Source-linked AI summary
BUT System Description to VoxCeleb Speaker Recognition Challenge 2019
Hossein Zeinali, Shuai Wang, Anna Silnova, Pavel Matějka, Oldřich Plchot
TL;DR
The report addresses BUT’s submissions to the Fixed and Open tracks of the VoxSRC 2019 speaker-recognition challenge. It combines DNN systems based on x-vector and ResNet34 embeddings, with condition-specific training data and fusion strategies. The resulting evaluation performance was 1.42% EER for Fixed and 1.26% EER for Open.
Problem
The report evaluates speaker-recognition systems under Fixed and Open VoxSRC conditions with different constraints on available training data.
Method
BUT fuses four CNN-based systems using x-vector and ResNet34 embeddings, with AAM fine-tuning and condition-specific fusion strategies.
Results
1.42% EER was achieved for the Fixed evaluation system and 1.26% EER for the Open evaluation fusion.
Takeaways & Limitations
The Open fusion outperformed the Fixed fusion, which the authors associate with additional training data used for Open systems.
Abstract
from arXiv · showhide
In this report, we describe the submission of Brno University of Technology (BUT) team to the VoxCeleb Speaker Recognition Challenge (VoxSRC) 2019. We also provide a brief analysis of different systems on VoxCeleb-1 test sets. Submitted systems for both Fixed and Open conditions are a fusion of 4 Convolutional Neural Network (CNN) topologies. The first and second networks have ResNet34 topology and use two-dimensional CNNs. The last two networks are one-dimensional CNN and are based on the x-vector extraction topology. Some of the networks are fine-tuned using additive margin angular softmax. Kaldi FBanks and Kaldi PLPs were used as features. The difference between Fixed and Open systems lies in the used training data and fusion strategy. The best systems for Fixed and Open conditions achieved 1.42% and 1.26% ERR on the challenge evaluation set respectively.
1. Introduction
The report presents BUT’s VoxSRC 2019 submissions across Fixed and Open tracks, using fused DNN speaker-embedding systems based on x-vector and ResNet34 topologies.
- The challenge separates Fixed training, restricted to VoxCeleb-2 development data, from Open training, which permits any data.
- All BUT systems use DNN-based speaker embeddings, including one-dimensional CNNs in an x-vector extraction topology.The x-vector topology was modified with more neurons and residual connections.
- The second embedding approach uses a deep ResNet34 structure with two-dimensional CNN layers and residual connections.Residual connections are described as supporting training robustness.
- The report covers challenge setup, x-vector and ResNet34 systems, backends, fusion strategies, and results analysis.
2. Experimental Setup
The experiments vary training data, augmentation, features, and evaluation trials across the Fixed and Open conditions, using VoxCeleb-derived development data and additional speaker datasets for Open systems.
- Fixed embedding training combines original speech with approximately 5 million augmented segments, including reverberation, noise, music, and babble conditions.
- Open training adds VoxCeleb-1 development data and, for some systems, speakers from LibriSpeech and DeepMine.The setup uses around 1,152 VoxCeleb-1 speakers, 2,338 LibriSpeech speakers, and 1,735 DeepMine speakers.
- Evaluation reports results on cleaned VoxCeleb1, VoxCeleb1-E, and VoxCeleb1-H trial lists, with VoxCeleb1-O denoting the original test set.
- The systems use 30-dimensional Kaldi PLP or 40-dimensional Kaldi FBank features with three-second sliding-window mean normalization.
3. DNN based Systems
BUT develops DNN speaker embeddings with TDNN x-vector and ResNet34 architectures, uses Kaldi features and VAD, and fine-tunes selected networks with additive angular margin loss.
- All DNN embeddings use the Energy-based VAD from the Kaldi SRE16 recipe.The report states that different VAD choices had no significant performance impact within the DNN embedding paradigm.
- x-vectors: Table 1 defines x-vector topology symbols K, T, and N as feature dimensionality, training-segment frames, and speaker count.
- x-vectors: The x-vector systems use a TDNN topology trained with modified feature, epoch, segment-generation, and network-size settings.The modifications include PLP or FBANK features, six epochs, 200-frame segments, more neurons, and a larger network.
- ResNet34: The ResNet embedding uses two-dimensional features, two-dimensional CNN layers, and both mean and standard-deviation statistics, producing r-vectors.The networks were trained with SGD and used L2 regularization.
- AAM fine-tuning: AAM fine-tuning removes layers after the embedding layer and fine-tunes the remaining ResNet or TDNN network.The networks are first trained with normal Softmax before AAM fine-tuning.
- ResNet34: Table 2 describes the ResNet34 architecture, with input dimensions representing filter-bank count and frame count and N representing speakers.
4. Backend
The backend uses PLDA or cosine scoring, while fusion differs by condition: hand-weighted averaging for Fixed and trained logistic-regression fusion for Open.
- Backend scoring: The PLDA backend is trained on embeddings from 500,000 original VoxCeleb-2 utterances and applies centering, non-reducing LDA, length normalization, and 312-dimensional speaker and channel subspaces.
- Backend scoring: Fine-tuned ResNet embeddings use cosine scoring after centering, without additional preprocessing of the 256- or 160-dimensional embeddings.
- Score normalization: Adaptive symmetric score normalization selects the top 300 cohort scores for normalization, using a cohort formed by averaging x-vectors per speaker.
- Fixed fusion: Fixed fusion averages four system scores with hand-picked weights, assigning 0.4 to each ResNet cosine system and 0.1 to each remaining system.
- Open fusion: Open fusion uses VoxCeleb1-O trials for logistic-regression calibration and fusion, with scalar offset and system-weight parameters optimized by crossentropy.Scores are precalibrated before fusion and recalibrated afterward.
5. Results and Analysis
The BUT team’s fixed and open submissions used closely related fusion-based systems, but the open condition added training data. Final evaluation performance was 1.42% EER for fixed and 1.26% EER for open.
- Fixed condition: 1.42% EER was achieved by the fixed-condition fusion on the evaluation set.The fusion combined four systems eligible under the fixed condition.
- Fixed condition: The fixed fusion improved only marginally over the previous two-system submission, which scored 1.49% EER.Adding two components did not dramatically improve results, and logistic-regression fusion offered little gain over a weighted average.
- Open condition: The open-condition systems used VoxCeleb1 and VoxCeleb2, with some systems additionally trained on LibriSpeech and DeepMine.The final open submission fused three open systems and one fixed ResNet160 system.
- Open condition: 1.26% EER was achieved by the open-condition fusion on the evaluation set.The reported open-condition fusion included the additional training data used by the open systems.
- Comparison: The fixed and open submissions were similar, while additional training data was identified as the main difference and a possible reason for the open fusion’s improved performance.The authors attribute the difference to the extra data used for the open systems.