数据集正确转换

.gitignore

@@ -1,4 +1,6 @@
 # log and data files
+common_voice_11_0/
+Common Voice Scripted Speech 25.0 - Japanese/
 *.model
 *.pkl
 #*.ipynb

convert_ja.py

@@ -0,0 +1,83 @@
+import os
+import sys
+from dataclasses import is_dataclass
+from datasets import load_dataset
+from datasets.utils.logging import set_verbosity_info, enable_progress_bar
+from omegaconf import OmegaConf
+
+# 开启 datasets 进度条和日志
+set_verbosity_info()
+enable_progress_bar()
+
+# 添加脚本所在路径到 sys.path
+sys.path.append(os.path.join(os.path.dirname(__file__), 'scripts', 'speech_recognition'))
+
+from convert_hf_dataset_to_nemo import HFDatasetConversionConfig, prepare_output_dirs, process_dataset
+
+def convert_ja():
+    # 本地数据集路径
+    dataset_path = os.path.join(os.path.dirname(__file__), "data", "common_voice_11_0")
+    # 输出路径
+    output_dir = os.path.join(os.path.dirname(__file__), "data", "nemo_ja")
+    
+    # 构建配置
+    cfg = HFDatasetConversionConfig(
+        output_dir=output_dir,
+        path=dataset_path,
+        name="ja",
+        ensure_ascii=False,
+        use_auth_token=False,
+        num_proc=4,  # 使用多进程加快速度
+    )
+    
+    # 转换为 OmegaConf
+    if is_dataclass(cfg):
+        cfg = OmegaConf.structured(cfg)
+        
+    # 准备输出目录
+    prepare_output_dirs(cfg)
+    
+    print(f"开始加载数据集 {cfg.path}，语言: {cfg.name}...")
+    print("此过程会进行数据的准备和解压，请耐心等待（可以通过命令行查看进度条）...")
+    
+    # 加载数据集
+    try:
+        dataset = load_dataset(
+            path=cfg.path,
+            name=cfg.name,
+            split=cfg.split,
+            cache_dir=None,
+            streaming=cfg.streaming,
+            token=cfg.use_auth_token,
+            trust_remote_code=True,
+            download_mode="force_redownload",
+        )
+        print("数据集加载完成！")
+    except Exception as e:
+        import traceback
+        print("Failed to load dataset. Traceback:")
+        print(traceback.format_exc())
+        sys.exit(1)
+        
+    print("开始进行格式转换 (HF -> NeMo)...")
+    # 处理数据集
+    if isinstance(dataset, dict):
+        print(f"\nMultiple splits found for dataset {cfg.path}: {list(dataset.keys())}")
+        keys = list(dataset.keys())
+        for key in keys:
+            ds_split = dataset[key]
+            print(f"Processing split {key} for dataset {cfg.path}")
+            cfg.split_output_dir = os.path.join(cfg.resolved_output_dir, key)
+            process_dataset(ds_split, cfg)
+            del dataset[key], ds_split
+            
+        cfg.split_output_dir = None
+    else:
+        print(f"Single split found for dataset {cfg.path} | Split chosen = {cfg.split}")
+        if cfg.split is not None:
+            cfg.split_output_dir = os.path.join(cfg.resolved_output_dir, cfg.split)
+            
+        process_dataset(dataset, cfg)
+
+if __name__ == '__main__':
+    convert_ja()

readme.txt

@@ -25,12 +25,16 @@ https://datacollective.mozillafoundation.org/datasets/cmn2hm68r01n4mm071qux43yu
     日语 asr 数据集，原来的 mozilla-foundation/common_voice_11_0 已失效
 
 https://ai.gitee.com/hf-datasets/mozilla-foundation/common_voice_11_0
+    https://ai.gitee.com/datasets/blbqxxol/common_voice_11_0 我的 fork
     这里可以下载数据集
-    git clone https://ai.gitee.com/zoesimin/common_voice_11_0
 
 https://modelscope.cn/datasets
     这里有方言数据集
 
+conda run -n NeMo pip install datasets hydra-core omegaconf librosa soundfile
+
+
+ git clone https://ai.gitee.com/hf-datasets/common_voice_11_0
 
 tar xzvf "/mnt/y/ai/Common Voice Scripted Speech 25.0 - Japanese.tar.gz" -C /mnt/e/huggingface_echodict/NeMo/data
     wsl 解压数据

requirements.txt

@@ -1 +1 @@
-datasets
+datasets==3.6.0

scripts/speech_recognition/convert_hf_dataset_to_nemo.py

@@ -91,7 +91,7 @@ import hydra
 import librosa
 import soundfile
 import tqdm
-from datasets import Audio, Dataset, IterableDataset, load_dataset
+from datasets import Audio, Dataset, IterableDataset, load_dataset, disable_progress_bar, enable_progress_bar
 from hydra.conf import HydraConf, RunDir
 from hydra.core.config_store import ConfigStore
 from omegaconf import OmegaConf
@@ -139,7 +139,7 @@ def prepare_output_dirs(cfg: HFDatasetConversionConfig):
     cfg.split_output_dir = None
 
 
-def infer_dataset_segments(batch):
+def infer_dataset_segments(audio_path):
     """
     Helper method to run in batch mode over a mapped Dataset.
 
@@ -149,9 +149,9 @@ def infer_dataset_segments(batch):
         A cleaned list of path segments
     """
     segments = []
-    segment, path = os.path.split(batch['audio']['path'])
+    segment, path = os.path.split(audio_path)
     segments.insert(0, path)
-    while segment not in ('', os.path.sep):
+    while segment and segment != os.path.sep and segment != os.path.dirname(segment):
         segment, path = os.path.split(segment)
         segments.insert(0, path)
 
@@ -204,22 +204,55 @@ def build_map_dataset_to_nemo_func(cfg: HFDatasetConversionConfig, basedir):
     """
 
     def map_dataset_to_nemo(batch):
-        # Write audio file to correct path
-        if cfg.streaming:
-            batch['audio_filepath'] = batch['audio']['path'].split("::")[0].replace("zip://", "")
+        # 检查是否是 batched 模式
+        is_batched = isinstance(batch['audio'], list)
+        
+        if is_batched:
+            audio_filepaths = []
+            durations = []
+            
+            for i in range(len(batch['audio'])):
+                audio_path = batch['audio'][i]['path']
+                audio_array = batch['audio'][i]['array']
+                audio_sr = batch['audio'][i]['sampling_rate']
+                
+                # Write audio file to correct path
+                if cfg.streaming:
+                    filepath = audio_path.split("::")[0].replace("zip://", "")
+                else:
+                    segments = infer_dataset_segments(audio_path)
+                    filepath = os.path.join(*segments)
+        
+                filepath = os.path.abspath(os.path.join(basedir, filepath))
+                filepath = prepare_audio_filepath(filepath)
+                audio_filepaths.append(filepath)  # update filepath with prepared path
+        
+                soundfile.write(filepath, audio_array, samplerate=cfg.sampling_rate, format='wav')
+        
+                durations.append(librosa.get_duration(y=audio_array, sr=audio_sr))
+                
+            batch['audio_filepath'] = audio_filepaths
+            batch['duration'] = durations
         else:
-            segments = infer_dataset_segments(batch)
-            audio_filepath = os.path.join(*segments)
-            batch['audio_filepath'] = audio_filepath
-
-        batch['audio_filepath'] = os.path.abspath(os.path.join(basedir, batch['audio_filepath']))
-        audio_filepath = batch['audio_filepath']
-        audio_filepath = prepare_audio_filepath(audio_filepath)
-        batch['audio_filepath'] = audio_filepath  # update filepath with prepared path
-
-        soundfile.write(audio_filepath, batch['audio']['array'], samplerate=cfg.sampling_rate, format='wav')
-
-        batch['duration'] = librosa.get_duration(y=batch['audio']['array'], sr=batch['audio']['sampling_rate'])
+            audio_path = batch['audio']['path']
+            audio_array = batch['audio']['array']
+            audio_sr = batch['audio']['sampling_rate']
+            
+            # Write audio file to correct path
+            if cfg.streaming:
+                filepath = audio_path.split("::")[0].replace("zip://", "")
+            else:
+                segments = infer_dataset_segments(audio_path)
+                filepath = os.path.join(*segments)
+    
+            filepath = os.path.abspath(os.path.join(basedir, filepath))
+            filepath = prepare_audio_filepath(filepath)
+            batch['audio_filepath'] = filepath  # update filepath with prepared path
+    
+            soundfile.write(filepath, audio_array, samplerate=cfg.sampling_rate, format='wav')
+    
+            batch['duration'] = librosa.get_duration(y=audio_array, sr=audio_sr)
+            
         return batch
 
     return map_dataset_to_nemo
@@ -245,9 +278,19 @@ def convert_offline_dataset_to_nemo(
     if num_proc < 0:
         num_proc = max(1, os.cpu_count() // 2)
 
-    dataset = dataset.map(build_map_dataset_to_nemo_func(cfg, basedir), num_proc=num_proc)
+    print(f"[{cfg.split}] 开始提取音频并保存为 wav 格式 (使用 {num_proc} 个进程)...")
+    enable_progress_bar()
+    dataset = dataset.map(
+        build_map_dataset_to_nemo_func(cfg, basedir), 
+        num_proc=num_proc,
+        batched=True, # 强制单条处理
+        desc=f"[{cfg.split}] 转换音频文件"
+    )
+    print(f"[{cfg.split}] 音频转换完成！")
+    
     ds_iter = iter(dataset)
 
+    print(f"[{cfg.split}] 开始生成 manifest JSON 文件...")
     with open(manifest_filepath, 'w') as manifest_f:
         for idx, sample in enumerate(
             tqdm.tqdm(
@@ -322,13 +365,16 @@ def process_dataset(dataset: IterableDataset, cfg: HFDatasetConversionConfig):
     if 'sentence' in dataset.features:
         dataset = dataset.rename_column("sentence", "text")
 
+    # 获取路径的最后一部分作为数据集名称，避免绝对路径导致文件名非法
+    dataset_name = os.path.basename(os.path.normpath(cfg.path))
+
     if cfg.split_output_dir is None:
         basedir = cfg.resolved_output_dir
-        manifest_filename = f"{cfg.path.replace('/', '_')}_manifest.json"
+        manifest_filename = f"{dataset_name.replace('/', '_')}_manifest.json"
     else:
         basedir = cfg.split_output_dir
         split = os.path.split(cfg.split_output_dir)[-1]
-        manifest_filename = f"{split}_{cfg.path.replace('/', '_')}_manifest.json"
+        manifest_filename = f"{split}_{dataset_name.replace('/', '_')}_manifest.json"
 
         if not os.path.exists(cfg.split_output_dir):
             os.makedirs(cfg.split_output_dir, exist_ok=True)