Update main.py

main.py

@@ -8,11 +8,13 @@ from fastapi.responses import JSONResponse
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
-from librosa.sequence import dtw
 import tempfile
 import uuid
 import shutil
 
+# Disable numba JIT to avoid caching issues
+os.environ["NUMBA_DISABLE_JIT"] = "1"
+
 # Initialize FastAPI app
 app = FastAPI(
     title="Quran Recitation Comparison API",
@@ -114,11 +116,60 @@ def get_deep_embedding(audio, sr=16000):
     except Exception as e:
         raise HTTPException(status_code=500, detail=f"Error extracting embeddings: {e}")
 
+# Custom DTW implementation to avoid librosa.sequence.dtw issues
+def custom_dtw(X, Y, metric='euclidean'):
+    """
+    Custom implementation of DTW to avoid librosa.sequence.dtw issues.
+    
+    Parameters:
+    X, Y : numpy.ndarray
+        The two sequences to be aligned
+    metric : str, optional
+        The distance metric to use
+        
+    Returns:
+    D : numpy.ndarray
+        The accumulated cost matrix
+    wp : list
+        The warping path
+    """
+    # Initialize cost matrix
+    n, m = len(X[0]), len(Y[0])
+    D = np.zeros((n+1, m+1))
+    D[0, :] = np.inf
+    D[:, 0] = np.inf
+    D[0, 0] = 0
+    
+    # Fill cost matrix
+    for i in range(1, n+1):
+        for j in range(1, m+1):
+            if metric == 'euclidean':
+                cost = np.sqrt(np.sum((X[:, i-1] - Y[:, j-1])**2))
+            elif metric == 'cosine':
+                cost = 1 - np.dot(X[:, i-1], Y[:, j-1]) / (np.linalg.norm(X[:, i-1]) * np.linalg.norm(Y[:, j-1]))
+            else:
+                cost = np.sum(np.abs(X[:, i-1] - Y[:, j-1]))  # Manhattan by default
+                
+            D[i, j] = cost + min(D[i-1, j], D[i, j-1], D[i-1, j-1])
+    
+    # Backtrack to find warping path
+    i, j = n, m
+    wp = [(i, j)]
+    while i > 1 or j > 1:
+        candidates = [(i-1, j-1), (i-1, j), (i, j-1)]
+        valid_candidates = [(ii, jj) for ii, jj in candidates if ii > 0 and jj > 0]
+        i, j = min(valid_candidates, key=lambda x: D[x[0], x[1]])
+        wp.append((i, j))
+    
+    wp.reverse()
+    return D, wp
+
 # Compute DTW distance
 def compute_dtw_distance(features1, features2):
     """Compute the DTW distance between two sequences of features."""
     try:
-        D, wp = dtw(X=features1, Y=features2, metric='euclidean')
+        # Use custom DTW implementation instead of librosa's
+        D, wp = custom_dtw(features1, features2, metric='euclidean')
         distance = D[-1, -1]
         normalized_distance = distance / len(wp)
         return normalized_distance