Awesome
Deep Speaker: An End-to-End Neural Speaker Embedding System.
Unofficial Tensorflow/Keras implementation of Deep Speaker | Paper | Pretrained Models.
Tested with Tensorflow 2.3, 2.4, 2.5 and 2.6.
Sample Results
Models were trained on clean speech data. Keep in mind that the performance will be lower on noisy data. It is advised to remove silence and background noise before computing the embeddings (by using Sox for example). There is a discussion on the topic: Silence / Background Noise similarity.
| Model name | Testing dataset | Num speakers | F | TPR | ACC | EER | Training Logs | Download model |
|---|---|---|---|---|---|---|---|---|
| ResCNN Softmax trained | LibriSpeech all(*) | 2484 | 0.789 | 0.733 | 0.996 | 0.043 | Click | Click |
| ResCNN Softmax+Triplet trained | LibriSpeech all(*) | 2484 | 0.843 | 0.825 | 0.997 | 0.025 | Click | Click |
(*) all includes: dev-clean, dev-other, test-clean, test-other, train-clean-100, train-clean-360, train-other-500.
The Softmax+Triplet checkpoint is also available on the Chinese cloud - WeiYun.
Overview
Deep Speaker is a neural speaker embedding system that maps utterances to a hypersphere where speaker similarity is measured by cosine similarity. The embeddings generated by Deep Speaker can be used for many tasks, including speaker identification, verification, and clustering.
Getting started
Install dependencies
Requirements
- tensorflow>=2.0
- keras>=2.3.1
- python>=3.6
pip install -r requirements.txt
If you see this error: libsndfile not found, run this: sudo apt-get install libsndfile-dev.
Training
The code for training is available in this repository. It takes a bit less than a week with a GTX1070 to train the models.
System requirements for a complete training are:
- At least 300GB of free disk space on a fast SSD (250GB just for all the uncompressed + processed data)
- 32GB of memory and at least 32GB of swap (can create swap with SSD space).
- A NVIDIA GPU such as the 1080Ti.
pip uninstall -y tensorflow && pip install tensorflow-gpu
./deep-speaker download_librispeech # if the download is too slow, consider replacing [wget] by [axel -n 10 -a] in download_librispeech.sh.
./deep-speaker build_mfcc # will build MFCC for softmax pre-training and triplet training.
./deep-speaker build_model_inputs # will build inputs for softmax pre-training.
./deep-speaker train_softmax # takes ~3 days.
./deep-speaker train_triplet # takes ~3 days.
NOTE: If you want to use your own dataset, make sure you follow the directory structure of librispeech. Audio files have to be in .flac. format. If you have .wav, you can use ffmpeg to make the conversion. Both formats are flawless (FLAC is compressed WAV).
Test instruction using pretrained model
- Download the trained models
| Model name | Used datasets for training | Num speakers | Model Link |
|---|---|---|---|
| ResCNN Softmax trained | LibriSpeech train-clean-360 | 921 | Click |
| ResCNN Softmax+Triplet trained | LibriSpeech all | 2484 | Click |
Note: the pre-training was performed on a subset of all the speakers we have. This is to match the philosophy of the paper where they first trained the model with softmax and then trained it on the whole dataset (bigger than this repo!) with triplets.
- Run with pretrained model
import random
import numpy as np
from deep_speaker.audio import read_mfcc
from deep_speaker.batcher import sample_from_mfcc
from deep_speaker.constants import SAMPLE_RATE, NUM_FRAMES
from deep_speaker.conv_models import DeepSpeakerModel
from deep_speaker.test import batch_cosine_similarity
# Reproducible results.
np.random.seed(123)
random.seed(123)
# Define the model here.
model = DeepSpeakerModel()
# Load the checkpoint. https://drive.google.com/file/d/1F9NvdrarWZNktdX9KlRYWWHDwRkip_aP.
# Also available here: https://share.weiyun.com/V2suEUVh (Chinese users).
model.m.load_weights('ResCNN_triplet_training_checkpoint_265.h5', by_name=True)
# Sample some inputs for WAV/FLAC files for the same speaker.
# To have reproducible results every time you call this function, set the seed every time before calling it.
# np.random.seed(123)
# random.seed(123)
mfcc_001 = sample_from_mfcc(read_mfcc('samples/PhilippeRemy/PhilippeRemy_001.wav', SAMPLE_RATE), NUM_FRAMES)
mfcc_002 = sample_from_mfcc(read_mfcc('samples/PhilippeRemy/PhilippeRemy_002.wav', SAMPLE_RATE), NUM_FRAMES)
# Call the model to get the embeddings of shape (1, 512) for each file.
predict_001 = model.m.predict(np.expand_dims(mfcc_001, axis=0))
predict_002 = model.m.predict(np.expand_dims(mfcc_002, axis=0))
# Do it again with a different speaker.
mfcc_003 = sample_from_mfcc(read_mfcc('samples/1255-90413-0001.flac', SAMPLE_RATE), NUM_FRAMES)
predict_003 = model.m.predict(np.expand_dims(mfcc_003, axis=0))
# Compute the cosine similarity and check that it is higher for the same speaker.
print('SAME SPEAKER', batch_cosine_similarity(predict_001, predict_002)) # SAME SPEAKER [0.81564593]
print('DIFF SPEAKER', batch_cosine_similarity(predict_001, predict_003)) # DIFF SPEAKER [0.1419204]
- Commands to reproduce the test results after the training
NOTE: For some reasons, the test-model does not work with tensorflow>2.3. Make sure to run pip install tensorflow==2.3 if you want to run those two commands below.
$ export CUDA_VISIBLE_DEVICES=0; python cli.py test-model --working_dir ~/.deep-speaker-wd/triplet-training/ --
checkpoint_file checkpoints-softmax/ResCNN_checkpoint_102.h5
f-measure = 0.789, true positive rate = 0.733, accuracy = 0.996, equal error rate = 0.043
$ export CUDA_VISIBLE_DEVICES=0; python cli.py test-model --working_dir ~/.deep-speaker-wd/triplet-training/ --checkpoint_file checkpoints-triplets/ResCNN_checkpoint_265.h5
f-measure = 0.849, true positive rate = 0.798, accuracy = 0.997, equal error rate = 0.025
When the triplet loss select the hard examples, then the training loss does not really decrease. Because the hard samples are always hard meaning they are on average above alpha. The test set should however decreased.
Further work
- LSTM model: https://github.com/philipperemy/deep-speaker/pull/53
- Fusion score: https://github.com/philipperemy/deep-speaker/pull/76