This is a complete implementation of 'Generalized End-to-End loss for speaker verification (GE2E Loss)'
- Generalized End-to-End Loss for Speaker Verification
- Please note:
- Data used here is very elementary.
- 10 speakers from
librispeech_test-otherand 4 additional speakers. Total 14 speakers.
- Please prepare you own data to train the model for your usecase.
- Since the model is trained on 14 people's voice, it can only identify those 14 people.
- If you wish, you could collect sample voices for 'N' different people and retrain the model to be able to identify those 'N' voices.
- For a given utterance (audio) of a speaker, this model produces a vector of length 256 (1x256).
- Therefore, a batch of inference might look something like below
-
[-0.0205, 0.0271, 0.0419, ..., -0.1035, 0.0387, 0.0905], [-0.0078, 0.0203, 0.0275, ..., -0.0943, 0.0301, 0.0814], ..., [ 0.0220, 0.0717, -0.0553, ..., 0.1109, -0.1149, 0.0084], [ 0.0164, 0.0770, -0.0502, ..., 0.1053, -0.1108, 0.0152], [ 0.0287, 0.0664, -0.0607, ..., 0.1179, -0.1123, 0.0064]], grad_fn=<DivBackward0>) embeddings.shape -> torch.Size([32, 256]) - The output produced by this model can be used to
- Voice detection
- Identify different speakers
- Voice cloning
- High fidelity voice generation
- etc.
- Python 3
- Numpy
- PyTorch
- librosa
| Embedding Model (GE2E) |
|---|
| Pre-trained embedding model |
-
Gather audio data -> different utterances from different people.
-
Create spectrogram of those audios.
-
Train the Speaker Embedding model.
-
Use the code from
step1_train_embedding_model.py- Results Step 1 (Embedding model using GE2E loss)
- With 4 Speakers
- With 6 Speakers
- With 10 Speakers
- Results Step 1 (Embedding model using GE2E loss)
