Update custom_interface_app.py

custom_interface_app.py

@@ -100,39 +100,12 @@ class ASR(Pretrained):
                     seq.append(token)
             output = []
         return seq
-    
-
-    def increase_volume(self, waveform, threshold_db=-25):
-        # Measure loudness using RMS
-        loudness_vector = librosa.feature.rms(y=waveform)
-        average_loudness = np.mean(loudness_vector)
-        average_loudness_db = librosa.amplitude_to_db(average_loudness)
-
-        print(f"Average Loudness: {average_loudness_db} dB")
-
-        # Check if loudness is below threshold and apply gain if needed
-        if average_loudness_db < threshold_db:
-            # Calculate gain needed
-            gain_db = threshold_db - average_loudness_db
-            gain = librosa.db_to_amplitude(gain_db)  # Convert dB to amplitude factor
-
-            # Apply gain to the audio signal
-            waveform = waveform * gain
-            loudness_vector = librosa.feature.rms(y=waveform)
-            average_loudness = np.mean(loudness_vector)
-            average_loudness_db = librosa.amplitude_to_db(average_loudness)
-
-            print(f"Average Loudness: {average_loudness_db} dB")
-        return waveform
 
 
     def classify_file_w2v2(self, path, device):
         # Load the audio file
         waveform, sr = librosa.load(path, sr=16000)
 
-        # increase the volume if needed
-        # waveform = self.increase_volume(waveform)
-
         # Get audio length in seconds
         audio_length = len(waveform) / sr
         
@@ -142,15 +115,16 @@ class ASR(Pretrained):
             max_duration = 20 * sr  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
-            end = 0
             for i in range(num_segments):
-                start = start + end
-                end = start + max_duration * sr
+                end = start + max_duration
+                if end > len(waveform):
+                    end = len(waveform)
                 segment_part = waveform[start:end]
                 segment_len = len(segment_part) / sr
                 if segment_len < 1:
                     continue
                 segments.append(segment_part)
+                start = end
 
             for segment in segments:
                 segment_tensor = torch.tensor(segment).to(device)
@@ -171,14 +145,14 @@ class ASR(Pretrained):
             outputs = self.encode_batch_w2v2(device, batch, rel_length)
             yield outputs
 
+        
+
 
     def classify_file_whisper_mkd(self, path, device):
         # Load the audio file
+        # path = "long_sample.wav"
         waveform, sr = librosa.load(path, sr=16000)
 
-        # increase the volume if needed
-        # waveform = self.increase_volume(waveform)
-
         # Get audio length in seconds
         audio_length = len(waveform) / sr
         
@@ -188,22 +162,23 @@ class ASR(Pretrained):
             max_duration = 20 * sr  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
-            end = 0
             for i in range(num_segments):
-                start = start + end
-                end = start + max_duration * sr
+                end = start + max_duration
+                if end > len(waveform):
+                    end = len(waveform)
                 segment_part = waveform[start:end]
                 segment_len = len(segment_part) / sr
                 if segment_len < 1:
                     continue
                 segments.append(segment_part)
+                start = end
 
             for segment in segments:
                 segment_tensor = torch.tensor(segment).to(device)
 
                 # Fake a batch for the segment
                 batch = segment_tensor.unsqueeze(0).to(device)
-                rel_length = torch.tensor([1.0]).to(device)  # Adjust if necessary
+                rel_length = torch.tensor([1.0]).to(device)
 
                 # Pass the segment through the ASR model
                 segment_output = self.encode_batch_whisper(device, batch, rel_length)
@@ -228,9 +203,6 @@ class ASR(Pretrained):
         # Load the audio file
         waveform, sr = librosa.load(path, sr=16000)
 
-        # increase the volume if needed
-        # waveform = self.increase_volume(waveform)
-
         # Get audio length in seconds
         audio_length = len(waveform) / sr
         
@@ -240,15 +212,16 @@ class ASR(Pretrained):
             max_duration = 20 * sr  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
-            end = 0
             for i in range(num_segments):
-                start = start + end
-                end = start + max_duration * sr
+                end = start + max_duration
+                if end > len(waveform):
+                    end = len(waveform)
                 segment_part = waveform[start:end]
                 segment_len = len(segment_part) / sr
                 if segment_len < 1:
                     continue
                 segments.append(segment_part)
+                start = end
 
             for segment in segments:
                 segment_tensor = torch.tensor(segment).to(device)