Transformer Speaker Encoder

Implemented a transformer-based speaker recognition system using Pytorch.

The model is trained on LibriSpeech and evaluated on Equal Error Rate (EER). After training, the encoder embeddings can be used to identify if two audios are from the same speaker.

Update model hyperparameter

open config.py to edit and adjust the necessary model parameters as needed

Train a model

Run:

python3 encoder.py


To begin training the model on train data specified by path TRAIN_DATA_DIR in config.py

Once training is done, the trained model will be saved to disk, based on the path specified by SAVED_MODEL_PATHin config.py.

Evaluate the model

Once the model has been trained and saved to disk

Run:

python3 eval.py


to evaluate the Equal Error Rate (EER) of the model on test data.

Adjust NUM_EVAL_TRIPLETS in config.py to control how many positive and negative trials you want to evaluate.

When the evaluation job finishes, the EER will be printed to screen

‍

View source code on GitHub

References

1. Speaker Recognition, Wang Q.

‍