Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/juanmc2005/speakerembeddinglosscomparison

Companion repository for the paper "A Comparison of Metric Learning Loss Functions for End-to-End Speaker Verification" published at SLSP 2020
https://github.com/juanmc2005/speakerembeddinglosscomparison

additive-angular-margin-loss end-to-end-machine-learning metric-learning pytorch sincnet speaker-embedding speaker-verification x-vector

Last synced: 11 days ago
JSON representation

Companion repository for the paper "A Comparison of Metric Learning Loss Functions for End-to-End Speaker Verification" published at SLSP 2020

Awesome Lists containing this project

README 

        
# A Comparison of Metric Learning Loss Functions for End-to-End Speaker Verification



This is the companion repository for the paper [A Comparison of Metric Learning Loss Functions for End-to-End Speaker Verification](https://arxiv.org/abs/2003.14021), published at the [SLSP 2020](https://irdta.eu/slsp2020/) conference. It hosts our best model trained with additive angular margin loss, and contains instructions for reproducing our results and using the model.



## Architecture



![Architecture](images/architecture.png?raw=true "Architecture")



The architecture of our model consists of [SincNet](https://arxiv.org/abs/1808.00158) for feature extraction followed by [x-vector](https://www.danielpovey.com/files/2018_icassp_xvectors.pdf).



## Training



You can train our model from scratch using the configuration file `config.yml` that we provide. All you need to do is run the following commands in your terminal:



```console

$ export EXP=models/AAM # Replace with the new path to config.yml

$ export PROTOCOL=VoxCeleb.SpeakerVerification.VoxCeleb2

$ pyannote-audio emb train --parallel=10 --gpu --to=1000 $EXP $PROTOCOL 

```



Note that you may need to change parameters based on your setup.



## Evaluation



We provide a step-by-step guide on reproducing our equal error rates alongside their 95% confidence intervals. The guide first evaluates the pretrained model using raw cosine distances, and then improves it with adaptive s-norm score normalization.



If you want to reproduce our results, check out [this notebook](https://github.com/juanmc2005/SpeakerEmbeddingLossComparison/blob/master/reproduce.ipynb)



## Fine-tuning



You can fine-tune our model to your dataset with the following commands:



```console

$ export WEIGHTS=models/AAM/train/VoxCeleb.SpeakerVerification.VoxCeleb2.train/weights/0560.pt

$ export EXP=

$ export PROTOCOL=

$ pyannote-audio emb train --pretrained $WEIGHTS --gpu --to=1000 $EXP $PROTOCOL

```



## Inference in Python



The default pyannote model for speaker embedding on `torch.hub` is our AAM loss model trained on variable length audio chunks. If you want to use the model right away, you can do so easily in a Python script:



```python

# load pretrained model from torch.hub

import torch

model = torch.hub.load('pyannote/pyannote-audio', 'emb')



# extract embeddings for the whole files

emb1 = model({'audio': '/path/to/file1.wav'})

emb2 = model({'audio': '/path/to/file2.wav'})



# compute distance between embeddings

from scipy.spatial.distance import cdist

import numpy as np

distance = np.mean(cdist(emb1, emb2, metric='cosine'))

```



You can also replace the call to `torch.hub.load` with a pyannote `Pretrained` instance pointing to the model in this repo:



```python

from pyannote.audio.features import Pretrained

model = Pretrained(

    'models/AAM/train/VoxCeleb.SpeakerVerification.VoxCeleb2.train/validate_equal_error_rate/'

    'VoxCeleb.SpeakerVerification.VoxCeleb1_X.development', step=0.0333)



print(f'Embeddings of {model.sliding_window.duration:g}s duration and of dimension {model.dimension:d}, '

      f'extracted every {1000 * model.sliding_window.step:g}ms')

```



## Some Compatibility Notes



This project depends on the [pyannote-audio](https://github.com/pyannote/pyannote-audio) toolkit, so make sure you install it before running any code.



Under normal circumstances, everything should work with the newest version of pyannote. However, given that pyannote is constantly evolving, some compatibility issues may appear. To make sure these don't happen, you can install the version at [this commit](https://github.com/pyannote/pyannote-audio/commit/562b74e4e2c7c2e97ad8eaabd1c95015c7e41e16) from the `develop` branch.



## Citation



If our work has been useful to you, please cite our paper:



```bibtex

@InProceedings{10.1007/978-3-030-59430-5_11,

    author="Coria, Juan M.

    and Bredin, Herv{\'e}

    and Ghannay, Sahar

    and Rosset, Sophie",

    editor="Espinosa-Anke, Luis

    and Mart{\'i}n-Vide, Carlos

    and Spasi{\'{c}}, Irena",

    title="{A Comparison of Metric Learning Loss Functions for End-To-End Speaker Verification}",

    booktitle="Statistical Language and Speech Processing",

    year="2020",

    publisher="Springer International Publishing",

    address="Cham",

    pages="137--148",

    isbn="978-3-030-59430-5"

}

```