How do we enable noise and background speech filtering? For example if I start talking to it, it needs to ignore any other voices it hears in the background like the devices from the big tech do. Right now, I can't use the voice on home assistant at all if anything else is talking in the background or it will just record indefinitely.
#noise filter and background speech?
indigo sleet
main skiff
How do we enable noise and background speech filtering? For example if I start t...
This is a known issue and not just something that can be "enabled" at the moment.
This is likely something that will need to be done on the audio processing before the audio gets to the pipeline. I know options are being looked at for this for the VPE but once a solution starts appearing for a specific setup it will become generalised.
This is just one of those things that the current voice setup options are behind the "big tech" devices in.
earnest wraith
How do we enable noise and background speech filtering? For example if I start t...
Noise is easy to suppress, but voices are not. We'll have to use resource-intensive neural networks that perform diarization (which requires modern graphics cards). A good move on the hardware side would be to release a new version of the satellite with beamforming for the microphone array. However, there are difficulties, as the OHF team doesn't have such technologies, and we'll have to rely again on closed-source products (for example, the XVF3800 from seeed).
If you launch the linux-voice-assist project yourself, you'll be able to choose a high-quality speakerphone (not entry-level).
frank token
Noise is easy to suppress, but voices are not. We'll have to use resource-intens...
To be honest, XVF3800 doesn't use its beam as well. It clearly has beam formed - but doesn't use it to cut background noise.
I reported this to Seeed, maybe they'll do something about it.
earnest wraith
To be honest, XVF3800 doesn't use its beam as well. It clearly has beam formed -...
Frankly, I wouldn't want OHF to use this module (I'm also not impressed by ReSpeaker's algorithms). Ideally, the team would start a project to develop a microphone array with open-source algorithms, so they could have complete control over this component, rather than relying on a subcontractor. But given that they haven't even decided to implement a standalone OTBR based on the ZBT-2, this probably isn't happening anytime soon.