Add Files

alm_pipeline.py

@@ -1,8 +1,28 @@
+import os
+import warnings
+
 import whisper
 import librosa
 import numpy as np
 import tensorflow_hub as hub
 
+# Reduce TensorFlow log noise and avoid attempting GPU / oneDNN on CPU-only envs
+os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2")  # hide INFO/WARNING logs
+os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")  # disable oneDNN custom ops
+os.environ.setdefault("CUDA_VISIBLE_DEVICES", "-1")  # don't try to use CUDA GPUs
+
+# Suppress specific library warnings that are expected in this setup
+warnings.filterwarnings(
+    "ignore",
+    category=UserWarning,
+    message="FP16 is not supported on CPU; using FP32 instead",
+)
+warnings.filterwarnings(
+    "ignore",
+    category=FutureWarning,
+    module="librosa",
+)
+
 # Load ASR
 asr_model = whisper.load_model("small")
 
@@ -21,7 +41,8 @@ def estimate_emotion(activation):
 
 
 def speech_to_text(audio):
-    result = asr_model.transcribe(audio)
+    # Force FP32 on CPU to avoid FP16 warnings and ensure compatibility
+    result = asr_model.transcribe(audio, fp16=False)
     return result["text"]
 
 
@@ -31,8 +52,15 @@ def detect_sound(audio):
     waveform = waveform.astype(np.float32)
     scores, embeddings, _ = yamnet(waveform)
     mean_scores = np.mean(scores.numpy(), axis=0)
-    top_idx = np.argmax(mean_scores)
-    return class_map[top_idx].decode("utf-8"), mean_scores.max()
+    top_idx = int(np.argmax(mean_scores))
+    # class_map may contain integers or byte strings depending on TF Hub version;
+    # convert robustly to a human-readable label.
+    label = class_map[top_idx]
+    if isinstance(label, bytes):
+        label = label.decode("utf-8")
+    else:
+        label = str(label)
+    return label, float(mean_scores.max())
 
 
 def analyze_audio(audio_file):