runngezhang / srpipeline

A Speaker Recognition Pipeline

A Basic Speaker Recognition Pipeline

dataset

In this pipeline I use THCHS-30 open-source dataset, you can get free access to it via http://openslr.org/18/

For higher accuracy, try http://openslr.org/33/ and http://openslr.org/38/ with slight change of data pre-processing.

models

ResCNN and GRU model from Baidu(https://arxiv.org/abs/1705.02304) and 3D-CNN model by A.Torfi et al.(https://arxiv.org/abs/1705.09422) are used here.

usage

feature extraction

python main.py extract

training

for softmax pre-training, use:

python main.py train --triplet=False

for fine-tunning based on triplets, use:

python main.py train --triplet=True

BTW, to understand triplets loss and its application, you can take a look at this video: https://www.youtube.com/watch?v=d2XB5-tuCWU

test

So far only ResCNN and GRU models are supported.

python main.py test

ONNX model exportation and deployment

first, run this

python main.py export

then, follow the instructions here(https://oracle.github.io/graphpipe/#/guide/servers/serving) to serve your model and do inference.

TODO

• try larger dataset
• hyper-parameter tuning
• visualization
• rewrite some parts(e.g. Variable -> Tensor)
• fix bugs
• ...