feat: implement wav2vec2 model loading and inference functions for audio processing

local_demo.py

@@ -166,6 +166,47 @@ def _is_model_compatible(model: Any, n_features: int) -> bool:
     return expected in (None, n_features)
 
 
+def _load_wav2vec2() -> Any:
+    """Load trained wav2vec2 model from .pt checkpoint. Returns None if unavailable."""
+    import torch
+    pt_path = MODELS_DIR / "wav2vec2_auris_v1.pt"
+    if not pt_path.exists():
+        return None
+    try:
+        config = Wav2Vec2Config()
+        model = Wav2Vec2MusicClassifier(config)
+        state = torch.load(str(pt_path), map_location="cpu", weights_only=True)
+        model.load_state_dict(state)
+        model.eval()
+        print(f"wav2vec2 loaded: {pt_path.name}")
+        return model
+    except Exception as exc:  # noqa: BLE001
+        print(f"wav2vec2 skipped ({exc})")
+        return None
+
+
+def _wav2vec2_predict(model: Any, audio_path: str) -> float | None:
+    """Run wav2vec2 inference on a raw audio file. Returns AI probability or None."""
+    import torch
+    try:
+        import librosa
+        config: Wav2Vec2Config = model.config
+        y, _ = librosa.load(audio_path, sr=config.sample_rate, mono=True)
+        max_samples = int(config.max_audio_sec * config.sample_rate)
+        if len(y) > max_samples:
+            y = y[:max_samples]
+        elif len(y) < max_samples:
+            import numpy as _np
+            y = _np.pad(y, (0, max_samples - len(y)))
+        tensor = torch.tensor(y, dtype=torch.float32).unsqueeze(0)  # (1, samples)
+        with torch.no_grad():
+            probs = model.predict_proba(tensor)
+        return float(probs[0])
+    except Exception as exc:  # noqa: BLE001
+        print(f"wav2vec2 inference failed: {exc}")
+        return None
+
+
 def _load_artifacts() -> DemoArtifacts:
     scaler_path = MODELS_DIR / "feature_scaler_v1.pkl"
     columns_path = MODELS_DIR / "feature_columns_v1.json"