update

cosyvoice/tokenizer/tokenizer.py

@@ -240,6 +240,8 @@ def get_tokenizer(
 
 class QwenTokenizer():
     def __init__(self, token_path, skip_special_tokens=True):
+        super().__init__()
+        # NOTE: non-chat model, all these special tokens keep randomly initialized.
         special_tokens = {
             'eos_token': '<|endoftext|>',
             'pad_token': '<|endoftext|>',
@@ -248,6 +250,9 @@ class QwenTokenizer():
                 '[breath]', '<strong>', '</strong>', '[noise]',
                 '[laughter]', '[cough]', '[clucking]', '[accent]',
                 '[quick_breath]',
+                "<laughter>", "</laughter>",
+                "[hissing]", "[sigh]", "[vocalized-noise]",
+                "[lipsmack]", "[mn]"
             ]
         }
         self.tokenizer = AutoTokenizer.from_pretrained(token_path)