1nF0rmed/speech_recog

Speech Recognition for Tensorflow challenge

The goal is to assign one of the following 12 labels to each command: yes, no, up, down, left, right, on, off, stop, go, silence, unknown.

I used mel-scaled spectrograms and mel-frequency cepstral coefficients as inputs for two NASNet-A Convolutional Neural Networks and then averaged their predictions.

My solution

I used PyTorch to train two NASNet-A Convolutional Neural Networks. The First network was trained on mel-scaled spectrograms, the second - on mel-frequency cepstral coefficients. Then I averaged their predictions to make a final submission.

Examples of mel-scaled spectrograms for speech commands:

Requirements

• Python 3.6
• PyTorch 0.4.0a0. If you want to use version 0.3, you need to modify train.py and predict.py files:
  • remove with torch.no_grad and pass volatile=True when you are creating Variables while running forward_pass in a validation mode).
  • remove with torch.no_grad and pass volatile=True when you are creating a Variable.
• Libraries from requirements.txt

To run the code

• Adjust config variables in config.py
• Execute run.sh file