3D-Speaker

3D-Speaker (modelscope/3D-Speaker, ICASSP 2025 toolkit paper, arXiv 2306.15354) is Alibaba DAMO Academy's open toolkit for speaker verification, recognition and diarization — notable for being one of the few open projects that does audio-visual diarization (face + voice) as well as audio-only. The model zoo is its calling card: CAM++ (7.2M params, 0.65% EER on VoxCeleb1-O), ERes2NetV2 (17.8M, 0.61% EER), ERes2Net-large (22.5M, 0.52% EER) — all competitive with the best published numbers. For diarization specifically, the repo's own benchmark table reports DER of 10.30% on Aishell-4 (vs 12.2% pyannote.audio, 11.7% DiariZen_WavLM), 11.75% on VoxConverse (vs 11.3% pyannote), 19.73% on Alimeeting (vs 24.4% pyannote), plus internal Chinese meeting sets where it leads comfortably. Those are provider-reported, but the head-to-head framing against named competitors is unusually transparent. It's also bundled with the large-scale 3D-Speaker corpus (multi-device, multi-distance, multi-dialect) for representation-disentanglement research. Everything is Apache 2.0, including models and corpus — the cleanest licensing story in this batch. Strong pick if your audio is Mandarin-heavy or you want audio-visual diarization without license anxiety.

• Speaker verification: CAM++ 0.65% EER, ERes2NetV2 0.61% EER, ERes2Net-large 0.52% EER on VoxCeleb1-O (provider-reported)\n- Diarization DER: Aishell-4 10.30% (pyannote 12.2%, DiariZen 11.7%); VoxConverse 11.75% (pyannote 11.3%); Alimeeting 19.73% (pyannote 24.4%); AMI_SDM 21.76% (provider-reported head-to-head)\n- No independent replication found