subho406 / TF-Speech-Recognition-Challenge-Solution

Source code of the model used in Tensorflow Speech Recognition Challenge (https://www.kaggle.com/c/tensorflow-speech-recognition-challenge). The solution ranked in top 5% in private leaderboard.

TF Speech Recognition Challenge

Tensorflow Speech Recognition Challenge was a Kaggle competition organised by Google Brain to use the Speech Commands Dataset to build an algorithm that understands simple spoken commands. https://www.kaggle.com/c/tensorflow-speech-recognition-challenge

This solution achieved a rank of 63 on private leaderboard (top 5%).

Project Structure

• data
  • raw
    • train (Training audio files)
    • test (Test audio files used for evaluation
• libs
  • classification (All scripts used for training and evaluation)
• notebooks
• scripts (Executable scripts)
• models (Pretrained Models)

Requirements

1. Tensorflow 1.4
2. librosa
3. scikit-learn
4. Python 3.x

Running

Download the Speech Commands Dataset and extract the dataset in the train folder. Test Audio can be placed in data/test/audio folder.

The notebooks can be run individually using Jupyter. To run the scripts from command line edit the notebooks using Jupyter and run:

./script/execute_notebook.py

and select the notebook to run. The results are stored in results/notebook_name.log

P0 Predict Test WAV.ipynb can be used to predict audio files using a trained graphdef model.

Architecture

Models used

1. A variant of Convolutional LSTM (https://arxiv.org/pdf/1610.00277.pdf)
2. LSTM-L (https://arxiv.org/pdf/1711.07128.pdf)
3. C-RNN (https://arxiv.org/pdf/1711.07128.pdf)
4. GRU-L (https://arxiv.org/pdf/1711.07128.pdf)
5. Resnet

Training

The model was trained using a GCP instance with the following specifications:

• NVIDIA Tesla P100 X 1
• 16 GB RAM
• 35 GB SSD

Most of the models converged in 30k steps. Pseudo Labelling on test data was used to improve the model performance.

Prediction

The final model was a ensemble 13 models. Weighted Averaging and Stacking was used to generate the final predictions.

Aknowledgements

1. ML-KWS-for-MCU (https://github.com/ARM-software/ML-KWS-for-MCU)
2. Very Deep Convolutional Neural Network for Robust Speech Recognition (https://arxiv.org/pdf/1610.00277.pdf)
3. Speech Commands Dataset (https://research.googleblog.com/2017/08/launching-speech-commands-dataset.html)

If you like this project or have any queries don't hesitate to send an email to [e-mail]