Use speech_tokenizer_v3.onnx for CosyVoiceFrontEnd

cosyvoice/cli/frontend.py

@@ -26,26 +26,39 @@ import inflect
 from cosyvoice.utils.file_utils import logging, load_wav
 from cosyvoice.utils.frontend_utils import contains_chinese, replace_blank, replace_corner_mark, remove_bracket, spell_out_number, split_paragraph, is_only_punctuation
 
-
 class CosyVoiceFrontEnd:
-
-    def __init__(self,
-                 get_tokenizer: Callable,
-                 feat_extractor: Callable,
-                 campplus_model: str,
-                 speech_tokenizer_model: str,
-                 spk2info: str = '',
-                 allowed_special: str = 'all'):
+    def __init__(
+        self,
+        get_tokenizer: Callable,
+        feat_extractor: Callable,
+        campplus_model: str,
+        speech_tokenizer_model: str,
+        spk2info: str = '',
+        allowed_special: str = 'all',
+    ):
         self.tokenizer = get_tokenizer()
         self.feat_extractor = feat_extractor
         self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
         option = onnxruntime.SessionOptions()
         option.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
         option.intra_op_num_threads = 1
-        self.campplus_session = onnxruntime.InferenceSession(campplus_model, sess_options=option, providers=["CPUExecutionProvider"])
-        self.speech_tokenizer_session = onnxruntime.InferenceSession(speech_tokenizer_model, sess_options=option,
-                                                                     providers=["CUDAExecutionProvider" if torch.cuda.is_available() else
-                                                                                "CPUExecutionProvider"])
+
+        self.campplus_session = onnxruntime.InferenceSession(
+            campplus_model, sess_options=option, providers=["CPUExecutionProvider"]
+        )
+
+        # Force correct tokenizer file under same directory
+        import os
+        tokenizer_dir = os.path.dirname(speech_tokenizer_model)
+        speech_tokenizer_model = os.path.join(tokenizer_dir, "speech_tokenizer_v3.onnx")
+
+        self.speech_tokenizer_session = onnxruntime.InferenceSession(
+            speech_tokenizer_model,
+            sess_options=option,
+            providers=["CUDAExecutionProvider" if torch.cuda.is_available() else "CPUExecutionProvider"],
+        )
+    
+
         if os.path.exists(spk2info):
             self.spk2info = torch.load(spk2info, map_location=self.device, weights_only=True)
         else: