WeSpeaker 1.2.0

What's Changed

• Add a recipe for the validation set of VoxSRC-23's diarization track by @xx205 in #166
• Support the SphereFace2 loss function by @Hunterhuan in #173
• Support the Self-Supervised Learning (SSL) recipe on Voxceleb dataset, including DINO, MoCo, and SimCLR, by @czy97 and @Hunterhuan in #180
• Support the NIST SRE16 recipe by @czy97 in #177
• Support the kaldi-compatible PLDA and unsupervised adaptation by @wsstriving in #186

WeSpeaker 1.1.0

What's Changed

• Support the RepVGG model by @cdliang11 in #102
• Support Automatic Mixed Precision (AMP) training by @czy97 in #103
• Add Triton GPU deployment for diarization pipeline by @wd929 in #113
• Support Multi-Query Multi-Head Attentive Pooling (MQMHASTP) and Intertopk-Subcenter loss by @Hunterhuan in #115
• Support C++ onnxruntime by @cdliang11 in #135
• Support the CAM++ model by @JiJiJiang in #153
• Add more pretrained models by @wsstriving, @czy97 and @JiJiJiang

WeSpeaker 1.0.0

Highlight

• Competitive results: compared with SpeechBrain, ASV-Subtools, etc
• Light-weight: clean and simple codes, no Kaldi dependency
• Unified IO (UIO): designed for large-scale training data
• On-the-fly feature preparation: provide different data augmentation methods
• Distributed training: adopted for multi-node multi-GPU scalability
• Production ready: support TensorRT or ONNX exporting format, with a triton inference server demo
• Pre-trained models: provide the python bindings, and a Hugging face interactive demo on speaker verification

Overall Structure

Recipes

We provide three well-structured recipes:

• Speaker Verification: VoxCeleb an CNCeleb (SOTA results)
• Speaker Diarization: VoxConverse (An example of using pre-trained speaker model)

Support List

• SOTA Models: TDNN-based x-vector, ResNet-based r-vector, and ECAPA_TDNN
• Pooling Functions: statistics-based TAP/TSDP/TSTP, and attention-based ASTP
• Criteria: standard Softmax, and margin-based A-/AM-/AAM-Softmax
• Scoring: Cosine, PLDA, and Score Normalization (AS-Norm)
• Metric: EER, minDCF (DET curve), and DER
• Online Augmentation: Resample, Noise && RIR, Speed Perturb, and SpecAug
• Training strategies: Well-designed learning-rate and margin schedulers, Large margin fine-tuning