- What is deepaudio-speaker?
- Installation
- Get Started
- Model Architecture
- How to contribute to deepaudio-speaker?
- Acknowledge
Deepaudio-speaker is a framework for training neural network based speaker embedders. It supports online audio augmentation thanks to torch-audiomentation. It inlcudes or will include popular neural network architectures and losses used for speaker embedder.
To make it easy to use various functions such as mixed-precision, multi-node training, and TPU training etc, I introduced PyTorch-Lighting and Hydra in this framework (just like what pyannote-audio and openspeech do).
Deepaudio-tts is coming soon.
conda create -n deepaudio python=3.8.5
conda activate deepaudio
conda install numpy cffi
conda install libsndfile=1.0.28 -c conda-forge
git clone https://github.com/deepaudio/deepaudio-speaker.git
cd deepaudio-speaker
pip install -e .
####Voxceleb2
- Download VoxCeleb dataset and follow this script to obtain this kind of directory structure:
/path/to/voxceleb/voxceleb1/dev/wav/id10001/1zcIwhmdeo4/00001.wav
/path/to/voxceleb/voxceleb1/test/wav/id10270/5r0dWxy17C8/00001.wav
/path/to/voxceleb/voxceleb2/dev/aac/id00012/21Uxsk56VDQ/00001.m4a
/path/to/voxceleb/voxceleb2/test/aac/id00017/01dfn2spqyE/00001.m4a
- Example1: Train the
ecapa-tdnnmodel withfbankfeatures on GPU.
$ deepaudio-speaker-train \
dataset=voxceleb2 \
dataset.dataset_path=/your/path/to/voxceleb2/dev/wav/ \
model=clovaai_ecapa \
model.channels=1024 \
feature=fbank \
lr_scheduler=reduce_lr_on_plateau \
trainer=gpu \
criterion=pyannote_aamsoftmax
- Example2: Train ecapa model to get eer around 1.13% for voxceleb 1 trials ( original version, without norm operation).
$ git clone https://github.com/deepaudio/deepaudio-database.git
$ cd deepaudio-database
$ vim database.yml # edit the list path and wav path
$ deepaudio-speaker-train \
dataset=dataframe \
dataset.database_yml=/your/path/to/deepaudio-database/database.yml \
dataset.dataset_name=voxceleb2_dev \
model=clovaai_ecapa \
model.channels=1024 \
model.embed_dim=256 \
model.min_num_frames=200 \
model.max_num_frames=300 \
feature=fbank \
lr_scheduler=warmup_adaptive_reduce_lr_on_plateau \
lr_scheduler.warmup_steps=30000 \
lr_scheduler.lr_factor=0.8 \
trainer=gpu \
trainer.batch_size=128 \
trainer.max_epochs=30 \
trainer.num_checkpoints=30 \
criterion=adaptive_aamsoftmax \
criterion.increase_steps=300000 \
augment.apply_spec_augment=True\
augment.time_mask_num=1 \
augment.apply_noise_augment=True \
augment.apply_reverb_augment=True \
augment.apply_noise_reverb_augment=True \
augment.noise_augment_weight=2 \
augment.noise_dataset_dir=/your/path/to/musan \
augment.rir_dataset_dir=/your/path/to/RIRS_NOISES/simulated_rirs/ \
- Example3: Compute the equal error rate (EER)
from deepaudio.speaker.datasets.dataframe.utils import load_trial_dataframe, get_dataset_items
from deepaudio.speaker.models.inference import Inference
from deepaudio.speaker.metrics.eer import model_eer
trial_meta = get_dataset_items('/your/path/to/deepaudio-database/database.yml',
'voxceleb1_o', 'trial')
wav_dir, trial_path = trial_meta[0]
trials = load_trial_dataframe(wav_dir, trial_path)
inference = Inference('/your/path/to/checkpoint.ckpt')
eer, thresh = model_eer(inference, trials)- Example4: Export torchscript model
from deepaudio.speaker.models.inference import Inference
model = Inference('/your/path/to/checkpoint.ckpt').model
model.to_torchscript('filepath/to/model')Wespeaker Models from wespeaker.
ECAPA-TDNN This is an unofficial implementation from @lawlict. Please find more details in this link.
ECAPA-TDNN This is implemented by @joonson. Please find more details in this link.
ResNetSE34L This is borrowed from voxceleb trainer.
ResNetSE34V2 This is borrowed from voxceleb trainer.
Resnet101 This is proposed by BUT for speaker diarization. Please note that the feature used in this framework is different from VB-HMM
It is a personal project. So I don't have enough gpu resources to do a lot of experiments. I appreciate any kind of feedback or contributions. Please feel free to make a pull requsest for some small issues like bug fixes, experiment results. If you have any questions, please open an issue.
I borrow a lot of codes from openspeech and pyannote-audio