Description

Implementation of the work presented in "CNN based Query by Example Spoken Term Detection".

We have included some example groundtruth files for training as well as development set.

The posteriors features are extracted using the setup presented in the following paper: "High-performance query-by-example spoken term detection on the SWS 2013 evaluation".

The input feature files for training/evaluation are in pytorch readable format which are saved as python dictionaries. The keys are the names of the files in 'groundtruth files' and values are the features in matrix format.

Training

python query_detection_dtw_cnn.py -optim adam -learning_rate 0.0001 -input_size 152 -batch_size 50 -layers 9 -depth 30 -dropout 0.2 -loss_threshold 0.1 -n_valid 50 -max_batch_dev 250 -max_batch_train 1000

Evaluation

python query_detection_dtw_cnn_evaluation.py -input_size 152 -depth 30 -load_model -modelpath cnn_qbe_std_model.pt -outdir outpath -query_list dev_queries_sample_list.txt -search_list search_utterances_sample_list.txt

Reference

@inproceedings{ram2018cnn,
  title={CNN based Query by Example Spoken Term Detection},
  author={Ram, Dhananjay and Miculicich, Lesly and Bourlard, Herv{\'e}},
  booktitle={Nineteenth Annual Conference of the International Speech Communication Association (INTERSPEECH)},
  year={2018}
}

Contact:

dhananjay.ram@idiap.ch