An end to end deep learning model for speaker identification. The model exploits the idea of One-shot learning.

Requirements

This project assumes that you have the following python modules installed:

• Tensorflow V1.
• Pandas
• pysptk
• hickle

Training

Creating features

This model uses MFCCs, their deltas and double deltas as input features. To create these features, run the command

python Utils/features.py "path_to_audio_clips" "path_to_train_tsv_file"

The above command will create the features and store in path features/data/.

Creating the One Shot Dataset

Run the command python Utils/datasets.py "path_to_train_tsv_file". This will create SiameseDataset.npy at path Utils/.

Training the model

Run the command python spk_identification.py "path_to_train_tsv_file" "SiameseDataset.npy".

Evaluating the model

Run the command python Evaluation.py "path_to_train_tsv_file". I trained this model on training set of 2432 voice samples having 28 different speakers. These examples were taken from spanish Common Voice dataset. After training for 10 epochs, both the training and validation accuracy was ~0.97.

Plots showing model performance

Embeddings

We plotted the embeddings we got from our testing set of 400 voice samples taken from the trained model. This dataset also had the same 28 speakers. We did tSNE over the embeddings.

Author: Anurag Kumar