WeSpeaker

WeSpeaker (wenet-e2e/wespeaker, ICASSP 2023, arXiv 2210.17016) is the speaker-modeling sibling of the WeNet speech-recognition toolkit, built by the same open-source community (X-LANCE Lab at Shanghai Jiao Tong, Tencent, AISpeech, Horizon Robotics, NVIDIA contributors). It focuses on one thing — learning high-quality speaker embeddings — and does it thoroughly: modern architectures (ResNet, ECAPA-TDNN, CAM++ and friends), a full menu of loss functions and scoring back-ends, structured recipes whose results were good enough for winning systems in speaker-verification challenges, and diarization shown as a downstream application of the embeddings. The production angle is real: a CLI (wespeaker --task diarization --audio_file audio.wav), a Python API (model.diarize('audio.wav')), CPU- and GPU-compatible deployment code, and an ONNX runtime, plus a Hugging Face demo. The license situation needs one careful read: the code is Apache 2.0, but the pretrained weights follow the license of their training dataset — e.g. VoxCeleb-trained models are CC-BY-4.0. If you want a Chinese-strong or challenge-grade embedding backbone for your own diarization or verification stack, this is the most complete open starting point.

• Competitive results across VoxCeleb and CNCeleb verification recipes, adopted in challenge-winning systems (author-reported, ICASSP 2023)\n- Diarization demonstrated as a downstream recipe; no single published DER headline for it — evaluate on your own data