C-STAR-database approach

来自cslt Wiki
2019年10月23日 (三) 23:10Cslt讨论 | 贡献的版本

跳转至: 导航搜索

C-STAR 中华名人音频数据收集

成员:王东,蔡云麒,李蓝天,周子雅,李开诚,陈浩林,程思潼,张鹏远,范悦,康嘉文

目标

  • Collect audio data of 1,000 Chinese celebrities.
  • Automatically clip videoes through a pipeline including face detection, face recognition, speaker validation and speaker diarization.
  • Create a database.

未来计划

  • Augment the database to 10,000 people.
  • Build a model between SyncNet and Speaker_Diarization based on LSTM, which can learn the relationship of them.


基本方法

  • Tensorflow, PyTorch, Keras, MxNet 实现
  • 检测、识别人脸的RetinaFace和ArcFace模型,说话人识别的SyncNet模型,Speaker Diarization的UIS-RNN模型
  • 输入为目标主人公的视频、目标主人公的面部图片
  • 输出为该视频中主人公声音片段的时间标签


项目GitHub地址

celebrity-audio-collection

项目报告

v1.0阶段性报告


参考文献

  • Deng et al., "RetinaFace: Single-stage Dense Face Localisation in the Wild", 2019. [1]
  • Deng et al., "ArcFace: Additive Angular Margin Loss for Deep Face Recognition", 2018, [2]
  • Wang et al., "CosFace: Large Margin Cosine Loss for Deep Face Recognition", 2018, [3]
  • Liu et al., "SphereFace: Deep Hypersphere Embedding for Face Recognition", 2017[4]
  • Zhong et al., "GhostVLAD for set-based face recognition", 2018. link
  • Chung et al., "Out of time: automated lip sync in the wild", 2016.link
  • Xie et al., "UTTERANCE-LEVEL AGGREGATION FOR SPEAKER RECOGNITION IN THE WILD", 2019. link
  • Zhang1 et al., "FULLY SUPERVISED SPEAKER DIARIZATION", 2018. link