Fix bug causing TTS to redownload models every time

src/ovtk_audiencekit/core/Audio.py

@@ -36,7 +36,11 @@ class Clip:
 
     @classmethod
     def empty(cls, channels, samplerate, length=0):
-        blank = np.zeros((channels, length), dtype='float32')
+        samples = int(samplerate * length)
+        if channels > 1:
+            blank = np.zeros((channels, samples), dtype='float32')
+        else:
+            blank = np.zeros((samples,), dtype='float32')
         return cls(blank, samplerate)
 
     def copy(self):
@@ -87,7 +91,7 @@ class Clip:
 
     def trim(self, aggressive=False):
         """Remove leading and trailing silence"""
-        self.samples, _ = librosa.effects.trim(self.samples, top_db=10 if aggressive else 20)
+        self.samples, _ = librosa.effects.trim(self.samples, top_db=10 if aggressive else 30)
         return self
 
     def stretch(self, speed):
@@ -121,6 +125,7 @@ class Clip:
             self.samples, fix=False, scale=False,
             orig_sr=self.samplerate * speed, target_sr=self.samplerate,
         )
+        return self
 
     def _opcheck(self, other):
         if not isinstance(other, Clip):

src/ovtk_audiencekit/plugins/TTS/TTS.py

@@ -23,9 +23,12 @@ class TextToSpeechPlugin(PluginBase):
         conf_overrides = {k[2:]: v for k, v in kwargs.items() if k.startswith('o_')}
 
         self.cache_dir = os.path.join(CACHE_DIR, 'tts')
-        self.cache = {}
         os.makedirs(os.path.dirname(self.cache_dir), exist_ok=True)
 
+        self.output_dir = os.path.join(self.cache_dir, 'outputs')
+        os.makedirs(self.output_dir, exist_ok=True)
+        self.cache = {}
+
         self.cuda = cuda
 
         manager = ModelManager(output_prefix=CACHE_DIR) # HACK: coqui automatically adds 'tts' subdir
@@ -60,7 +63,7 @@ class TextToSpeechPlugin(PluginBase):
         for task in self.tasks:
             task.cancel()
         self.output.close()
-        shutil.rmtree(self.cache_dir)
+        shutil.rmtree(self.output_dir)
 
     async def text_to_clip(self, text):
         # Force punctuation (keeps the models from acting unpredictably)
@@ -73,7 +76,7 @@ class TextToSpeechPlugin(PluginBase):
             return Clip.from_file(cached)
         else:
             self.logger.info(f'Generating TTS "{text}"...')
-            filename = os.path.join(self.cache_dir, f'{uuid.uuid1()}.wav')
+            filename = os.path.join(self.output_dir, f'{uuid.uuid1()}.wav')
 
             if self.speaker_wav:
                 fn = lambda _text: self.synthesizer.tts(_text, None, 'en', self.speaker_wav)