Code Repository for the paper Few-Shot Keyword Spotting with Prototypical Networks.
-
Clone the repository:
git clone https://github.com/ArchitParnami/Few-Shot-KWS -
Create a conda environment:
conda create -n FS-KWS -
If pip not installed, install pip by:
conda install pip -
Install the required packages:
pip install -r requirements.text -
Install the protonets package:
cd Few-Shot-KWS python setup.py develop
cd Few-Shot-KWS/data/
python download_prepare_data.py
To train a simple 2-way 1-shot experiment.
cd Few-Shot-KWS/scripts/train/few-shot/fewshotspeech
./train.sh 2 1 0 mymodel
Specify arguments to train.sh in the following manner
train.sh num_ways num_shots exp_type exp_id
- num_ways
- Number of classes
- Eg. 2 or 4
- num_shots
- Number of samples per class.
- Eg. 1,5
- exp_type
- Number indicating the type of experimental setup
- 0 = Simple N-Way K-Shot Setup. No background, silence or unknown keywords.
- 1 = Include Background
- 2 = Include Silence
- 3 = Include Unknown
- 4 = Background + Silence
- 5 = Background + Unkown
- 6 = Unknown + Silence
- 7 = Background + Silence + Unknown
- Number indicating the type of experimental setup
- exp_id
- identifier = directory name
- results are saved in
Few-Shot-KWS/scripts/train/few-shot/fewshotspeech/results/[exp_id]
cd Few-Shot-KWS/scripts/predict/few-shot
python eval_results.py ../../train/few_shot/fewshotspeech/results/
The evaluation can be found in:
cat Few-Shot-KWS/scripts/train/few-shot/fewshotspeech/results/[exp-id]/[timestamp]/eval.txt
Comaring test accuracy of different embedding networks on 4-way FS-KWS as we increase the number of support examples. The results are presented for four different cases discussed in the paper.
The code in this repository has been adapted from:
- https://github.com/jakesnell/prototypical-networks
- https://github.com/hyperconnect/TC-ResNet
- https://github.com/tensorflow/docs/blob/master/site/en/r1/tutorials/sequences/audio_recognition.md
@misc{parnami2020fewshot,
title={Few-Shot Keyword Spotting With Prototypical Networks},
author={Archit Parnami and Minwoo Lee},
year={2020},
eprint={2007.14463},
archivePrefix={arXiv},
primaryClass={eess.AS}
}