Update app.py

app.py

@@ -2,6 +2,9 @@
 # Active models: AST x3 and PaSST (AudioSet-527); optional CLAP and Silero VAD.
 # This intentionally does not use the separate DCASE 2024 CRNN app/checkpoint.
 
+# Required by Hugging Face ZeroGPU. Keep this before importing torch.
+import spaces
+
 import os
 import sys
 import glob
@@ -18,6 +21,7 @@ import matplotlib.pyplot as plt
 import gradio as gr
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"ZeroGPU-aware inference device: {DEVICE}")
 MAX_SEC = int(os.getenv("MAX_SEC", "60"))
 MAX_CELLS = int(os.getenv("MAX_CELLS", "1500"))
 
@@ -97,9 +101,9 @@ try:
     def clap_logits(batch48, prompts):
         kw = dict(text=prompts, return_tensors="pt", padding=True)
         try:
-            inp = clap_proc(audio=list(batch48), **kw)                       # recent transformers
+            inp = clap_proc(audio=list(batch48), sampling_rate=48000, **kw)  # recent transformers
         except TypeError:
-            inp = clap_proc(audios=list(batch48), **kw)                      # older transformers
+            inp = clap_proc(audios=list(batch48), sampling_rate=48000, **kw) # compatibility fallback
         inp = {k: v.to(DEVICE) for k, v in inp.items()}
         return clap_model(**inp).logits_per_audio.float().cpu().numpy()      # scaled audio-text similarity (B, K)
 
@@ -384,6 +388,7 @@ def tag_windows(wavs, win_s, hop_s, dur):
     return {f: np.mean(v, 0) for f, v in fam.items()}
 
 
+@spaces.GPU(duration=120)
 def detect(audio_path, sens=0.0, win_s=2.0, hop_s=0.25, min_votes=2, use_vad=True, denoise=False,
            discover=True, extra=""):
     if audio_path is None: