fix

app.py

@@ -1,6 +1,18 @@
 import os
-# Force CPU usage only to avoid Hugging Face free tier CUDA initialization errors
-os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
+
+# ── Memory / thread constraints (MUST be set before importing TF) ──────────
+# These prevent TensorFlow from spawning dozens of threads and over-allocating
+# memory, which causes SIGSEGV (exit 139) on memory-limited containers like
+# the Hugging Face free-tier Spaces.
+os.environ["CUDA_VISIBLE_DEVICES"] = "-1"          # Force CPU only
+os.environ["TF_ENABLE_ONEDNN_OPTS"] = "0"          # Disable oneDNN (saves ~200 MB)
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"           # Suppress INFO/WARNING logs
+os.environ["OMP_NUM_THREADS"] = "1"                 # Single OpenMP thread
+os.environ["TF_NUM_INTEROP_THREADS"] = "1"
+os.environ["TF_NUM_INTRAOP_THREADS"] = "1"
+os.environ["OPENBLAS_NUM_THREADS"] = "1"
+os.environ["MKL_NUM_THREADS"] = "1"
+
 import numpy as np
 import librosa
 import scipy.signal
@@ -9,6 +21,10 @@ import gradio as gr
 import tensorflow as tf
 import warnings
 
+# Lock TF threading at runtime as well (belt-and-suspenders).
+tf.config.threading.set_inter_op_parallelism_threads(1)
+tf.config.threading.set_intra_op_parallelism_threads(1)
+
 # Must be called before any model is loaded.
 # keras.config.enable_unsafe_deserialization() unlocks Lambda layer loading
 # across standalone Keras 2.x / 3.x. The tf.keras path is a fallback for
@@ -167,15 +183,28 @@ def load_model():
             return None, str(e)
 
 
-# Load at startup
-model, model_error = load_model()
-if model is None:
-    print(f"Warning: Could not load model. Error: {model_error}")
+# ── Lazy model loading ─────────────────────────────────────────────────────
+# Loading the model at import time doubles peak memory (TF init + model
+# deserialization happen simultaneously). Deferring until the first request
+# keeps startup lightweight and avoids the OOM → SIGSEGV on free-tier HF.
+_model = None
+_model_error = ""
+
+def _get_model():
+    """Return the cached model, loading it on first call."""
+    global _model, _model_error
+    if _model is None and _model_error == "":
+        print("Lazy-loading model on first request…")
+        _model, _model_error = load_model()
+        if _model is None:
+            print(f"Warning: Could not load model. Error: {_model_error}")
+    return _model, _model_error
 
 preprocessor = MachineListenerPreprocessor()
 
 
 def predict(audio_filepath):
+    model, model_error = _get_model()
     if model is None:
         return f"Model not loaded properly. Error: {model_error}"
     if audio_filepath is None:

requirements.txt

@@ -1,5 +1,5 @@
-tensorflow-cpu
+tensorflow-cpu==2.16.*
 gradio
 librosa
-numpy
+numpy<2
 scipy