hope to see the support of personalized deep noise suppression

[wendongj]

thanks for such a great work, in DNS challenge, personalized speech enhancement is gradually replaced the non-personalized speech enhancement, this is a challenge and interesting task, since it need to use speaker recognition model to extract speaker embedding, which is already existed in We-Speaker

[StuartIanNaylor]

https://github.com/BUTSpeechFIT/speakerbeam

There was this paper that looks really interesting but can not find any code examples https://arxiv.org/abs/2106.02934

Also with Esspresif's laterial thought on BSS you could replace the KWS on each signal output with VAD that maybe uses ondevice training of a smaller model to bias to a voice profile?
VAD can often be lighter than KWS and maybe post process KWS or upstream.

I agree totally that extraction of a known than cancellation of the unknown seems to be the way forward and the above paper seems the only alternative to Googles VoiceFilterLite (Please opensource :) ) so that it can run downstream on lite hardware or concurrent multiple instances.

Also https://arxiv.org/abs/2106.04129 but again no code.