3D-Speaker is an open-source toolkit for single- and multi-modal speaker verification, speaker recognition, and speaker diarization. All pretrained models are accessible on [ModelScope](https://www.modelscope.cn/models?page=1&tasks=speaker-verification&type=audio). Furthermore, we present a large-scale speech corpus also called [3D-Speaker-Dataset](https://3dspeaker.github.io/) to facilitate the research of speech representation disentanglement.
 Please support our community by starring it 感谢大家支持
## Benchmark
The EER results on VoxCeleb, CNCeleb and 3D-Speaker datasets for fully-supervised speaker verification.
| Model | Params | VoxCeleb1-O | CNCeleb | 3D-Speaker |
|:-----:|:------:| :------:|:------:|:------:|
| Res2Net | 4.03 M | 1.56% | 7.96% | 8.03% |
| ResNet34 | 6.34 M | 1.05% | 6.92% | 7.29% |
| ECAPA-TDNN | 20.8 M | 0.86% | 8.01% | 8.87% |
| ERes2Net-base | 6.61 M | 0.84% | 6.69% | 7.21% |
| CAM++ | 7.2 M | 0.65% | 6.78% | 7.75% |
| ERes2NetV2 | 17.8M | 0.61% | **6.14%** | 6.52% |
| ERes2Net-large | 22.46 M | **0.52%** | 6.17% | **6.34%** |
The DER results on public and internal multi-speaker datasets for speaker diarization.
| Test | 3D-Speaker | [pyannote.audio](https://github.com/pyannote/pyannote-audio) | [DiariZen_WavLM](https://github.com/BUTSpeechFIT/DiariZen) |
|:-----:|:------:|:------:|:------:|
|[Aishell-4](https://arxiv.org/abs/2104.03603)|**10.30%**|12.2%|11.7%|
|[Alimeeting](https://www.openslr.org/119/)|19.73%|24.4%|**17.6%**|
|[AMI_SDM](https://groups.inf.ed.ac.uk/ami/corpus/)|21.76%|22.4%|**15.4%**|
|[VoxConverse](https://github.com/joonson/voxconverse)|11.75%|**11.3%**|28.39%|
|Meeting-CN_ZH-1|**18.91%**|22.37%|32.66%|
|Meeting-CN_ZH-2|**12.78%**|17.86%|18%|
## Quickstart
### Install 3D-Speaker
``` sh
git clone https://github.com/modelscope/3D-Speaker.git && cd 3D-Speaker
conda create -n 3D-Speaker python=3.8
conda activate 3D-Speaker
pip install -r requirements.txt
```
### Running experiments
``` sh
# Speaker verification: ERes2NetV2 on 3D-Speaker dataset
cd egs/3dspeaker/sv-eres2netv2/
bash run.sh
# Speaker verification: CAM++ on 3D-Speaker dataset
cd egs/3dspeaker/sv-cam++/
bash run.sh
# Speaker verification: ECAPA-TDNN on 3D-Speaker dataset
cd egs/3dspeaker/sv-ecapa/
bash run.sh
# Self-supervised speaker verification: SDPN on VoxCeleb dataset
cd egs/voxceleb/sv-sdpn/
bash run.sh
# Audio and multimodal Speaker diarization:
cd egs/3dspeaker/speaker-diarization/
bash run_audio.sh
bash run_video.sh
# Language identification
cd egs/3dspeaker/language-idenitfication
bash run.sh
```
### Inference using pretrained models from Modelscope
All pretrained models are released on [Modelscope](https://www.modelscope.cn/models?page=1&tasks=speaker-verification&type=audio).
``` sh
# Install modelscope
pip install modelscope
# ERes2Net trained on 200k labeled speakers
model_id=iic/speech_eres2net_sv_zh-cn_16k-common
# ERes2NetV2 trained on 200k labeled speakers
model_id=iic/speech_eres2netv2_sv_zh-cn_16k-common
# CAM++ trained on 200k labeled speakers
model_id=iic/speech_campplus_sv_zh-cn_16k-common
# Run CAM++ or ERes2Net inference
python speakerlab/bin/infer_sv.py --model_id $model_id
# Run batch inference
python speakerlab/bin/infer_sv_batch.py --model_id $model_id --wavs $wav_list
# SDPN trained on VoxCeleb
model_id=iic/speech_sdpn_ecapa_tdnn_sv_en_voxceleb_16k
# Run SDPN inference
python speakerlab/bin/infer_sv_ssl.py --model_id $model_id
# Run diarization inference
python speakerlab/bin/infer_diarization.py --wav [wav_list OR wav_path] --out_dir $out_dir
# Enable overlap detection
python speakerlab/bin/infer_diarization.py --wav [wav_list OR wav_path] --out_dir $out_dir --include_overlap --hf_access_token $hf_access_token
```
## Overview of Content
- **Supervised Speaker Verification**
- [CAM++](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-cam%2B%2B), [ERes2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-eres2net), [ERes2NetV2](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-eres2netv2), [ECAPA-TDNN](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-ecapa), [ResNet](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-resnet) and [Res2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-res2net) training recipes on [3D-Speaker](https://3dspeaker.github.io/).
- [CAM++](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-cam%2B%2B), [ERes2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-eres2net), [ERes2NetV2](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-eres2netv2), [ECAPA-TDNN](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-ecapa), [ResNet](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-resnet) and [Res2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-res2net) training recipes on [VoxCeleb](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/).
- [CAM++](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-cam%2B%2B), [ERes2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-eres2net), [ERes2NetV2](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-eres2netv2), [ECAPA-TDNN](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-ecapa), [ResNet](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-resnet) and [Res2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-res2net) training recipes on [CN-Celeb](http://cnceleb.org/).
- **Self-supervised Speaker Verification**
- [RDINO](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-rdino) and [SDPN](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-sdpn) training recipes on [VoxCeleb](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/)
- [RDINO](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-rdino) training recipes on [3D-Speaker](https://3dspeaker.github.io/).
- [RDINO](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-rdino) training recipes on [CN-Celeb](http://cnceleb.org/).
- **Speaker Diarization**
- [Speaker diarization](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/speaker-diarization) inference recipes which comprise multiple modules, including overlap detection[optional], voice activity detection, speech segmentation, speaker embedding extraction, and speaker clustering.
- **Language Identification**
- [Language identification](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/language-identification) training recipes on [3D-Speaker](https://3dspeaker.github.io/).
- **3D-Speaker Dataset**
- Dataset introduction and download address: [3D-Speaker](https://3dspeaker.github.io/)
- Related paper address: [3D-Speaker](https://arxiv.org/pdf/2306.15354.pdf)
## What‘s new :fire:
- [2024.12] Update [diarization](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/speaker-diarization) recipes and add results on multiple diarization benchmarks.
- [2024.8] Releasing [ERes2NetV2](https://modelscope.cn/models/iic/speech_eres2netv2_sv_zh-cn_16k-common) and [ERes2NetV2_w24s4ep4](https://modelscope.cn/models/iic/speech_eres2netv2w24s4ep4_sv_zh-cn_16k-common) pretrained models trained on 200k-speaker datasets.
- [2024.5] Releasing [SDPN](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-sdpn) model and [X-vector](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-xvector) model training and inference recipes for VoxCeleb.
- [2024.5] Releasing [visual module](https://github.com/modelscope/3D-Speaker/tree/main/egs/ava-asd/talknet) and [semantic module](https://github.com/modelscope/3D-Speaker/tree/main/egs/semantic_speaker/bert) training recipes.
- [2024.4] Releasing [ONNX Runtime](https://github.com/modelscope/3D-Speaker/tree/main/runtime/onnxruntime) and the relevant scripts for inference.
- [2024.4] Releasing [ERes2NetV2](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-eres2netv2) model with lower parameters and faster inference speed on VoxCeleb datasets.
- [2024.2] Releasing [language identification](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/language-identification) integrating phonetic information recipes for more higher recognition accuracy.
- [2024.2] Releasing [multimodal diarization](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/speaker-diarization) recipes which fuses audio and video image input to produce more accurate results.
- [2024.1] Releasing [ResNet34](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-resnet) and [Res2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-res2net) model training and inference recipes for 3D-Speaker, VoxCeleb and CN-Celeb datasets.
- [2024.1] Releasing [large-margin finetune recipes](https://github.com/modelscope/3D-Speaker/blob/main/egs/voxceleb/sv-eres2net/run.sh) in speaker verification and adding [diarization recipes](https://github.com/modelscope/3D-Speaker/blob/main/egs/3dspeaker/speaker-diarization/run.sh).
- [2023.11] [ERes2Net-base](https://modelscope.cn/models/damo/speech_eres2net_base_200k_sv_zh-cn_16k-common/summary) pretrained model released, trained on a Mandarin dataset of 200k labeled speakers.
- [2023.10] Releasing [ECAPA model](https://github.com/modelscope/3D-Speaker/tree/main/egs/voxceleb/sv-ecapa) training and inference recipes for three datasets.
- [2023.9] Releasing [RDINO](https://github.com/modelscope/3D-Speaker/tree/main/egs/cnceleb/sv-rdino) model training and inference recipes for [CN-Celeb](http://cnceleb.org/).
- [2023.8] Releasing [CAM++](https://modelscope.cn/models/damo/speech_campplus_sv_cn_cnceleb_16k/summary), [ERes2Net-Base](https://modelscope.cn/models/damo/speech_eres2net_base_sv_zh-cn_cnceleb_16k/summary) and [ERes2Net-Large](https://modelscope.cn/models/damo/speech_eres2net_large_sv_zh-cn_cnceleb_16k/summary) benchmarks in [CN-Celeb](http://cnceleb.org/).
- [2023.8] Releasing [ERes2Net](https://modelscope.cn/models/damo/speech_eres2net_base_lre_en-cn_16k/summary) annd [CAM++](https://modelscope.cn/models/damo/speech_campplus_lre_en-cn_16k/summary) in language identification for Mandarin and English.
- [2023.7] Releasing [CAM++](https://modelscope.cn/models/damo/speech_campplus_sv_zh-cn_3dspeaker_16k/summary), [ERes2Net-Base](https://modelscope.cn/models/damo/speech_eres2net_base_sv_zh-cn_3dspeaker_16k/summary), [ERes2Net-Large](https://modelscope.cn/models/damo/speech_eres2net_large_sv_zh-cn_3dspeaker_16k/summary) pretrained models trained on [3D-Speaker](https://3dspeaker.github.io/).
- [2023.7] Releasing [Dialogue Detection](https://modelscope.cn/models/damo/speech_bert_dialogue-detetction_speaker-diarization_chinese/summary) and [Semantic Speaker Change Detection](https://modelscope.cn/models/damo/speech_bert_semantic-spk-turn-detection-punc_speaker-diarization_chinese/summary) in speaker diarization.
- [2023.7] Releasing [CAM++](https://modelscope.cn/models/damo/speech_campplus_lre_en-cn_16k/summary) in language identification for Mandarin and English.
- [2023.6] Releasing [3D-Speaker](https://3dspeaker.github.io/) dataset and its corresponding benchmarks including [ERes2Net](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-eres2net), [CAM++](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-cam%2B%2B) and [RDINO](https://github.com/modelscope/3D-Speaker/tree/main/egs/3dspeaker/sv-rdino).
- [2023.5] [ERes2Net](https://modelscope.cn/models/damo/speech_eres2net_sv_zh-cn_16k-common/summary) and [CAM++](https://www.modelscope.cn/models/damo/speech_campplus_sv_zh-cn_16k-common/summary) pretrained model released, trained on a Mandarin dataset of 200k labeled speakers.
## Contact
If you have any comment or question about 3D-Speaker, please contact us by
- email: {yfchen97, wanghuii}@mail.ustc.edu.cn, {dengchong.d, zsq174630, shuli.cly}@alibaba-inc.com
## License
3D-Speaker is released under the [Apache License 2.0](LICENSE).
## Acknowledge
3D-Speaker contains third-party components and code modified from some open-source repos, including:
[Speechbrain](https://github.com/speechbrain/speechbrain), [Wespeaker](https://github.com/wenet-e2e/wespeaker), [D-TDNN](https://github.com/yuyq96/D-TDNN), [DINO](https://github.com/facebookresearch/dino), [Vicreg](https://github.com/facebookresearch/vicreg), [TalkNet-ASD
](https://github.com/TaoRuijie/TalkNet-ASD), [Ultra-Light-Fast-Generic-Face-Detector-1MB](https://github.com/Linzaer/Ultra-Light-Fast-Generic-Face-Detector-1MB), [pyannote.audio](https://github.com/pyannote/pyannote-audio)
## Citations
If you find this repository useful, please consider giving a star :star: and citation :t-rex::
```BibTeX
@article{chen20243d,
title={3D-Speaker-Toolkit: An Open Source Toolkit for Multi-modal Speaker Verification and Diarization},
author={Chen, Yafeng and Zheng, Siqi and Wang, Hui and Cheng, Luyao and others},
booktitle={ICASSP},
year={2025}
}
```
