• Search by Module
• Search by Words
• Search Projects

• Java
• Python
• JavaScript
• TypeScript
• C++
• Scala
• Blog

• Spoken-language-identification-master
  • make_submission.py
  • create_spectrograms.py
  • ensembling
    • ensemble.theano.py
    • get_output_layers.py
  • prototxt
    • main_32r-2-64r-2-64r-2-128r-2-128r-2-256r-2-1024rd0.5-1024rd0.5_DLR.prototxt
    • solver.main.adadelta.prototxt
    • deploy.augm_32r-2-64r-2-64r-2-128r-2-128r-2-256r-2-1024rd0.3-1024rd0.3.prototxt
    • deploy.main_32r-2-64r-2-64r-2-128r-2-128r-2-256r-2-1024rd0.5-1024rd0.5_DLR.prototxt
    • solver.augm.nolrcoef.prototxt
    • augm_32r-2-64r-2-64r-2-128r-2-128r-2-256r-2-1024r-1024r_DLR_nolrcoef.prototxt
    • augm_32r-2-64r-2-64r-2-128r-2-128r-2-256r-2-1024rd0.3-1024rd0.3.prototxt
  • concatenate_csvs.py
  • theano
    • networks
      • tc_net_rnn.py
      • tc_net_deeprnn_shared_pad.py
      • tc_net_rnn_onernn_notimepool.py
      • base_network.py
      • tc_net_rnn_nodense.py
      • tc_net_mod_5khz_small.py
      • rnn_2layers.py
      • tc_net_rnn_shared_pad.py
      • tc_net.py
      • rnn_2layers_5khz.py
      • tc_net_rnn_onernn.py
      • __init__.py
      • tc_net_mod.py
      • rnn.py
      • tc_net_rnn_shared.py
      • tc_net_rnn_shared_pad_augm.py
    • plot.py
    • main.py
    • README.md
  • get_score_from_top3_prediction.py
  • LICENSE
  • augment_data.py
  • choose_equal_split.py
  • majority_vote_ensembling.py
  • get_score_from_probabilities.py
  • test_augm_network.py
  • test_main_network.py
  • README.md
  • get_sum_of_csvs.py
  • .gitignore

Spoken language identification with deep learning

Read more in the following blog posts:

• About TopCoder contest and our CNN-based solution implemented in Caffe (October 2015)
• About combining CNN and RNN using Theano/Lasagne (June 2016)

Theano/Lasagne models are here. The basic steps to run them are:

• Download the dataset from here or use your own dataset.
• Create spectrograms for recording using create_spectrograms.py or augment_data.py. The latter will also augment the data by randomly perturbing the spectrograms and cropping a random interval of length 9s from the recording.
• Create listfiles for training set and validation set, where each row of the a listfile describes one example and has 2 values seperated by a comma. The first one is the name of the example, the second one is the label (counting starts from 0). A typical listfile will look like this.
• Change the png_folder and listfile paths in theano/main.py.
• Run theano/main.py.