This program classifies a given MP3-file of human speech into 2 types: American English or British English.

You need to put an MP3-file of someone's speech into the data/inference/mp3/ directory. The speech is better to be 17-25 seconds long.

ONE MAJOR RESTRICTION IS REQUIRED: SPEECH MUST BE EITHER AMERICAN OR BRITISH. This means that if you try to predict German/Spanish/French/Russian/dog-barking/noise the neural network will not be able to distinguish them and you will get American or British prediction anyway.

How does it work?

The algorithm can be described as following:

1. Convert an mp3 file into a wav file via FFmpeg.
2. Trim the first n and m seconds (default: n = 1, m = 1).
3. Remove all the silent gaps and noises between sentences and words so it becomes a continuous speech.
4. Cut this sound into pieces of a fixed size (default: 4 seconds) and save them.
5. Create spectrograms on different scales and combine them into one image. Short-time Fourier Transform and its variation are used to get spectrograms. Here's how it looks:

6. Load a network that is already trained out by me. Basically, I took the ResNet50 network pre-trained on the ImageNet dataset and added a custom output layer (transfer learning). So, it's a convolutional model:

7. Calculate an average prediction based on all pieces of the cut mp3.
8. Additionally, a saliency map can be generated.

An accuracy during the training stage was 0.68 - 0.73.