Update app.py

app.py

@@ -9,6 +9,7 @@ from fastapi import FastAPI, UploadFile, File, HTTPException
 from fastapi.middleware.cors import CORSMiddleware
 from typing import Optional, Dict, Any, List
 import warnings
+import noisereduce as nr
 
 warnings.filterwarnings('ignore')
 
@@ -208,6 +209,36 @@ except Exception as e:
     yamnet_model = None
 
 
+# ========================================
+# NOISE REDUCTION AND AUDIO PREPROCESSING
+# ========================================
+def clean_and_prepare_audio(y, sr):
+    """
+    Step 1: Apply noise reduction to clean the audio
+    Step 2: Format audio to proper shape for models
+    """
+    print("   🧹 Applying noise reduction...")
+    try:
+        # Apply noise reduction using spectral gating method
+        y_clean = nr.reduce_noise(y=y, sr=sr, stationary=True, prop_decrease=0.8)
+        print("   ✅ Noise reduction completed")
+    except Exception as e:
+        print(f"   ⚠️ Noise reduction failed: {e}, using original audio")
+        y_clean = y
+    
+    # Normalize audio to proper amplitude range
+    print("   📐 Formatting audio shape for models...")
+    y_clean = y_clean.astype(np.float32)
+    
+    # Normalize to [-1, 1] range
+    max_val = np.max(np.abs(y_clean))
+    if max_val > 0:
+        y_clean = y_clean / max_val
+    
+    print("   ✅ Audio prepared and shaped correctly")
+    return y_clean
+
+
 # ========================================
 # KAGGLE METHOD - FINE-TUNED MODEL
 # ========================================
@@ -587,7 +618,6 @@ def analyze_audio_segments(y, sr, model, segment_duration=10):
 def count_precise_apneas(y, sr, segment_results, min_apnea=1, max_apnea=60):
     """
     ACCURATE APNEA DETECTION
-
     RULES:
     1. Apnea is only valid when a confirmed snoring segment (HIGH or MEDIUM)
        exists IMMEDIATELY before AND immediately after the silence gap.
@@ -601,12 +631,10 @@ def count_precise_apneas(y, sr, segment_results, min_apnea=1, max_apnea=60):
        b) ML score check: gap audio scored by the fine-tuned model must be
           < 0.40 (model must agree it is NOT snoring).
     4. If total snoring count = 0, apnea is forced to 0.
-
     Risk Levels:
     - 10-15 seconds: LOW
     - 15-20 seconds: MEDIUM
     - 20+ seconds: HIGH
-
     Returns: (apnea_count, apnea_events_for_annotations)
     """
     print("\n🎯 APNEA DETECTION (Accurate - Immediate snoring before & after required)")
@@ -845,7 +873,6 @@ def validate_sleep_recording(y, sr):
 def calculate_apnea_risk(apnea_events: List[Dict[str, Any]]) -> str:
     """
     Calculate apnea risk based on the longest apnea duration
-
     Rules:
     - NO: No apnea detected
     - LOW: Apnea duration 10-15 seconds
@@ -880,7 +907,8 @@ async def root():
         "counting_feature": "Enabled - counts individual snores within detected segments" if new_counting_model else "Disabled",
         "apnea_detection": "HYBRID (Fast screening + ML validation)",
         "model_accuracy": "95.33%",
-        "threshold_note": "Fine-tuned threshold: 0.55 (NO LOW confidence detection)"
+        "threshold_note": "Fine-tuned threshold: 0.55 (NO LOW confidence detection)",
+        "noise_reduction": "Enabled - audio cleaned before analysis"
     }
 
 
@@ -911,9 +939,9 @@ async def analyze_audio(file: UploadFile = File(...)):
         print(f"   {'✅ PASS' if is_valid else '⚠️ WARNING'}: {validation_reason}")
         validation_warning = None if is_valid else validation_reason
         
-        # 3. SKIP NOISE REDUCTION (REMOVED)
-        y_clean = y
-        print("⏭️ Noise reduction skipped (removed from pipeline)")
+        # 3. CLEAN AND PREPARE AUDIO (NEW STEP)
+        print("\n🧹 Cleaning and preparing audio...")
+        y_clean = clean_and_prepare_audio(y, sr)
         
         # 4. ANALYZE with STRICT DUAL MODELS + NEW COUNTING
         print("\n🛡️ Using STRICT detection (HIGH and MEDIUM confidence only)...")
@@ -1011,6 +1039,7 @@ async def analyze_audio(file: UploadFile = File(...)):
             "fine_tuned_threshold": 0.55,
             "validation_approach": "Strict - only detecting HIGH and MEDIUM confidence snores",
             "apnea_detection_method": "Hybrid (Fast screening + ML validation)",
+            "noise_reduction_applied": True,
             "events": annotations
         }