diff --git a/app.py b/app.py
index 15448281..2f1b2c29 100644
--- a/app.py
+++ b/app.py
@@ -225,10 +225,17 @@
 
 
 class VoxCPMDemo:
-    def __init__(self, model_id: str = "openbmb/VoxCPM2", device: str = "auto") -> None:
+    def __init__(
+        self,
+        model_id: str = "openbmb/VoxCPM2",
+        device: str = "auto",
+        optimize: bool = False,
+    ) -> None:
         self.device = resolve_runtime_device(device, "cuda")
         logger.info(f"Running VoxCPM on device: {self.device}")
-        self.optimize = self.device.startswith("cuda")
+        # Allow enabling torch.compile via param or VOXCPM_OPTIMIZE=1
+        env_optimize = os.environ.get("VOXCPM_OPTIMIZE", "0").lower() in ("1", "true")
+        self.optimize = (optimize or env_optimize) and self.device.startswith("cuda")
 
         self.asr_model_id = "iic/SenseVoiceSmall"
         self.asr_device = "cuda:0" if self.device.startswith("cuda") else "cpu"
@@ -560,8 +567,9 @@ def run_demo(
     show_error: bool = True,
     model_id: str = "openbmb/VoxCPM2",
     device: str = "auto",
+    optimize: bool = False,
 ):
-    demo = VoxCPMDemo(model_id=model_id, device=device)
+    demo = VoxCPMDemo(model_id=model_id, device=device, optimize=optimize)
     interface = create_demo_interface(demo)
     interface.queue(max_size=10, default_concurrency_limit=1).launch(
         server_name=server_name,
@@ -597,10 +605,17 @@ def run_demo(
         default="auto",
         help="Runtime device: auto, cpu, mps, cuda, or cuda:N (default: auto)",
     )
+    parser.add_argument(
+        "--optimize",
+        action="store_true",
+        default=False,
+        help="Enable torch.compile optimization (requires compatible GPU/PyTorch)",
+    )
     args = parser.parse_args()
     run_demo(
         model_id=args.model_id,
         server_name=args.host,
         server_port=args.port,
         device=args.device,
+        optimize=args.optimize,
     )
diff --git a/src/voxcpm/model/utils.py b/src/voxcpm/model/utils.py
index f6d7463c..37ae1e96 100644
--- a/src/voxcpm/model/utils.py
+++ b/src/voxcpm/model/utils.py
@@ -240,3 +240,43 @@ def resolve_runtime_device(device: Optional[str], configured_device: str = "cuda
         f"Unsupported device '{device}'. Supported values are 'auto', 'cpu', 'mps', "
         "'cuda', or indexed CUDA devices like 'cuda:0'."
     )
+
+
+def load_audio_file(file_path: str, target_sr: int = 16000) -> torch.Tensor:
+    """Load any audio file (WAV, MP3, M4A, AAC, FLAC, OGG, WEBM, etc.) and return a mono 1D float32 tensor."""
+    # Try librosa / soundfile first
+    try:
+        import librosa
+        audio, _ = librosa.load(file_path, sr=target_sr, mono=True)
+        return torch.from_numpy(audio.astype(torch.float32.numpy_dtype() if hasattr(torch.float32, 'numpy_dtype') else "float32"))
+    except Exception:
+        pass
+
+    # Fallback to ffmpeg stream conversion
+    try:
+        import io
+        import subprocess
+        import soundfile as sf
+
+        try:
+            import imageio_ffmpeg
+            ffmpeg_exe = imageio_ffmpeg.get_ffmpeg_exe()
+        except ImportError:
+            ffmpeg_exe = "ffmpeg"
+
+        cmd = [
+            ffmpeg_exe,
+            "-nostdin",
+            "-threads", "0",
+            "-i", str(file_path),
+            "-f", "wav",
+            "-ar", str(target_sr),
+            "-ac", "1",
+            "pipe:1",
+        ]
+        proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, check=True)
+        audio_np, _ = sf.read(io.BytesIO(proc.stdout), dtype="float32")
+        return torch.from_numpy(audio_np)
+    except Exception as e:
+        raise RuntimeError(f"Failed to load audio file '{file_path}': {e}") from e
+
diff --git a/src/voxcpm/model/voxcpm2.py b/src/voxcpm/model/voxcpm2.py
index 174dea3b..49f65696 100644
--- a/src/voxcpm/model/voxcpm2.py
+++ b/src/voxcpm/model/voxcpm2.py
@@ -48,6 +48,7 @@
 from .utils import (
     apply_generation_seed,
     get_dtype,
+    load_audio_file,
     materialize_generation_seed,
     mask_multichar_chinese_tokens,
     next_and_close,
@@ -414,8 +415,7 @@ def _encode_wav(
         Returns:
             audio_feat: (T, P, D) tensor of latent patches.
         """
-        audio, _ = librosa.load(wav_path, sr=self._encode_sample_rate, mono=True)
-        audio = torch.from_numpy(audio).unsqueeze(0)
+        audio = load_audio_file(wav_path, target_sr=self._encode_sample_rate).unsqueeze(0)
         if trim_silence_vad:
             audio = _trim_audio_silence_vad(audio, self._encode_sample_rate, max_silence_ms=200.0)
         patch_len = self.patch_size * self.chunk_size