diff --git a/projects/.gitignore b/projects/.gitignore
new file mode 100644
index 0000000..f10c0ff
--- /dev/null
+++ b/projects/.gitignore
@@ -0,0 +1,52 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+
+# C extensions
+*.so
+
+# Distribution / packaging
+bin/
+build/
+develop-eggs/
+dist/
+eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+.tox/
+.coverage
+.cache
+nosetests.xml
+coverage.xml
+
+# Translations
+*.mo
+
+# Mr Developer
+.mr.developer.cfg
+.project
+.pydevproject
+
+# Rope
+.ropeproject
+
+# Django stuff:
+*.log
+*.pot
+
+# Sphinx documentation
+docs/_build/
+
+.venv
\ No newline at end of file
diff --git a/projects/vad/files/SI499.wav b/projects/vad/files/SI499.wav
new file mode 100644
index 0000000..6538a81
Binary files /dev/null and b/projects/vad/files/SI499.wav differ
diff --git a/projects/vad/files/checkpoint_19 b/projects/vad/files/checkpoint_19
new file mode 100644
index 0000000..81ebb0b
Binary files /dev/null and b/projects/vad/files/checkpoint_19 differ
diff --git a/projects/vad/inference.py b/projects/vad/inference.py
new file mode 100644
index 0000000..9cbd8d6
--- /dev/null
+++ b/projects/vad/inference.py
@@ -0,0 +1,73 @@
+import torch
+import argparse
+import numpy as np
+
+from model.models import DNN
+from utils.audio_processing import signal_to_melspec
+from utils.utils import load_wav, get_config
+from utils.dataloader import process_neighbor
+
+
+def load_checkpoint(model, checkpoint_path):
+    checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
+    model.load_state_dict(checkpoint_dict['state_dict_model'])
+    return model
+
+def get_sample_all_positions(config, mel_file):
+    neighbors = process_neighbor(config["p"], config["f"])
+
+    if isinstance(mel_file, str):
+        mel = np.load(mel_file)
+    else:
+        mel = mel_file
+    
+    mel = np.pad(mel, ((0, 0), (config["p"], config["f"])))
+    C, L = mel.shape
+
+    all_positions = np.arange(config["p"], L-config["f"])
+    mel_neighbors = [[mel[:, position+n] for n in neighbors] for position in all_positions]
+    mel_neighbors = np.asarray(mel_neighbors).reshape(-1, C*len(neighbors))
+    mel = torch.from_numpy(mel_neighbors)
+
+    return mel.to(config["device"])
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-c', '--yaml-dir', type=str, default='./config/base.yaml',
+                        help="YAML file for config")
+    parser.add_argument('-cp', '--check-point', type=str, default='./files/checkpoint_19',
+                        help="Checkpoint file for model")
+    parser.add_argument('-i', '--input-audio', type=str, default='./files/SI499.wav',
+                        help="Audio file for input")
+    args = parser.parse_args()
+    
+    model_config = get_config(args.yaml_dir, "model")
+    preprocess_config = get_config(args.yaml_dir, "preprocess")
+    load_checkpoint_file = args.check_point
+    input_audio_file = args.input_audio
+
+    model = DNN(model_config['in_features'], model_config['hidden_features_list'], model_config['dropout']).to(model_config['device'])
+
+    if load_checkpoint_file is not None:
+        model = load_checkpoint(model, load_checkpoint_file)
+    
+
+    signal_wav = load_wav(input_audio_file, target_sr=preprocess_config['sr_model'])
+    mel = signal_to_melspec(signal_wav,
+                            sr=preprocess_config['sr_model'],
+                            n_fft=preprocess_config['n_fft'],
+                            hop_length=preprocess_config['hop_length'],
+                            win_length=preprocess_config['win_length'],
+                            window=preprocess_config['fn_window'],
+                            n_mel_channels=preprocess_config['n_mel_channels'],
+                            mel_fmin=preprocess_config['mel_fmin'],
+                            mel_fmax=preprocess_config['mel_fmax'])
+
+    signal_wav = signal_wav[:mel.shape[-1]*preprocess_config['hop_length']]
+    mel = mel[:signal_wav.shape[-1]//preprocess_config['hop_length']]
+
+    mels = get_sample_all_positions(model_config, mel)
+    scores, preds = model.infer(mels)
+    TP = TN = FP = FN = 0
+    for pred in zip(preds.reshape(-1)):
+        print(pred)
\ No newline at end of file