philipperemy/deep-speaker

Implements a ResNet-CNN architecture trained with softmax pre-training followed by triplet loss to generate fixed 512-dimensional speaker embeddings on a hypersphere, enabling speaker identification, verification, and clustering via cosine similarity. Built on TensorFlow/Keras with pretrained models available on LibriSpeech (achieving EER of 0.025 on 2484 speakers), it processes audio through MFCC feature extraction and supports end-to-end training from raw FLAC/WAV files or inference with provided checkpoints.

939 stars. No commits in the last 6 months.