Update app.py

app.py

@@ -55,28 +55,38 @@ def predict_voice_score(wav):
 
 
 def detect_singing(audio_path):
-    """检测唱歌片段"""
     wav, _ = librosa.load(audio_path, sr=SAMPLE_RATE)
-    duration = len(wav) / SAMPLE_RATE
-    raw_segments = []
+    total_len = len(wav)
+    chunk_len = SAMPLE_RATE * CHUNK_SIZE
+    step_len = SAMPLE_RATE * STEP
+    results = []
 
-    for start in np.arange(0, max(0, duration - CHUNK_SIZE), STEP):
-        end = start + CHUNK_SIZE
-        snippet = wav[int(start * SAMPLE_RATE):int(end * SAMPLE_RATE)]
+    for start_idx in range(0, max(1, total_len - chunk_len), step_len):
+        end_idx = start_idx + chunk_len
+        snippet = wav[start_idx:end_idx]
 
+        # --- 修正：强制长度一致 ---
+        if len(snippet) < chunk_len:
+            snippet = np.pad(snippet, (0, chunk_len - len(snippet)))
+
+        # --- 模型推理 ---
         music_score = predict_music_score(snippet)
         voice_score = predict_voice_score(snippet)
 
+        start_t = start_idx / SAMPLE_RATE
+        end_t = start_t + CHUNK_SIZE
+
         if music_score > MUSIC_THRESHOLD and voice_score > VOICE_THRESHOLD:
-            raw_segments.append((float(start), float(end)))
+            results.append((float(start_t), float(end_t)))
 
-    # === 合并连续窗口 ===
+    # 合并相邻片段
     merged = []
-    for seg in raw_segments:
+    for seg in results:
         if not merged or seg[0] > merged[-1][1]:
             merged.append(list(seg))
         else:
             merged[-1][1] = seg[1]
+
     merged = [(s, e) for s, e in merged if e - s >= MIN_SEG_DURATION]
     return merged