buddy23333 / kws_pytorch

Keyword spotting, Speech wake_up, by pytorch, DNN, CNN, TDNN, DFSMN, LSTM

KWS_pytorch

Keyword spotting, Speech wake_up, pytorch, DNN, CNN, TDNN, DFSMN, LSTM

Introduction

• The project is based on ICASSP 2014 paper Small-footprint keyword spotting using deep neural networks.

• We implement the idea with various deep neural network architecture, e.g.,DNN, CNN, TDNN, DFSMN, LSTM.

• The project can be applied to several tasks, such as key-word spotting and speech wake-up.

Documents

• command_loader.py: CommandLoader is defined for data extraction. The data is structured as follow

  • path/key words/audio file (.wav)

• model.py: Implementation of several backbones: DNN, CNN, TDNN, DFSMN, LSTM.

• train.py: Definition of training & testing process.

• run.py: Main program for training & testing. Possible parameters are explained below: +

Datasets

Speech wake-up：

• MobvoiHotwords: A corpus of wake-up words collected from a commercial smart speaker of Mobvoi.
  • Containing audio of "Hi xiaowen" and " Nihao Wenwen", as well as noise speech.
  • Homepage

Key-word spotting：

• Synthetic Speech Commands Dataset.
  • Consisted of key-words audio of thirty categories, e.g., "bed", "bird", "cat", "dog", "eight", "five", "stop", "wow", "zero".
  • Download link

Visualization of Results

Key-word spotting：

• Batchnums-Accuracy curve with STFT：

• Batchnums-Accuracy curve with Deep KWS:

• Accuracy:

Module	Epoch1	Epoch2	epoch3	epoch4	epoch5	text
DNN	38.57%	52.85%	58.81%	67.48%	71.00%	62.59%
CNN	95.30%	96.12%	96.30%	97.20%	96.75%	95.17%
TDNN	70.10%	69.02%	74.35%	77.87%	80.76%	76.50%
LSTM	57.36%	74.35%	75.16%	79.31%	81.39%	78.75%
DFSMN	91.15%	92.14%	94.94%	93.86%	94.04%	90.34%
DNN(KWS)	87.97%	90.37%	91.04%	91.18%	90.44%	89.67%

Speech wake-up：

• Accuracy with Deep_KWS:

• Loss with Deep_KWS: