Combining metric-learning and multi-clss classification for speaker dependent representation learning from speech
Python >= 3.6
Pytorch >= 1.2
Scikit-learn >=0.19
tqdm
h5py
Data preparation scripts are provided and features in Kaldi format are exepected.
Pre-processed data will consist of hdf files such that features for each recording are stored as datasets of shape [1, nfeat, nframes] and further stored under a group labeled with the speaker ID.
Prepare Kaldi features with data_prep.py. Arguments:
--path-to-data Path to scp files with features
--data-info-path Path to spk2utt and utt2spk
--spk2utt Path to spk2utt
--utt2spk Path to utt2spk
--path-to-more-data Path to extra scp files with features
--more-data-info-path Path to spk2utt and utt2spk
--more-spk2utt Path Path to spk2utt
--more-utt2spk Path Path to utt2spk
--out-path Path Path to output hdf file
--out-name Path Output hdf file name
--min-recordings Minimum number of train recordings for speaker to be included
Train and development data hdfs are expected.
Train models with train.py. Arguments:
--batch-size input batch size for training (default: 64)
--epochs number of epochs to train (default: 500)
--lr learning rate (default: 0.001)
--momentum Momentum paprameter (default: 0.9)
--l2 Weight decay coefficient (default: 0.00001)
--margin margin fro triplet loss (default: 0.3)
--lamb Entropy regularization penalty (default: 0.001)
--swap Swaps anchor and positive depending on distance to negative example
--patience Epochs to wait before decreasing LR by a factor of 0.5 (default: 10)
--checkpoint-epoch epoch to load for checkpointing. If None, training starts from scratch
--checkpoint-path Path for checkpointing
--pretrained-path Path for pre trained model
--train-hdf-file Path to hdf data
--valid-hdf-file Path to hdf data
--model {resnet_mfcc,resnet_lstm,resnet_stats,resnet_large,resnet_small}
--delta Enables extra data channels
--workers number of data loading workers
--seed random seed (default: 1)
--save-every how many epochs to wait before logging training status. Default is 1
--ncoef number of MFCCs (default: 23)
--latent-size latent layer dimension (default: 200)
--n-frames maximum number of frames per utterance (default: 800)
--n-cycles cycles over speakers list to complete 1 epoch
--valid-n-cycles cycles over speakers list to complete 1 epoch
--softmax {none,softmax,am_softmax}
--pretrain Adds softmax layer for speaker identification and train exclusively with CE minimization
--mine-triplets Enables distance mining for triplets
--no-cuda Disables GPU use
--no-cp Disables checkpointing
--verbose Verbose is activated if > 0
Scoring with a back-end of choice can be performed on top of embeddings which can be computed with embedd.py.
Arguments:
--path-to-data Path to input data
--cp-path Path for file containing model
--out-path Path to output hdf file
--model {resnet_mfcc,resnet_lstm,resnet_stats,resnet_large}
--latent-size latent layer dimension (default: 200)
--ncoef number of MFCCs (default: 23)
--no-cuda Disables GPU use
Simple cosine similarity evaluation of a give set of test trials can be performed with eval.py.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent