This is a model for identifying the language spoken in a short audio segment.

To install the required libraries (tested on Ubuntu 17.11) run:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

1. Convert an audio file to a spectrogram:

    python data/dataset_gen.py -z speech.wav -o .
   

2. Obtain the prediction using a pre-trained model:

    python main.py --model-dir your-trained-model/ --params your-trained-model/params.json --model combo --predict speech.png
   

1. Prepare a dataset:

   • Place your spectrograms in a folder
   • Create a test set CSV file containing "Filename,Language" pairs
   • Create an evaluation set CSV file (same format as the test)

2. Train the model:

    python main.py --model-dir your-trained-model/ --params your-trained-model/params.json --model combo --image-dir your-data/ --train-set your-data/train-set.csv --eval-set your-data/eval-set.csv
   

This project was developed by Rimvydas Naktinis during Pi School's AI programme in Fall 2017.