Upload folder using huggingface_hub

app.py

@@ -46,14 +46,21 @@ STT = EmbeddingGemmaSTT(
 
 
 def _estimate_duration(audio_path, *args, **kwargs):
-    """Reserve GPU time from the input length (extras swallowed: Gradio injects progress=)."""
+    """Reserve GPU time from the input length (extras swallowed: Gradio injects progress=).
+
+    Measured on ZeroGPU (zero-a10g): 5.4 s of audio -> 0.42 s, 6.9 s -> 1.27 s,
+    44.7 s -> 1.18 s and the 180 s cap -> 5.46 s of actual compute, plus roughly
+    0.4 s of per-request worker overhead in steady state. 2 s + 0.05 s per audio
+    second (floored at 6 s so a cold worker still fits) tracks that with a
+    comfortable margin without wasting visitor quota.
+    """
     secs = MAX_SECONDS
     if isinstance(audio_path, str) and os.path.exists(audio_path):
         try:
             secs = sf.info(audio_path).duration
         except Exception:
             pass
-    return int(min(90.0, 15.0 + secs * 0.35))
+    return max(6, min(18, int(round(2.0 + secs * 0.05))))
 
 
 @spaces.GPU(duration=_estimate_duration)
@@ -112,7 +119,7 @@ CSS = """
 .dark .gradio-container { color: var(--body-text-color); }
 """
 
-with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
+with gr.Blocks() as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(
             "# Notesoop-STT-Ko-v1 — Korean Speech Recognition\n"
@@ -177,4 +184,9 @@ with gr.Blocks(theme=gr.themes.Citrus(), css=CSS) as demo:
     )
 
 if __name__ == "__main__":
-    demo.launch(mcp_server=True)
+    demo.launch(
+        theme=gr.themes.Citrus(),
+        css=CSS,
+        mcp_server=True,
+        allowed_paths=[os.path.join(APP_DIR, "examples")],
+    )