Source-linked AI summary
Clova Baseline System for the VoxCeleb Speaker Recognition Challenge 2020
Hee Soo Heo, Bong-Jin Lee, Jaesung Huh, Joon Son Chung
TL;DR
Speaker recognition in VoxSRC 2020 must handle domain shift and shorter test utterances. The report evaluates ResNet variants with multiple losses, including combined Angular Prototypical and softmax objectives, and reports strong challenge performance without ensemble or post-processing while releasing code and models as unofficial baselines.
Problem
VoxSRC 2020 evaluates speaker recognition under explicit training-test domain shift and test utterances shorter than training segments.
Method
The system evaluates ResNet speaker models with metric-learning and classification losses, including Angular Prototypical loss combined with vanilla softmax.
Results
The best model outperforms all single-model systems and all but one ensemble system from the previous challenge, producing 5.19% EER and 0.314 MinDCF on VoxSRC 2020.
Takeaways & Limitations
The released training code and pre-trained models provide unofficial baselines for the VoxSRC 2020 challenge.
Abstract
from arXiv · showhide
This report describes our submission to the VoxCeleb Speaker Recognition Challenge (VoxSRC) at Interspeech 2020. We perform a careful analysis of speaker recognition models based on the popular ResNet architecture, and train a number of variants using a range of loss functions. Our results show significant improvements over most existing works without the use of model ensemble or post-processing. We release the training code and pre-trained models as unofficial baselines for this year's challenge.
1. Introduction
VoxSRC 2020 evaluates speaker recognition in unconstrained, “in the wild” data under explicit domain shift and shorter test utterances than training segments.
- VoxSRC 2020 assesses speaker recognition technology on unconstrained or “in the wild” data.
- The challenge introduces explicit domain shift between training and test data.
- The test set contains utterances shorter than the segments used during training.
2. Model
The system uses ResNet-based speaker models with alternative trunk architectures and combines metric-learning and classification objectives, including AP+Softmax.
- Trunk architecture: Two 34-layer ResNet variants trade computational cost against performance through different channel widths and pooling mechanisms.The speed-optimised Q / SAP model has 1.4 million parameters, while the performance-optimised H / ASP model has 8.0 million parameters and uses attentive statistics pooling.
- Trunk architecture: The performance-optimised model uses attentive statistics pooling, which combines weighted means with channel-wise weighted standard deviations.
- Loss function: AM-Softmax and AAM-Softmax introduce margins between classes to increase inter-class variance.
- Loss function: Angular Prototypical loss provides an angular objective for prototypical-network training without manually defined hyper-parameters.
- Loss function: Combining Angular Prototypical loss with vanilla softmax improves performance over using either loss function alone.
3. Experiments
The experiments train models on VoxCeleb2 with speech augmentations, evaluate them using crop-averaged cosine scores and EER/MinDCF, and find strong results for combined losses and output batch normalisation.
- Data and augmentation: Models are trained on the VoxCeleb2 development set containing 5,994 speakers, with VoxCeleb1 and the previous VoxSRC test set used for validation.
- Data and augmentation: Training augmentation uses additive speech, music, background noise, and simulated room impulse responses selected randomly at each training step.
- Evaluation protocol: Evaluation averages 100 cosine similarities computed from ten 4-second crops sampled from each test segment.
- Evaluation protocol: The reported metrics are Equal Error Rate and minimum detection cost using the specified NIST SRE and VoxSRC cost function.
- Results: The sum of metric-learning and classification-based losses works best in most scenarios, while output batch normalisation significantly improves classification objectives.
- Results: 5.19% EER and 0.314 MinDCF are obtained on VoxSRC 2020 by the performance-optimised AP+Softmax model without embedding batch normalisation.
4. Conclusion
The report presents a VoxSRC 2020 baseline trained with metric-learning and classification objectives, and releases code and pre-trained models as unofficial baselines.
- The proposed system combines metric-learning and classification-based training objectives.
- The best model outperforms all single-model systems and all but one ensemble system from the previous year’s challenge.
- The authors release full training code and pre-trained models as unofficial challenge baselines.