Update custom_interface_app.py

custom_interface_app.py

@@ -133,10 +133,11 @@ class ASR(Pretrained):
         # Get audio length in seconds
         audio_length = len(waveform) / 16000
         
-        if audio_length >= 20:
+        if audio_length >= 30:
             # split audio every 20 seconds
             segments = []
-            max_duration = 20 * 16000  # Maximum segment duration in samples (20 seconds)
+            all_segments = []
+            max_duration = 30 * 16000  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
             for i in range(num_segments):
@@ -159,7 +160,14 @@ class ASR(Pretrained):
 
                 # Pass the segment through the ASR model
                 segment_output = self.encode_batch_w2v2(device, batch, rel_length)
-                yield segment_output
+                segment_output = [" ".join(segment) for segment in segment_output]
+                all_segments.append(segment_output)
+            
+            segments = ""
+            for segment in all_segments:
+                segment = segment[0]
+                segments += segment + " "
+            return [segments]
         else:
             waveform = torch.tensor(waveform).to(device)
             waveform = waveform.to(device)
@@ -167,7 +175,7 @@ class ASR(Pretrained):
             batch = waveform.unsqueeze(0)
             rel_length = torch.tensor([1.0]).to(device)
             outputs = self.encode_batch_w2v2(device, batch, rel_length)
-            yield outputs
+            return [" ".join(out) for out in outputs]
 
         
 
@@ -179,10 +187,11 @@ class ASR(Pretrained):
         # Get audio length in seconds
         audio_length = len(waveform) / 16000
         
-        if audio_length >= 20:
+        if audio_length >= 30:
             # split audio every 20 seconds
             segments = []
-            max_duration = 20 * 16000  # Maximum segment duration in samples (20 seconds)
+            all_segments = []
+            max_duration = 30 * 16000  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
             for i in range(num_segments):
@@ -205,21 +214,28 @@ class ASR(Pretrained):
 
                 # Pass the segment through the ASR model
                 segment_output = self.encode_batch_whisper(device, batch, rel_length)
-                yield segment_output
+                # segment_output = [" ".join(segment) for segment in segment_output]
+                all_segments.append(segment_output)
+            
+            segments = ""
+            for segment in all_segments:
+                segment = segment[0]
+                segments += segment + " "
+            return [segments]
         else:
             waveform = torch.tensor(waveform).to(device)
             waveform = waveform.to(device)
             batch = waveform.unsqueeze(0)
             rel_length = torch.tensor([1.0]).to(device)
             outputs = self.encode_batch_whisper(device, batch, rel_length)
-            yield outputs
+            return outputs
 
 
 
     def classify_file_whisper(self, waveform, pipe, device):
         # waveform, sr = librosa.load(path, sr=16000)
         transcription = pipe(waveform, generate_kwargs={"language": "macedonian"})["text"]
-        return transcription
+        return [transcription]
        
 
     def classify_file_mms(self, waveform, processor, model, device):
@@ -229,10 +245,11 @@ class ASR(Pretrained):
         # Get audio length in seconds
         audio_length = len(waveform) / 16000
         
-        if audio_length >= 20:
+        if audio_length >= 30:
             # split audio every 20 seconds
             segments = []
-            max_duration = 20 * 16000  # Maximum segment duration in samples (20 seconds)
+            all_segments = []
+            max_duration = 30 * 16000  # Maximum segment duration in samples (20 seconds)
             num_segments = int(np.ceil(len(waveform) / max_duration))
             start = 0
             for i in range(num_segments):
@@ -255,7 +272,13 @@ class ASR(Pretrained):
                 outputs = model(**inputs).logits
                 ids = torch.argmax(outputs, dim=-1)[0]
                 segment_output = processor.decode(ids)
-                yield segment_output
+                # segment_output = [" ".join(segment) for segment in segment_output]
+                all_segments.append(segment_output)
+            
+            segments = ""
+            for segment in all_segments:
+                segments += segment + " "
+            return [segments]
         else:
             waveform = torch.tensor(waveform).to(device)
             inputs = processor(waveform, sampling_rate=16_000, return_tensors="pt").to(device)
@@ -263,4 +286,4 @@ class ASR(Pretrained):
             outputs = model(**inputs).logits
             ids = torch.argmax(outputs, dim=-1)[0]
             transcription = processor.decode(ids)
-            yield transcription
+            return [transcription]