Fixing FFT issue with odd signals. Making state default to None in in…

…stantiate. (#50)

* Fixing FFT issue with odd signals. Making state default to None in instantiate.

* If loudness_cutoff is None, use faster excerpt routine.

Co-authored-by: pseeth <[email protected]>

audiotools/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.3.10"
+__version__ = "0.3.11"
 from .core import AudioSignal, STFTParams, Meter, util
 from . import metrics
 from . import data

audiotools/core/audio_signal.py

@@ -111,16 +111,18 @@ def excerpt(cls, audio_path, offset=None, duration=None, state=None, **kwargs):
     def salient_excerpt(
         cls, audio_path, loudness_cutoff=None, num_tries=None, state=None, **kwargs
     ):
-        loudness_cutoff = -np.inf if loudness_cutoff is None else loudness_cutoff
         state = util.random_state(state)
-        loudness = -np.inf
-        num_try = 0
-        while loudness <= loudness_cutoff:
+        if loudness_cutoff is None:
             excerpt = cls.excerpt(audio_path, state=state, **kwargs)
-            loudness = excerpt.loudness()
-            num_try += 1
-            if num_tries is not None and num_try >= num_tries:
-                break
+        else:
+            loudness = -np.inf
+            num_try = 0
+            while loudness <= loudness_cutoff:
+                excerpt = cls.excerpt(audio_path, state=state, **kwargs)
+                loudness = excerpt.loudness()
+                num_try += 1
+                if num_tries is not None and num_try >= num_tries:
+                    break
         return excerpt
 
     @classmethod

audiotools/core/effects.py

@@ -78,15 +78,16 @@ def convolve(self, other, start_at_max=True):
         delta = torch.zeros_like(other.audio_data)
         delta[..., 0] = 1
 
-        delta_fft = torch.fft.rfft(delta)
-        other_fft = torch.fft.rfft(other.audio_data)
-        self_fft = torch.fft.rfft(self.audio_data)
+        length = self.signal_length
+        delta_fft = torch.fft.rfft(delta, length)
+        other_fft = torch.fft.rfft(other.audio_data, length)
+        self_fft = torch.fft.rfft(self.audio_data, length)
 
         convolved_fft = other_fft * self_fft
-        convolved_audio = torch.fft.irfft(convolved_fft)
+        convolved_audio = torch.fft.irfft(convolved_fft, length)
 
         delta_convolved_fft = other_fft * delta_fft
-        delta_audio = torch.fft.irfft(delta_convolved_fft)
+        delta_audio = torch.fft.irfft(delta_convolved_fft, length)
 
         # Use the delta to rescale the audio exactly as needed.
         delta_max = delta_audio.abs().max(dim=-1, keepdims=True)[0]

audiotools/data/transforms.py

@@ -71,7 +71,7 @@ def __call__(self, *args, **kwargs):
 
     def instantiate(
         self,
-        state: RandomState,
+        state: RandomState = None,
         signal: AudioSignal = None,
     ):
         state = util.random_state(state)
@@ -105,7 +105,7 @@ def instantiate(
 
     def batch_instantiate(
         self,
-        states: list,
+        states: list = None,
         signal: AudioSignal = None,
     ):
         kwargs = []

setup.py

@@ -6,7 +6,7 @@
 
 setup(
     name="audiotools",
-    version="0.3.10",
+    version="0.3.11",
     classifiers=[
         "Intended Audience :: Developers",
         "Intended Audience :: Education",

tests/core/test_audio_signal.py

@@ -128,6 +128,11 @@ def test_salient_excerpt(loudness_cutoff):
         signal = AudioSignal.salient_excerpt(
             f.name, loudness_cutoff=np.inf, duration=1, num_tries=10
         )
+        signal = AudioSignal.salient_excerpt(
+            f.name,
+            loudness_cutoff=None,
+            duration=1,
+        )
 
 
 def test_arithmetic():