cantabile-kwok / VAD-LSTM

Perform Voice Activity Detection with LSTM

VAD Project Task2: VAD Based on BiLSTM and MFCC

cantabile-kwok

Code Environment & Required Packages

Code tested on the following environments:

• Linux 3.10.0
• CUDA 10.1
• Python 3.6.13

Required Python Packages:

• PyTorch 1.4.0
• NumPy 1.19.2
• sklearn 0.24.2
• spafe 0.1.0
• librosa 0.8.0
• matplotlib 3.3.4

Note: spafe may be a niche package. Use pip install spafe to install it (requires SciPy and librosa). See https://spafe.readthedocs.io/en/latest/ for document of this package.

---

Guidelines for Running the Code

Similar to ESPNet recipes, we have explicit stages for running the code.

• stage 0: Preparing data and extracting feature. After this stage, features are stored in feats/
• stage 1: Model training. After this stage, the models are stored as epoch_{}.pth. The RoC curve is stored in fig/
• stage 2: Model evaluating.
• stage 3: Model predicting (generates test_label_task2.txt)

For example, to start from raw wave files, use

python main.py --stage 0

Note that his stage only has to be run once.

Other arguments include:

• --fs: sampling rate
• --win_len, --win_hop frame size and shift in seconds
• --hidden_size, --num_layers: LSTM hyperparameters
• --lr: learning rate
• --num_epoch: number of epochs
• --report_interval: number of batches of each report. E.g. if set to 50, then training loss will be printed to console every 50 batches.
• --L: adjust length in evaluating VACC when calculating SBA, EBA, according to Evaluating VAD for Automatic Speech Recognition, ICSP2014, Sibo Tong et al.

Note: The number of epochs is set to 2 by default. We use the last epoch to evaluate and predict after training. If you want to specify this, modify line 357.

---

Project Structure

vad (root)

├──data

│ ├──dev_label.txt, train_label.txt

│ ├──dev_label_dict.json （存储dev中的音频和对应label的字典）

│ ├──train_label_dict.json （同上）

│ └──test_label_task2.txt (预测结果)

├──feats

│ ├──dev

│ ├──test

│ └──train

├──fig （可视化的结果，包括一些样本的预测值和真值对比、RoC曲线）

├──latex, wavs文件夹

├──__init__.py, evaluate.py, vad_utils.py

├──main.py (主函数)

├──epoch_0.pth, epoch_1.pth等（保存模型）

├──script (可sbatch提交之)

└──README.md，README.html