#56 GPU-accelerated Guided Source Separation for Meeting Transcription

Описание: Authors: Desh Raj, Daniel Povey, Sanjeev Khudanpur
https://arxiv.org/pdf/2212.05271.pdf
  / gpu-accelerated-guided-source-separation-f...  
00:00 Could you tell us a little more about this paper?
00:09 My contribution was very modest
00:30 Source separation is basically if you have two speakers and you want to separate out the audio into speaker A and speaker B.
00:42 Source guided speech recognition, which is where you are told somehow which speaker you want to recognize.
01:23 MVDR is something that is used in a lot of source separation algorithms.
03:16 Time frequency mask is fed into the MVDR beamformer
07:00 How are you distinguishing these two people?
08:46 What's being transcribed when two people are talking at the same time?
10:56 Do we only want to decode the speaker with the red waves in this diagram?
#danielpovey #asr #speechprocessing