eval/thirdparty/UniSpeech/src/examples/speaker_verification/README.md

Pre-training Representations for Speaker Verification

Pre-trained models

Model	Fix pre-train	Vox1-O	Vox1-E	Vox1-H
ECAPA-TDNN	-	1.080	1.200	2.127
HuBERT large	Yes	0.888	0.912	1.853
Wav2Vec2.0 (XLSR)	Yes	0.915	0.945	1.895
UniSpeech-SAT large	Yes	0.771	0.781	1.669
HuBERT large	No	0.585	0.654	1.342
Wav2Vec2.0 (XLSR)	No	0.564	0.605	1.23
UniSpeech-SAT large	No	0.564	0.561	1.23

How to use?

Environment Setup

1. pip install -r requirements.txt
2. Install fairseq code
   • For HuBERT_Large and Wav2Vec2.0 (XLSR), we should install the official fairseq.
   • For UniSpeech-SAT large, we should install the Unispeech-SAT fairseq code.

Example

Take unispeech_sat and ecapa_tdnn for example:

1. First, you should download the pre-trained model in the above table to checkpoint_path.
2. Then, run the following codes:
   • The wav files are sampled from voxceleb1.

python verification.py --model_name unispeech_sat --wav1 vox1_data/David_Faustino/hn8GyCJIfLM_0000012.wav --wav2 vox1_data/Josh_Gad/HXUqYaOwrxA_0000015.wav --checkpoint $checkpoint_path
# output: The similarity score between two audios is 0.0317 (-1.0, 1.0).

python verification.py --model_name unispeech_sat --wav1 vox1_data/David_Faustino/hn8GyCJIfLM_0000012.wav --wav2 vox1_data/David_Faustino/xTOk1Jz-F_g_0000015.wav --checkpoint --checkpoint $checkpoint_path
# output: The similarity score between two audios is 0.5389 (-1.0, 1.0).

python verification.py --model_name ecapa_tdnn --wav1 vox1_data/David_Faustino/hn8GyCJIfLM_0000012.wav --wav2 vox1_data/Josh_Gad/HXUqYaOwrxA_0000015.wav --checkpoint $checkpoint_path
# output: The similarity score between two audios is 0.2053 (-1.0, 1.0).

python verification.py --model_name ecapa_tdnn --wav1 vox1_data/David_Faustino/hn8GyCJIfLM_0000012.wav --wav2 vox1_data/David_Faustino/xTOk1Jz-F_g_0000015.wav --checkpoint --checkpoint $checkpoint_path
# output: he similarity score between two audios is 0.5302 (-1.0, 1.0).