diff --git a/speech/.gitignore b/speech/.gitignore
old mode 100644
new mode 100755
diff --git a/speech/.gitmodules b/speech/.gitmodules
old mode 100644
new mode 100755
diff --git a/speech/asset/dingding.png b/speech/asset/dingding.png
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/__init__.py b/speech/cosyvoice/__init__.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/bin/export_jit.py b/speech/cosyvoice/bin/export_jit.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/bin/export_onnx.py b/speech/cosyvoice/bin/export_onnx.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/cli/__init__.py b/speech/cosyvoice/cli/__init__.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/cli/cosyvoice.py b/speech/cosyvoice/cli/cosyvoice.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/cli/frontend.py b/speech/cosyvoice/cli/frontend.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/cli/model.py b/speech/cosyvoice/cli/model.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/dataset/__init__.py b/speech/cosyvoice/dataset/__init__.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/dataset/dataset.py b/speech/cosyvoice/dataset/dataset.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/dataset/processor.py b/speech/cosyvoice/dataset/processor.py
old mode 100644
new mode 100755
index 1eec9762f7f94e445acdcf7dc123e76fa3088394..78e3904c745075bf1bbab9adbc5e886e96c1f08b
--- a/speech/cosyvoice/dataset/processor.py
+++ b/speech/cosyvoice/dataset/processor.py
@@ -21,11 +21,146 @@ import torchaudio
 from torch.nn.utils.rnn import pad_sequence
 import torch.nn.functional as F
 import pyworld as pw
-
+import glob
+import os
+import json
 
 AUDIO_FORMAT_SETS = {'flac', 'mp3', 'm4a', 'ogg', 'opus', 'wav', 'wma'}
 
 
+def individual_file_opener(data, mode='train', tts_data={}):
+    """Load data from individual files instead of parquet
+    
+    Args:
+        data: Iterable[{src}] where src is either:
+            - Path to a directory containing audio files
+            - Path to a JSON index file
+        mode: 'train' or 'test'
+        tts_data: Dict for TTS mode
+        
+    Yields:
+        Dict with all required fields for training
+    """
+    for sample in data:
+        assert 'src' in sample
+        src = sample['src']
+        
+        # Determine if src is a directory or index file
+        if src.endswith('.json'):
+            # Load from index file
+            with open(src, 'r') as f:
+                index_data = json.load(f)
+                file_list = index_data.get('data', [])
+        else:
+            # Scan directory for wav files
+            wav_files = glob.glob(os.path.join(src, '*/*/*wav'))
+            if not wav_files:
+                # Try different patterns
+                wav_files = glob.glob(os.path.join(src, '**/*.wav'), recursive=True)
+            
+            file_list = []
+            for wav_path in wav_files:
+                # Check if all required files exist
+                txt_path = wav_path.replace('.wav', '.normalized.txt')
+                embedding_path = wav_path.replace('.wav', '_embedding.pt')
+                token_path = wav_path.replace('.wav', '_tokens.pt')
+                
+                if not os.path.exists(txt_path):
+                    logging.warning(f'Text file not found for {wav_path}, skipping')
+                    continue
+                
+                # Extract metadata
+                utt = os.path.basename(wav_path).replace('.wav', '')
+                spk = utt.split('_')[0]
+                
+                # Find speaker embedding
+                spk_embed_dir = os.path.join(os.path.dirname(src), 'spk_embeddings')
+                if not os.path.exists(spk_embed_dir):
+                    spk_embed_dir = os.path.join(src, 'spk_embeddings')
+                spk_embedding_path = os.path.join(spk_embed_dir, f'{spk}_embedding.pt')
+                
+                file_info = {
+                    'utt': utt,
+                    'spk': spk,
+                    'wav': wav_path,
+                    'text_path': txt_path,
+                    'embedding_path': embedding_path,
+                    'token_path': token_path,
+                    'spk_embedding_path': spk_embedding_path
+                }
+                file_list.append(file_info)
+        
+        # Process each file
+        for file_info in file_list:
+            try:
+                # Read audio data
+                with open(file_info['wav'], 'rb') as f:
+                    audio_data = f.read()
+                
+                # Read text
+                with open(file_info['text_path'], 'r') as f:
+                    text = ''.join(l.strip() for l in f.readlines())
+                
+                # Load embeddings if they exist
+                if os.path.exists(file_info['embedding_path']):
+                    utt_embedding = torch.load(file_info['embedding_path'])
+                    if isinstance(utt_embedding, torch.Tensor):
+                        utt_embedding = utt_embedding.tolist()
+                else:
+                    logging.warning(f"Utterance embedding not found: {file_info['embedding_path']}")
+                    # Create a dummy embedding
+                    utt_embedding = [0.0] * 192  # Assuming 192-dim embeddings
+                
+                # Load tokens if they exist
+                if os.path.exists(file_info['token_path']):
+                    speech_token = torch.load(file_info['token_path'])
+                    if isinstance(speech_token, torch.Tensor):
+                        speech_token = speech_token.tolist()
+                else:
+                    logging.warning(f"Token file not found: {file_info['token_path']}")
+                    speech_token = []
+                
+                # Load speaker embedding
+                if os.path.exists(file_info['spk_embedding_path']):
+                    spk_embedding = torch.load(file_info['spk_embedding_path'])
+                    if isinstance(spk_embedding, torch.Tensor):
+                        spk_embedding = spk_embedding.tolist()
+                else:
+                    logging.warning(f"Speaker embedding not found: {file_info['spk_embedding_path']}")
+                    # Use utterance embedding as fallback
+                    spk_embedding = utt_embedding
+                
+                # Build sample dict
+                sample_dict = {
+                    'utt': file_info['utt'],
+                    'spk': file_info['spk'],
+                    'audio_data': audio_data,
+                    'text': text,
+                    'text_token': [],  # Will be filled by tokenize processor
+                    'utt_embedding': utt_embedding,
+                    'spk_embedding': spk_embedding,
+                    'speech_token': speech_token,
+                    'wav': file_info['wav'],  # Keep original path for reference
+                }
+                
+                # Copy over any additional fields from the original sample
+                for key, value in sample.items():
+                    if key not in sample_dict:
+                        sample_dict[key] = value
+                
+                if mode == 'train':
+                    yield sample_dict
+                else:
+                    # For TTS mode
+                    if file_info['utt'] in tts_data:
+                        for index, tts_text in enumerate(tts_data[file_info['utt']]):
+                            yield {**sample_dict, 'tts_index': index, 'tts_text': tts_text}
+                    else:
+                        yield sample_dict
+                        
+            except Exception as ex:
+                logging.warning(f'Failed to process {file_info["wav"]}: {ex}')
+
 def parquet_opener(data, mode='train', tts_data={}):
     """ Give url or local file, return file descriptor
         Inplace operation.
diff --git a/speech/cosyvoice/flow/decoder.py b/speech/cosyvoice/flow/decoder.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/flow/flow.py b/speech/cosyvoice/flow/flow.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/flow/flow_matching.py b/speech/cosyvoice/flow/flow_matching.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/flow/length_regulator.py b/speech/cosyvoice/flow/length_regulator.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/hifigan/discriminator.py b/speech/cosyvoice/hifigan/discriminator.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/hifigan/f0_predictor.py b/speech/cosyvoice/hifigan/f0_predictor.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/hifigan/generator.py b/speech/cosyvoice/hifigan/generator.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/hifigan/hifigan.py b/speech/cosyvoice/hifigan/hifigan.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/llm/llm.py b/speech/cosyvoice/llm/llm.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken b/speech/cosyvoice/tokenizer/assets/multilingual_zh_ja_yue_char_del.tiktoken
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/tokenizer/tokenizer.py b/speech/cosyvoice/tokenizer/tokenizer.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/__init__.py b/speech/cosyvoice/transformer/__init__.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/activation.py b/speech/cosyvoice/transformer/activation.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/attention.py b/speech/cosyvoice/transformer/attention.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/convolution.py b/speech/cosyvoice/transformer/convolution.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/decoder.py b/speech/cosyvoice/transformer/decoder.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/decoder_layer.py b/speech/cosyvoice/transformer/decoder_layer.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/embedding.py b/speech/cosyvoice/transformer/embedding.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/encoder.py b/speech/cosyvoice/transformer/encoder.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/encoder_layer.py b/speech/cosyvoice/transformer/encoder_layer.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/label_smoothing_loss.py b/speech/cosyvoice/transformer/label_smoothing_loss.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/positionwise_feed_forward.py b/speech/cosyvoice/transformer/positionwise_feed_forward.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/subsampling.py b/speech/cosyvoice/transformer/subsampling.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/transformer/upsample_encoder.py b/speech/cosyvoice/transformer/upsample_encoder.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/__init__.py b/speech/cosyvoice/utils/__init__.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/class_utils.py b/speech/cosyvoice/utils/class_utils.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/common.py b/speech/cosyvoice/utils/common.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/executor.py b/speech/cosyvoice/utils/executor.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/file_utils.py b/speech/cosyvoice/utils/file_utils.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/frontend_utils.py b/speech/cosyvoice/utils/frontend_utils.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/losses.py b/speech/cosyvoice/utils/losses.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/mask.py b/speech/cosyvoice/utils/mask.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/scheduler.py b/speech/cosyvoice/utils/scheduler.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice/utils/train_utils.py b/speech/cosyvoice/utils/train_utils.py
old mode 100644
new mode 100755
diff --git a/speech/cosyvoice2.yaml b/speech/cosyvoice2.yaml
old mode 100644
new mode 100755
index 551e42edc0258d9de2ee20a680e58ffe101af168..0f0b0c85466073525a506c62352653eedbbf4354
--- a/speech/cosyvoice2.yaml
+++ b/speech/cosyvoice2.yaml
@@ -129,6 +129,9 @@ hifigan: !new:cosyvoice.hifigan.hifigan.HiFiGan
         !ref <mel_spec_transform1>
     ]
 
+individual_file_opener: !name:cosyvoice.dataset.processor.individual_file_opener
+
+
 # processor functions
 parquet_opener: !name:cosyvoice.dataset.processor.parquet_opener
 get_tokenizer: !name:cosyvoice.tokenizer.tokenizer.get_qwen_tokenizer
@@ -176,25 +179,11 @@ padding: !name:cosyvoice.dataset.processor.padding
 
 # dataset processor pipeline
 data_pipeline: [
-    !ref <parquet_opener>,
-    !ref <tokenize>,
-    !ref <filter>,
-    !ref <resample>,
-    !ref <compute_fbank>,
-    !ref <parse_embedding>,
-    !ref <shuffle>,
-    !ref <sort>,
-    !ref <batch>,
-    !ref <padding>,
-]
-data_pipeline_gan: [
-    !ref <parquet_opener>,
+    !ref <individual_file_opener>,
     !ref <tokenize>,
     !ref <filter>,
     !ref <resample>,
-    !ref <truncate>,
     !ref <compute_fbank>,
-    !ref <compute_f0>,
     !ref <parse_embedding>,
     !ref <shuffle>,
     !ref <sort>,
diff --git a/speech/examples/magicdata-read/cosyvoice/local/prepare_data.py b/speech/examples/magicdata-read/cosyvoice/local/prepare_data.py
old mode 100644
new mode 100755
diff --git a/speech/examples/magicdata-read/cosyvoice/tts_text.json b/speech/examples/magicdata-read/cosyvoice/tts_text.json
old mode 100644
new mode 100755
diff --git a/speech/local/prepare_data.py b/speech/local/prepare_data.py
old mode 100644
new mode 100755
diff --git a/speech/test_train.sh b/speech/test_train.sh
new file mode 100755
index 0000000000000000000000000000000000000000..33ad670c31c05d7c5bba68a949e1479c72bf3d02
--- /dev/null
+++ b/speech/test_train.sh
@@ -0,0 +1,110 @@
+#!/bin/bash
+# Copyright 2024 Alibaba Inc. All Rights Reserved.
+
+stage=-1
+stop_stage=3
+
+data_url=www.openslr.org/resources/60
+data_dir=data
+pretrained_model_dir=./pretrained_models/CosyVoice2-0.5B
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+  echo "Data Download"
+  for part in test-clean; do
+    local/download_and_untar.sh ${data_dir} ${data_url} ${part}
+  done
+fi
+
+# if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+#   echo "Data preparation, prepare wav.scp/text/utt2spk/spk2utt"
+#   for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+#     mkdir -p data/$x
+#     python local/prepare_data.py --src_dir $data_dir/LibriTTS/$x --des_dir data/$x
+#   done
+# fi
+
+# if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+#   echo "Extract campplus speaker embedding, you will get spk2embedding.pt and utt2embedding.pt in data/$x dir"
+#   for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+#     tools/extract_embedding.py --dir data/$x \
+#       --onnx_path $pretrained_model_dir/campplus.onnx
+#   done
+# fi
+
+# if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+#   echo "Extract discrete speech token, you will get utt2speech_token.pt in data/$x dir"
+#   for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+#     tools/extract_speech_token.py --dir data/$x \
+#       --onnx_path $pretrained_model_dir/speech_tokenizer_v2.onnx
+#   done
+# fi
+
+# if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+#   echo "Prepare required parquet format data, you should have prepared wav.scp/text/utt2spk/spk2utt/utt2embedding.pt/spk2embedding.pt/utt2speech_token.pt"
+#   for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+#     mkdir -p data/$x/parquet
+#     tools/make_parquet_list.py --num_utts_per_parquet 1000 \
+#       --num_processes 10 \
+#       --src_dir data/$x \
+#       --des_dir data/$x/parquet
+#   done
+# fi
+
+# # train llm
+# export CUDA_VISIBLE_DEVICES="0,1,2,3"
+# num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+# job_id=1986
+# dist_backend="nccl"
+# num_workers=2
+# prefetch=100
+# train_engine=torch_ddp
+# if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+#   echo "Run train. We only support llm traning for now. If your want to train from scratch, please use conf/cosyvoice.fromscratch.yaml"
+#   if [ $train_engine == 'deepspeed' ]; then
+#     echo "Notice deepspeed has its own optimizer config. Modify conf/ds_stage2.json if necessary"
+#   fi
+#   cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
+#   cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
+#   # NOTE will update llm/hift training later
+#   for model in llm flow hifigan; do
+#     torchrun --nnodes=1 --nproc_per_node=$num_gpus \
+#         --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \
+#       cosyvoice/bin/train.py \
+#       --train_engine $train_engine \
+#       --config conf/cosyvoice2.yaml \
+#       --train_data data/train.data.list \
+#       --cv_data data/dev.data.list \
+#       --qwen_pretrain_path $pretrained_model_dir/CosyVoice-BlankEN \
+#       --model $model \
+#       --checkpoint $pretrained_model_dir/$model.pt \
+#       --model_dir `pwd`/exp/cosyvoice2/$model/$train_engine \
+#       --tensorboard_dir `pwd`/tensorboard/cosyvoice2/$model/$train_engine \
+#       --ddp.dist_backend $dist_backend \
+#       --num_workers ${num_workers} \
+#       --prefetch ${prefetch} \
+#       --pin_memory \
+#       --use_amp \
+#       --deepspeed_config ./conf/ds_stage2.json \
+#       --deepspeed.save_states model+optimizer
+#   done
+# fi
+
+# # average model
+# average_num=5
+# if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+#   for model in llm flow hifigan; do
+#     decode_checkpoint=`pwd`/exp/cosyvoice/$model/$train_engine/${model}.pt
+#     echo "do model average and final checkpoint is $decode_checkpoint"
+#     python cosyvoice/bin/average_model.py \
+#       --dst_model $decode_checkpoint \
+#       --src_path `pwd`/exp/cosyvoice/$model/$train_engine  \
+#       --num ${average_num} \
+#       --val_best
+#   done
+# fi
+
+# if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+#   echo "Export your model for inference speedup. Remember copy your llm or flow model to model_dir"
+#   python cosyvoice/bin/export_jit.py --model_dir $pretrained_model_dir
+#   python cosyvoice/bin/export_onnx.py --model_dir $pretrained_model_dir
+# fi
\ No newline at end of file
diff --git a/speech/tools/create_data_list.py b/speech/tools/create_data_list.py
old mode 100644
new mode 100755
diff --git a/speech/tools/inv_file_processor.py b/speech/tools/inv_file_processor.py
old mode 100644
new mode 100755
diff --git a/speech/tools/validate_data.py b/speech/tools/validate_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..d31ee7ee017f0d5f69945103cd2ee7ca72c23ce3
--- /dev/null
+++ b/speech/tools/validate_data.py
@@ -0,0 +1,99 @@
+#!/usr/bin/env python3
+"""Validate that all required files exist for training"""
+
+import argparse
+import glob
+import os
+from tqdm import tqdm
+
+def validate_data(src_dir):
+    """Check that all required files exist
+    
+    Args:
+        src_dir: Directory containing audio files
+    """
+    # Find all wav files
+    wav_files = glob.glob(os.path.join(src_dir, '*/*/*wav'))
+    if not wav_files:
+        wav_files = glob.glob(os.path.join(src_dir, '**/*.wav'), recursive=True)
+    
+    print(f"Found {len(wav_files)} WAV files")
+    
+    missing_txt = []
+    missing_embedding = []
+    missing_token = []
+    missing_spk_embedding = []
+    speakers = set()
+    
+    for wav_path in tqdm(wav_files, desc="Validating files"):
+        # Check text file
+        txt_path = wav_path.replace('.wav', '.normalized.txt')
+        if not os.path.exists(txt_path):
+            missing_txt.append(wav_path)
+        
+        # Check embedding file
+        embedding_path = wav_path.replace('.wav', '_embedding.pt')
+        if not os.path.exists(embedding_path):
+            missing_embedding.append(wav_path)
+        
+        # Check token file
+        token_path = wav_path.replace('.wav', '_tokens.pt')
+        if not os.path.exists(token_path):
+            missing_token.append(wav_path)
+        
+        # Extract speaker
+        utt = os.path.basename(wav_path).replace('.wav', '')
+        spk = utt.split('_')[0]
+        speakers.add(spk)
+    
+    # Check speaker embeddings
+    spk_embed_dir = os.path.join(src_dir, 'spk_embeddings')
+    if os.path.exists(spk_embed_dir):
+        for spk in speakers:
+            spk_embedding_path = os.path.join(spk_embed_dir, f'{spk}_embedding.pt')
+            if not os.path.exists(spk_embedding_path):
+                missing_spk_embedding.append(spk)
+    else:
+        print(f"Speaker embedding directory not found: {spk_embed_dir}")
+        missing_spk_embedding = list(speakers)
+    
+    # Report results
+    print("\n=== Validation Results ===")
+    print(f"Total WAV files: {len(wav_files)}")
+    print(f"Total speakers: {len(speakers)}")
+    print(f"Missing text files: {len(missing_txt)}")
+    print(f"Missing embedding files: {len(missing_embedding)}")
+    print(f"Missing token files: {len(missing_token)}")
+    print(f"Missing speaker embeddings: {len(missing_spk_embedding)}")
+    
+    if missing_txt:
+        print(f"\nFirst 5 missing text files:")
+        for f in missing_txt[:5]:
+            print(f"  {f}")
+    
+    if missing_embedding:
+        print(f"\nFirst 5 missing embedding files:")
+        for f in missing_embedding[:5]:
+            print(f"  {f}")
+    
+    if missing_token:
+        print(f"\nFirst 5 missing token files:")
+        for f in missing_token[:5]:
+            print(f"  {f}")
+    
+    if missing_spk_embedding:
+        print(f"\nFirst 5 missing speaker embeddings:")
+        for spk in list(missing_spk_embedding)[:5]:
+            print(f"  {spk}")
+    
+    # Return success if no missing files
+    return len(missing_txt) == 0 and len(missing_embedding) == 0 and len(missing_token) == 0
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--src_dir', type=str, required=True,
+                       help='Source directory to validate')
+    args = parser.parse_args()
+    
+    success = validate_data(args.src_dir)
+    exit(0 if success else 1)
\ No newline at end of file
diff --git a/speech/train.py b/speech/train.py
old mode 100644
new mode 100755