astorfi/3D-convolutional-speaker-recognition

Implements text-independent speaker verification using 3D-CNNs to jointly model temporal and spectral information from speech utterances, capturing both speaker identity and within-speaker variation. The architecture processes MFEC features (log-energies without DCT) extracted from overlapping 20ms windows, feeding multiple speaker utterances simultaneously through the network for direct speaker model creation rather than averaging d-vectors. Built on TensorFlow with Slim API, following a three-phase protocol: development (utterance-level speaker classification), enrollment (feature extraction for speaker model), and evaluation (test utterance comparison against stored models).

792 stars. No commits in the last 6 months.