Rumeysakeskin / Speaker-Verification

Verifying the identity of a person from characteristics of the voice independent from language via NVIDIA NeMo models (ECAPA-TDNN, SpeakerNet, TitaNet-L).

Speaker Verification

Speaker verification is verifying the identity of a person from characteristics of the voice independent from language via NVIDIA NeMo.

This reporisitory presents three NeMo speaker verification models:

• SpeakerNet
• TitaNet-L
• ECAPA-TDNN

Download Models - Save and Load Speaker Vectors

You can download Nemo models and speaker vectors for SpeakerNet, TitaNet-L, ECAPA-TDNN from files/.

Prediction

The cosine similarity metric was used for prediction.

from sklearn.metrics.pairwise import cosine_similarity
sims = cosine_similarity([vector[0]], speakers_vectors)[0]

To predict most similar speaker in test_voices/ refered to ref_voices run the following command:

python speaker_verification.py

Inference Benchmark

• Torch Model
• Onnx Model Run the script below to compare the inference time of SpeakerNet model.

speaker_verification_with_torch_and_onnx.ipynb