Update app.py

app.py

@@ -7,23 +7,53 @@ from datetime import datetime
 from transformers import pipeline
 import soundfile as sf
 import torch
+from tenacity import retry, stop_after_attempt, wait_fixed
+
+# Initialize local models with retry logic
+@retry(stop=stop_after_attempt(3), wait=wait_fixed(2))
+def load_whisper_model():
+    try:
+        # Whisper for speech-to-text (English-only)
+        model = pipeline(
+            "automatic-speech-recognition",
+            model="openai/whisper-tiny.en",
+            device=-1,  # CPU; use device=0 for GPU if available
+            model_kwargs={"use_safetensors": True}
+        )
+        print("Whisper model loaded successfully.")
+        return model
+    except Exception as e:
+        print(f"Failed to load Whisper model: {str(e)}")
+        raise
+
+@retry(stop=stop_after_attempt(3), wait=wait_fixed(2))
+def load_symptom_model():
+    try:
+        # Symptom-2-Disease for health analysis
+        model = pipeline(
+            "text-classification",
+            model="abhirajeshbhai/symptom-2-disease-net",
+            device=-1,  # CPU
+            model_kwargs={"use_safetensors": True}
+        )
+        print("Symptom-2-Disease model loaded successfully.")
+        return model
+    except Exception as e:
+        print(f"Failed to load Symptom-2-Disease model: {str(e)}")
+        raise
+
+whisper = None
+symptom_classifier = None
 
-# Initialize local models
 try:
-    # Whisper for speech-to-text (English-only)
-    whisper = pipeline("automatic-speech-recognition", model="openai/whisper-tiny.en", device=-1)  # CPU; use device=0 for GPU
-    print("Whisper model loaded successfully.")
+    whisper = load_whisper_model()
 except Exception as e:
-    print(f"Failed to load Whisper model: {str(e)}")
-    whisper = None
+    print(f"Whisper model initialization failed after retries: {str(e)}")
 
 try:
-    # Symptom-2-Disease for health analysis
-    symptom_classifier = pipeline("text-classification", model="abhirajeshbhai/symptom-2-disease-net", device=-1)  # CPU
-    print("Symptom-2-Disease model loaded successfully.")
+    symptom_classifier = load_symptom_model()
 except Exception as e:
-    print(f"Failed to load Symptom-2-Disease model: {str(e)}")
-    symptom_classifier = None
+    print(f"Symptom-2-Disease model initialization failed after retries: {str(e)}")
 
 def compute_file_hash(file_path):
     """Compute MD5 hash of a file to check uniqueness."""
@@ -36,16 +66,22 @@ def compute_file_hash(file_path):
 def transcribe_audio(audio_file):
     """Transcribe audio using local Whisper model."""
     if not whisper:
-        return "Error: Whisper model not loaded. Check logs for details."
+        return "Error: Whisper model not loaded. Check logs for details or ensure sufficient compute resources."
     try:
-        # Load and resample audio to 16,000 Hz
+        # Load and validate audio
         audio, sr = librosa.load(audio_file, sr=16000)
-        # Save as WAV for Whisper compatibility
+        if len(audio) < 1600:  # Less than 0.1s
+            return "Error: Audio too short. Please provide audio of at least 1 second."
+        if np.max(np.abs(audio)) < 1e-4:  # Too quiet
+            return "Error: Audio too quiet. Please provide clear audio describing symptoms in English."
+        
+        # Save as WAV for Whisper
         temp_wav = f"/tmp/{os.path.basename(audio_file)}.wav"
         sf.write(temp_wav, audio, sr)
         
-        # Transcribe
-        result = whisper(temp_wav)
+        # Transcribe with beam search for accuracy
+        with torch.no_grad():
+            result = whisper(temp_wav, generate_kwargs={"num_beams": 5})
         transcription = result.get("text", "").strip()
         print(f"Transcription: {transcription}")
         
@@ -57,6 +93,10 @@ def transcribe_audio(audio_file):
         
         if not transcription:
             return "Transcription empty. Please provide clear audio describing symptoms in English."
+        # Check for repetitive transcription
+        words = transcription.split()
+        if len(words) > 5 and len(set(words)) < len(words) / 2:
+            return "Error: Transcription appears repetitive. Please provide clear, non-repetitive audio describing symptoms."
         return transcription
     except Exception as e:
         return f"Error transcribing audio: {str(e)}"
@@ -64,11 +104,12 @@ def transcribe_audio(audio_file):
 def analyze_symptoms(text):
     """Analyze symptoms using local Symptom-2-Disease model."""
     if not symptom_classifier:
-        return "Error: Symptom-2-Disease model not loaded. Check logs for details.", 0.0
+        return "Error: Symptom-2-Disease model not loaded. Check logs for details or ensure sufficient compute resources.", 0.0
     try:
         if not text or "Error transcribing" in text:
             return "No valid transcription for analysis.", 0.0
-        result = symptom_classifier(text)
+        with torch.no_grad():
+            result = symptom_classifier(text)
         if result and isinstance(result, list) and len(result) > 0:
             prediction = result[0]["label"]
             score = result[0]["score"]
@@ -141,7 +182,7 @@ iface = gr.Interface(
     inputs=gr.Audio(type="filepath", label="Record or Upload Voice"),
     outputs=gr.Textbox(label="Health Assessment Feedback"),
     title="Health Voice Analyzer",
-    description="Record or upload a voice sample describing symptoms for preliminary health assessment. Supports English (transcription), with symptom analysis in English."
+    description="Record or upload a voice sample describing symptoms for preliminary health assessment. Supports English (transcription), with symptom analysis in English. Use clear audio (WAV, 16kHz) describing symptoms like 'I have a cough.'"
 )
 
 if __name__ == "__main__":