Update custom_interface_app.py

custom_interface_app.py

@@ -41,7 +41,7 @@ class ASR(Pretrained):
         # Forward encoder + decoder
         tokens = torch.tensor([[1, 1]]) * self.mods.whisper.config.decoder_start_token_id
         tokens = tokens.to(device)
-        enc_out, logits, _ = self.mods.whisper(wavs, tokens)
+        enc_out, logits, _ = self.mods.whisper(wavs.detach(), tokens.detach())
         log_probs = self.hparams.log_softmax(logits)
 
         hyps, _, _, _ = self.hparams.test_search(enc_out.detach(), wav_lens)
@@ -100,6 +100,30 @@ 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):
@@ -150,7 +174,6 @@ class ASR(Pretrained):
 
     def classify_file_whisper_mkd(self, path, device):
         # Load the audio file
-        # path = "long_sample.wav"
         waveform, sr = librosa.load(path, sr=16000)
 
         # Get audio length in seconds
@@ -178,7 +201,8 @@ class ASR(Pretrained):
 
                 # Fake a batch for the segment
                 batch = segment_tensor.unsqueeze(0).to(device)
-                rel_length = torch.tensor([1.0]).to(device)
+                batch = batch.to(torch.float16)
+                rel_length = torch.tensor([1.0], dtype=torch.float16).to(device)
 
                 # Pass the segment through the ASR model
                 segment_output = self.encode_batch_whisper(device, batch, rel_length)
@@ -186,13 +210,14 @@ class ASR(Pretrained):
         else:
             waveform = torch.tensor(waveform).to(device)
             waveform = waveform.to(device)
-            # Fake a batch:
             batch = waveform.unsqueeze(0)
-            rel_length = torch.tensor([1.0]).to(device)
+            batch = batch.to(torch.float16)
+            rel_length = torch.tensor([1.0], dtype=torch.float16).to(device)
             outputs = self.encode_batch_whisper(device, batch, rel_length)
             yield outputs
 
 
+
     def classify_file_whisper(self, path, pipe, device):
         waveform, sr = librosa.load(path, sr=16000)
         transcription = pipe(waveform, generate_kwargs={"language": "macedonian"})["text"]
@@ -228,6 +253,7 @@ class ASR(Pretrained):
 
                 # Pass the segment through the ASR model
                 inputs = processor(segment_tensor, sampling_rate=16_000, return_tensors="pt").to(device)
+                inputs['input_values'] = inputs['input_values'].to(torch.float16)
                 outputs = model(**inputs).logits
                 ids = torch.argmax(outputs, dim=-1)[0]
                 segment_output = processor.decode(ids)
@@ -235,6 +261,7 @@ class ASR(Pretrained):
         else:
             waveform = torch.tensor(waveform).to(device)
             inputs = processor(waveform, sampling_rate=16_000, return_tensors="pt").to(device)
+            inputs['input_values'] = inputs['input_values'].to(torch.float16)
             outputs = model(**inputs).logits
             ids = torch.argmax(outputs, dim=-1)[0]
             transcription = processor.decode(ids)