Open cgisky1980 opened 1 year ago
Hi, Do you use a universal or a personal model ? Also keep in mind that this fork (still) doesn't have any frontend processing, which means you likely need to add an additional automatic gain control and noise suppression step before feeding the data to snowman.
I tried personal model ,set Sensitivity to 0.4, it seems ok. but universal model is not so good.
Hi, Do you use a universal or a personal model ? Also keep in mind that this fork (still) doesn't have any frontend processing, which means you likely need to add an additional automatic gain control and noise suppression step before feeding the data to snowman.
anyway You did great job! I'll try to do someting for frontend processing.
Hi, Do you use a universal or a personal model ? Also keep in mind that this fork (still) doesn't have any frontend processing, which means you likely need to add an additional automatic gain control and noise suppression step before feeding the data to snowman.
pulseaudio has module-echo-cancel use webrtc
add something like :
load-module module-echo-cancel source_master=
it works great
I will try https://github.com/jagger2048/WebRtc_noise_suppression
it works great
Glad to hear that sorted out your issue.noise and echo cancellation are pretty important to avoid spurious, non voice activation. Long term it would be better to integrate them into the library as an optional preprocessing step similar to the original snowboy library, but for me it was never an issue because the mic arrays I use in my devices do AEC and noise supression in hardware.
I set it to 0.01, but there are still many wrong activations