Thanks VGGVox Oxford Model for the Speaker Recognition model generation. We would like to port to Python for realtime running

1> Overview • VGGVox is an end-to-end speaker verification model based on the VGG-M network architecture, trained on the VoxCeleb dataset. The code is based on the paper “VoxCeleb: a large-scale speaker identification dataset” by Nagrani et al and the MatLab verification evaluation script: https://github.com/a-nagrani/VGGVox. • The pretrained model weights are converted from the Matlab version downloaded from: https://www.robots.ox.ac.uk/~vgg/data/voxceleb/ • The project currently only contains evaluation code, not training. It supports 3 evaluation modes: offline (pre-recorded test samples), batch_offline (pre-recorded test samples, against a large number of enroll samples), online (real-time test sample). In the online mode, the script runs a loop that records a voice sample, computes score against all enrolled speakers, and outputs identification result.

2> Directory structure └── data/: data folder └── model_weights/model_0.h5: pretrained model weights └── wav/: contains wav files, each set organized in a folder └── lst/: contains index/list files └── test_list.csv: list of test utterances └── enroll_list.csv: list of enroll utterances └── res/: contains scoring results └── utils/: contains utility scripts └── compute_tests.py: main evaluation script └── constants.py: contains changeable parameters └── model.py: contains model architecture definition └── …

3> Instructions • Configurations ◦ Change list of test and enroll utterances in lst/. Each line has form: “filename speaker” ◦ Change parameters in constants.py. Notable ones are ▪ TEST_MODE: “online”, “offline”, or “batch_offline” ▪ COST_METRIC: scoring metric; “cosine” or “euclidean” ▪ MAX_SEC_ENROLL: maximum length of enroll samples ▪ ONLINE_RECORD_SEC: duration of real-time record • To start evaluation script: python3 compute_tests.py ◦ For offline and batch_offline test modes, results are stored in res/OFFLINE_RESULT_FILE. Results have form: [test utterance] [correct speaker] [scores against all enrolled speakers] [5 speakers with highest scores] [accuracy of top 1, top 1%, top 5%, top 10%] ◦ For online test mode, results will be printed to the console.