diff --git a/requirements-gpu.txt b/requirements-gpu.txt
index 617e84d..77d02b1 100644
--- a/requirements-gpu.txt
+++ b/requirements-gpu.txt
@@ -1 +1,2 @@
-onnxruntime-gpu
+onnxruntime-gpu<1.18
+numpy<2
diff --git a/requirements-test.txt b/requirements-test.txt
index 2bf7f76..3c8f4e1 100644
--- a/requirements-test.txt
+++ b/requirements-test.txt
@@ -17,4 +17,4 @@ matplotlib
 natsort
 torchaudio
 torch
-transformers
\ No newline at end of file
+transformers
diff --git a/requirements-zoo.txt b/requirements-zoo.txt
index 32e572b..5b95131 100644
--- a/requirements-zoo.txt
+++ b/requirements-zoo.txt
@@ -21,4 +21,4 @@ transformers
 click
 datasets
 evaluate
-pyannote.audio
\ No newline at end of file
+pyannote.audio
diff --git a/soundutils/langid/silero.py b/soundutils/langid/silero.py
index 0551cd3..ec6ceba 100644
--- a/soundutils/langid/silero.py
+++ b/soundutils/langid/silero.py
@@ -25,7 +25,8 @@ def _lang_group_dict_95():
 
 @lru_cache()
 def _open_model():
-    return open_onnx_model(hf_hub_download(repo_id=_REPO_ID, filename='lang_classifier_95.onnx'))
+    onnx_file = hf_hub_download(repo_id=_REPO_ID, filename='lang_classifier_95.onnx')
+    return open_onnx_model(onnx_file, mode='cpu')
 
 
 def _raw_langid(sound: SoundTyping, top_n: int = 5):
diff --git a/soundutils/utils/onnx.py b/soundutils/utils/onnx.py
index 22a37a1..cd5e595 100644
--- a/soundutils/utils/onnx.py
+++ b/soundutils/utils/onnx.py
@@ -74,7 +74,7 @@ def _open_onnx_model(ckpt: str, provider: str, use_cpu: bool = True) -> Inferenc
         providers.append("CPUExecutionProvider")
 
     logging.info(f'Model {ckpt!r} loaded with provider {provider!r}')
-    return InferenceSession(ckpt, options, providers=providers)
+    return InferenceSession(ckpt, sess_options=options, providers=providers)
 
 
 def open_onnx_model(ckpt: str, mode: str = None) -> InferenceSession:
@@ -93,4 +93,4 @@ def open_onnx_model(ckpt: str, mode: str = None) -> InferenceSession:
         on Linux, executing ``export ONNX_MODE=cpu`` will ignore any existing CUDA and force the model inference
         to run on CPU.
     """
-    return _open_onnx_model(ckpt, get_onnx_provider(mode or os.environ.get('ONNX_MODE', None)))
+    return _open_onnx_model(ckpt, provider=get_onnx_provider(mode or os.environ.get('ONNX_MODE', None)))