Speaker Recognition¶
Speaker Recognition (SR) can be categorized as two tasks: speaker identification (closed-set) and speaker verification (open set). The former classifies a speaker to a specific identity, while the latter determines whether a pair of utterances belongs to the same person. [1].
Currently Athena supports closed-set speaker recognition, i.e. speaker identification. The baseline system is built on ResNet-34 [2]. An example scipt for Voxceleb1 dataset and detailed results are ready at examples/speaker_identification/voxceleb. An example script for speaker verification on the Voxceleb dataset will be released soon.
To run experiments:
source tools/env.sh
bash examples/speaker_identification/voxceleb/run.sh
Core Stages:¶
(1) Data preparation¶
For data preparation, please run the script examples/speaker_identification/voxceleb/local/prepare_data.py. It downloads the voxceleb datasets, process the data (e.g. convert aac files to wav files, make speaker labels), and generates csv files.
(2) Data normalization¶
With the generated csv files, we should firstly compute the cmvn file for data normalization:
$ python athena/cmvn_main.py \
examples/speaker_identification/voxceleb/configs/speaker_resnet.json \
examples/speaker_identification/voxceleb/data/vox1_iden_train.csv
Or, it is okay to start training directly, the cmvn computation step will be done in the athena/main.py:
$ python athena/main.py examples/speaker_identification/voxceleb/configs/speaker_resnet.json
It will automatically compute the cmvn file before training if the cmvn file does not exist.
(3) Model training¶
You can train the model using json file:
$ python athena/main.py examples/speaker_identification/voxceleb/configs/speaker_resnet.json
(4) Inference¶
To test your model’s performance, run:
$ python athena/inference_main.py examples/speaker_identification/voxceleb/configs/speaker_resnet.json