Speech Commands Recognition

Training Deep Learning models using Google Speech Commands Dataset, implemented in PyTorch.

Features

• Training and testing basic ConvNets and TDNNs.
• Standard Train, Test, Valid folders for the Google Speech Commands Dataset v0.02.
• Dataset loader for standard Kaldi speech data folders (files and pipes).

Requirements

• Python 3.6+
• PyTorch
• SoX

To install SoX on Mac with Homebrew:

brew install sox

on Linux:

sudo apt-get install sox

Usage

Google Speech Commands Dataset (v0.02)

To download and extract the Google Speech Commands Dataset run the following command:

./download_audio.sh

Training

Use python3 run.py --help for more parameters and options.

python3 run.py --arc VGG16 --checkpoint VGG16 --num_workers 10

Results (Isolated word recognition, Speech Commands v0.02, 36 words)

Accuracy results for the validation and test sets using the default parameters (VGG16) and with data augmentation (VGG16 + sp)

Model	Valid acc.	Test acc.	parameters and options
VGG16	96.3%	96.4%	default
VGG16 + sp	96.6%	96.7%	--train_path data/train_training_sp

The augmented training dataset train_training_sp is an speed perturbed version of the train_training dataset. It was obtained using the Kaldi script perturb_data_dir_speed_3way.sh