Efficient Keyword Spotting by capturing long-range interactions with Temporal Lambda Networks

Abstract

Models based on attention mechanisms have shown unprecedented speech recognition performance. However, they are computationally expensive and unnecessarily complex for keyword spotting, a task targeted to small-footprint devices.

This work explores the application of Lambda networks, an alternative framework for capturing long-range interactions without attention, for the keyword spotting task. We propose a novel ResNet-based model by swapping the residual blocks by temporal Lambda layers. Furthermore, the proposed architecture is built upon uni-dimensional temporal convolutions that further reduce its complexity.

The presented model does not only reach state-of-the-art accuracies on the Google Speech Commands dataset, but it is 85% and 65% lighter than its Transformer-based (KWT) and convolutional (Res15) counterparts while being up to 100x faster. To the best of our knowledge, this is the first attempt to explore the Lambda framework within the speech domain and therefore, we unravel further research of new interfaces based on this architecture.

Requirements

Here we list the requirements needed to run the project (and the version we used). It is recomendded to install pytorch and torchaudio following the official installation instructions

easydict (1.9)
einops (0.3.0)
librosa (0.8.0)
pyyaml (5.4.1)

Configurating the model

The model configurations of the 3 tested models for the Google Speech Commands dataset are found in configs/google_commands/.

LambdaResnet18
TC Resnet 14
Resnet 15

To change the subtask, change the num_lables variable in the config .yml file. More parameters can be fine tuned just changing the configuration variables.

Training and evaluating the model

The training of the model can be performed by running the train.py file through the following command: (Choose the available GPU on your PC)

python -u train.py --config_exp configs/google_commands/'desired_config'.yml --gpu X

If gpu is not set, the model will be trained in DataParallel mode, using all the available GPUs and multipliying its batch size for the number of available GPUs.

To evaluate the trained model, the path to the pth.tar saved model must be given:

python -u eval.py --config_exp configs/google_commands/'desired_config'.yml --gpu X --model output/google_commands/'desired_model'/'model'.pth.tar

Plotting the results