = Audio.from_file(cswa_str)
+    mel_spec = MelSpectrogram.from_audio(audio)
+    assert math.isclose(mel_spec.spectrogram.max(), -30.914056301116943, abs_tol=1e-4)
+
+
 def test_to_image_shape(spec):
     img = spec.to_image(shape=[5, 6], channels=2, return_type="torch")
     assert list(img.shape) == [2, 5, 6]  # channels, height, width