search

wenet-e2e / wespeaker

Research and Production Oriented Speaker Verification, Recognition and Diarization Toolkit
Apache License 2.0
664 stars 112 forks source link

QMF and fusion in VoxCeleb Speaker Recognition Challenge 2023 papers #206

Closed wendongj closed 3 months ago

wendongj commented 10 months ago

greatly thanks to the project of speaker verification. in VoxSRC 2023, many team use QMF and ASnorm to improve the score, seems QMF is as good as ASnorm, if it is possible to add the QMF in this project? and they also use Fusion to fuse different model result, seems this can improve performance too. sorry to ask these requirement.

JiJiJiang commented 10 months ago
  1. QMF will be supported soon.
  2. There are several ways of fusions: score fusion, embedding average and score, etc. It is easy to implement I think.
wendongj commented 10 months ago
  1. QMF will be supported soon.
  2. There are several ways of fusions: score fusion, embedding average and score, etc. It is easy to implement I think.

really thanks for your reply, I will fuse them as you recommend.

wsstriving commented 10 months ago
  1. QMF will be supported soon.
  2. There are several ways of fusions: score fusion, embedding average and score, etc. It is easy to implement I think. @Hunterhuan Can you follow this?
wsstriving commented 10 months ago

@Hunterhuan Can you give a brief schedule on this topic?

Hunterhuan commented 10 months ago

@Hunterhuan Can you give a brief schedule on this topic?

sure, I am current working on it. I will immediately contribute the QMF, once the code is completed. This will take approximately two weeks.

wsstriving commented 9 months ago

@Hunterhuan Any update?

WhXmURandom commented 7 months ago

@Hunterhuan Can you give a brief schedule on this topic?

sure, I am current working on it. I will immediately contribute the QMF, once the code is completed. This will take approximately two weeks.

Hello, when will the code for QMF be updated?

wsstriving commented 5 months ago

@Hunterhuan I will pin this issue for high priority

Hunterhuan commented 3 months ago

greatly thanks to the project of speaker verification. in VoxSRC 2023, many team use QMF and ASnorm to improve the score, seems QMF is as good as ASnorm, if it is possible to add the QMF in this project? and they also use Fusion to fuse different model result, seems this can improve performance too. sorry to ask these requirement.

Score calibration is implemented for both voxceleb and cnceleb. Ref: https://github.com/wenet-e2e/wespeaker/pull/320

wsstriving commented 3 months ago

The PR https://github.com/wenet-e2e/wespeaker/pull/320 has been merged