Upload feature extractor

feature_extraction_gramt_ambisonics.py

@@ -106,6 +106,9 @@ class FeatureExtractor(torch.nn.Module):
         foa_aiv = self._get_foa_intensity_vectors(spec)
         return torch.cat([mel_spec, foa_aiv], dim = 1)
 
+    def put_to_device(self, device):
+        self.spectrogram.to(device)
+
 
 class AmbisonicsFeatureExtractor(SequenceFeatureExtractor):
     r"""
@@ -162,6 +165,8 @@ class AmbisonicsFeatureExtractor(SequenceFeatureExtractor):
         )
 
         waveform = torch.tensor(waveform.clone().detach())
+        melspec.to(waveform.device)
+        melspec.put_to_device(waveform.device)
         # If waveform has two channels, but the channel information is not the first dimension, transpose.
         if (waveform.ndim == 2) and (waveform.shape[0] > 100):
             waveform = waveform.transpose(1, 0)
@@ -184,16 +189,6 @@ class AmbisonicsFeatureExtractor(SequenceFeatureExtractor):
             raise Exception("Unknowm channel count")  
 
 
-    def _normalize_audio(self, audio_data, target_dBFS=-14.0):
-        rms = torch.sqrt(torch.mean(audio_data**2))  # Calculate the RMS of the audio
-        if rms == 0:  # Avoid division by zero in case of a completely silent audio
-            return audio_data
-        current_dBFS = 20 * torch.log10(rms)  # Convert RMS to dBFS
-        gain_dB = target_dBFS - current_dBFS  # Calculate the required gain in dB
-        gain_linear = 10 ** (gain_dB / 20)  # Convert gain from dB to linear scale
-        normalized_audio = audio_data * gain_linear  # Apply the gain to the audio data
-        return normalized_audio
-
     def __call__(
         self,
         raw_speech: Union[np.ndarray, list[float], list[np.ndarray], list[list[float]]],