Update audio_processing.py

audio_processing.py

@@ -47,7 +47,60 @@ def detect_language(audio):
     return detected_lang
 
 def process_long_audio(audio, task="transcribe", language=None):
-    # ... (rest of the function remains the same)
+    if audio[0] != SAMPLING_RATE:
+        # Save the input audio to a file for ffmpeg processing
+        ta.save("input_audio_1.wav", torch.tensor(audio[1]).unsqueeze(0), audio[0])
+
+        # Resample using ffmpeg
+        try:
+            resample_with_ffmpeg("input_audio_1.wav", "resampled_audio_2.wav", target_sr=SAMPLING_RATE)
+        except subprocess.CalledProcessError as e:
+            print(f"ffmpeg failed: {e.stderr}")
+            raise e
+
+        waveform, _ = ta.load("resampled_audio_2.wav")
+    else:
+        waveform = torch.tensor(audio[1]).float()
+    
+    # Ensure the audio is in the correct shape (mono)
+    if waveform.dim() == 2:
+        waveform = waveform.mean(dim=0)
+    
+    print(f"Waveform shape after processing: {waveform.shape}")
+
+    if waveform.numel() == 0:
+        raise ValueError("Waveform is empty. Please check the input audio file.")
+
+    input_length = waveform.shape[0]  # Since waveform is 1D, access the length with shape[0]
+    chunk_length = int(CHUNK_LENGTH_S * SAMPLING_RATE)
+
+    # Corrected slicing for 1D tensor
+    chunks = [waveform[i:i + chunk_length] for i in range(0, input_length, chunk_length)]
+
+    # Initialize the processor
+    processor = get_processor()
+    model = get_model()
+    device = get_device()
+
+    results = []
+    for chunk in chunks:
+        input_features = processor(chunk, sampling_rate=SAMPLING_RATE, return_tensors="pt").input_features.to(device)
+
+        with torch.no_grad():
+            if task == "translate":
+                forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task="translate")
+                generated_ids = model.generate(input_features, forced_decoder_ids=forced_decoder_ids)
+            else:
+                generated_ids = model.generate(input_features)
+
+        transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        results.extend(transcription)
+
+        # Clear GPU cache
+        torch.cuda.empty_cache()
+
+    return " ".join(results)
+
 
 def process_audio(audio):
     if audio is None: