Spoken Language Recognition Using Convolutional Neural Networks

Written by Joscha S. Rieber (Fraunhofer IAIS) in 2020

This project shows how to train a language recognizer from scratch that is able to distinguish between German and English. These notebooks build up a playground together with the data from Common Voice to build strong models. With the data, pre-processing and model the accuracy is 93.8 %.

Have a Look at my Article on Towards AI

This repository is also described in more detail in my article published by Towards AI.

Getting Started

• On Linux:
  • Download this repository from GitHub
  • Call "bash run.sh"
    • This script will first look if the environment is ready, if not, it will download Miniconda and create the conda environment. Please note that you will need "wget" to succeed.
  • Now go through the notebooks in the right order and follow the given instructions.

System Requirements

A fast CPU is recommended for data augmentation and pre-processing. For the model training, a well-suited GPU is necessary. I have tested the scripts with an Nvidia P5000 and an Nvidia Tesla G80. The dataset coming from Mozilla Common Voice has a huge size. It might take a lot of time to process all of the data.

References

• Bartz et al.: Language Identification Using Deep Convolutional Recurrent Neural Networks
• Paul-Louis Pietz Prove: Spoken Language Recognition
• Sarthak et al.: Spoken Language Identification using ConvNets
• Szegedy et al.: Rethinking the Inception Architecture for Computer Vision
• Husein Zolkepli: Sound Augmentation Librosa