同时使用多个 tar 数据集生成词表

text_tokenizer.py

@@ -9,14 +9,15 @@ sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "scripts
 # 为了避免冲突，我们直接调用其内部使用的函数，或者使用 os.system 来调用脚本
 # 考虑到动态计算 VOCAB_SIZE，我们首先读取 manifest，收集所有的字符来计算 VOCAB_SIZE
 
-def calculate_vocab_size(manifest_path):
+def calculate_vocab_size(manifest_paths):
     unique_chars = set()
-    with open(manifest_path, 'r', encoding='utf-8') as f:
-        for line in f:
-            item = json.loads(line)
-            text = item.get('text', '')
-            for char in text:
-                unique_chars.add(char)
+    for manifest_path in manifest_paths:
+        with open(manifest_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                item = json.loads(line)
+                text = item.get('text', '')
+                for char in text:
+                    unique_chars.add(char)
     
     # 词汇表大小 = 独立字符数 + 预留特殊 token (如 <s>, </s>, <pad>, <unk> 等) 的数量
     # 通常预留 10-20 个位置给特殊 token 和未见过的字符，这里我们加 20
@@ -25,11 +26,16 @@ def calculate_vocab_size(manifest_path):
     return vocab_size
 
 def main():
-    manifest_path = "data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json"
+    manifest_paths = [
+        "data/common_voice_11_0/ja/train/train_common_voice_11_0_manifest.json",
+        "data/common_voice_11_0/ja/test_tarred_1bk/tarred_audio_manifest.json",
+        "data/common_voice_11_0/ja/validation/validation_common_voice_11_0_manifest.json"
+    ]
+    manifest_paths_str = ",".join(manifest_paths)
     data_root = "data/common_voice_11_0/ja/tokenizers"
     
     print("Calculating dynamic VOCAB_SIZE...")
-    vocab_size = calculate_vocab_size(manifest_path)
+    vocab_size = calculate_vocab_size(manifest_paths)
     print(f"Calculated VOCAB_SIZE: {vocab_size}")
 
     # 调用 process_asr_text_tokenizer
@@ -42,7 +48,7 @@ def main():
     command = [
         "python",
         "scripts/tokenizers/process_asr_text_tokenizer.py",
-        f"--manifest={manifest_path}",
+        f"--manifest={manifest_paths_str}",
         f"--vocab_size={vocab_size}",
         f"--data_root={data_root}",
         "--tokenizer=spe",