sorry im complete new in ESP…looks like its working now
1 Like
I’ve had the Respeaker 3800 with the case for a few days now, and alongside it I also have the HA Voice Preview. I have to say I’m pretty disappointed with the Respeaker. Especially when there’s background noise, it performs much worse than the Voice. If the TV is loud, it doesn’t understand anything at all.
I don’t know if I might have done something wrong during setup, but the optimizations that Seeed Studio promises on their website don’t seem to have any effect, at least that’s how it feels to me.
I saw the same post on Seeed’s Discord server. Let’s see if they have something to say.
XVF3800 does behave badly when TV is on. That’s fact. I wish that beam forming would help distinguish the person speaking from other background noises.
On the other hand, XVF3800 hears me better from across my living room than Respeaker Lite or PE. When it’s reasonably quiet, I’m getting much better STT results with XVF3800. So now I use it in my living room despite the disadvantages (one of them is lack of buttons…).
Yes, I also hope they address the problem. I had been looking forward to it for a long time and was hoping it would make a difference. I have a large living area, and the ReSpeaker is about 8 meters away from the surround system and the TV in the kitchen. Even when I’m in the kitchen and the TV is at a medium volume, the problem already starts.
You’re right though — when it’s quiet, it works well and fast, probably even better than Home Assistant Voice.
1 Like
I got mine today, followed all the instructions, but I get no sound out of the device. It hears me, it runs HA commands no problem, but no sound comes out. I tested it by using it as a media source, and nothing.
Check that you didn’t alter config. And try both with JST speaker and 3.5mm jack, if you can.
Yeah, I tried with both, 'cause I have a set of spare Bose speakers that I set up next to where I place the voice assistant (had a Home Assistant Voice, found that it didn’t pick up my voice unless I turned around and yelled at it, was hoping the XVF3800 might pick me up better) and so after it didn’t work on its own, I tried plugging the speakers in to see if it was the internal speaker. No dice. And the only change I made to the Yaml config is to add Alexa as an option for the wake word, as I’m switching from Amazon and still automatically use that as my wake-up option.
What can I say… It seems like hardware problem to me… I never had any issues with speakers on this device. Well, after all the shenanigans with sample rate.
When I finish flashing it with the firmware, the LED’s pulse red. Could that have something to do with it?
Flashing DFU, or ESPHome?
If latter, try to flash former first.
I managed to figure out the problem. When I flashed the DFU firmware, it was failing to flash the part that handle the audio codec.
Hey @formatBCE, just got one of these and want to get it running. Sorry for being a noob, your readme.md isn’t super clear on how to start:
Edit: I found the link with the guide here: Smart Home Voice Control with Home Assistant | Seeed Studio Wiki
Just a request to add that to your repo’s readme.md then!
Finally installed, now getting:
[12:31:36][E][respeaker_xvf3800:040]: Could not find XVF3800 device on any tested address
[12:31:36][E][component:319]: respeaker_xvf3800 was marked as failed
[12:31:36][E][component:451]: respeaker_xvf3800 set Error flag: unspecified
Do I need to update the XVF firmware with DFU utils to solve this? Couldn’t find anybody else having this issue.
I had to solder on my own xiao chip (presoldered out of stock everywhere), i figured it was that at first, but multimeter tests seems to confirm everything is properly soldered…
Hey! You definitely want to flash DFU with dfuutil on just bought device. But the fact that it’s not found on I2C says that it’s not physically connected, most probably.
Ok will try that.
I can confirm they are connected though, Xiao GPIO5 and 6 go to XVF’s XOD37 and XOD38, straight on the chip, so it can’t be a connection issue.
Perhaps adress? That could be determined by firmware…
Well I2C address shouldn’t be different, it is dictated by firmware.
I still use automations to mute or take a snapshot of the TV volume , set the TV to a lower volume so it reliably understands me, then restores the snapshot using the same triggers as below. The trigger is “Satellite stared listening” and if you check mute check the checkbox after using the media player mute/unmute option. Then create another automation to unmute (unchecked) when the satellite starts thinking. The check mark is so it won’t mute your TV if it’s already muted and vice versa. That way it doesn’t hang on, if it has to go to my LLM for an answer it un-mutes while thinking so I don’t have to wait for it to start speaking before it un-mutes the TV
The top trigger is disabled as I created a template sensor before it was baked in as a trigger Honestly my [https://www.unfoldedcircle.com/](https://Unfolded circle) Remote 3has been super fast and accurate but it’s a remote control and you hold the button to talk then release and it’s not cheap but connects to the HA API to use almost all domains on the remote, along with other integrations but mostly setup via media players, scenes as activities and scripts to change any physical key functionality. Open API and all integrationsare open source. Hopefully the next model will be cheaper and more accessible.
8 have found overall the respeaker does the best at further distances but it’s also probably the same reason it holds onto the TV easier than the PE, not that the PE is great either. Honestly noise cancellation seems to be the last hurdle IMO. Wether it’s beam forming or some other method if you have 4 mics you should be able to isolate the person’s voice better. At the same time Amazon lost 25 billion on Alexa and Fire TV/tablets between 2018 and 2022 according to papers the New York Times got their hands on and most of it was from Alexa and keeping cloud servers up. Even with shifted hardware capabilities it’s still heavily cloud reliant. Running a GPU model of Whisper really helps for local only. PO iper can run on anything, it’s not resource intensive. Especially with streaming replies
I have blueprint for that.
1 Like
OK, got it working. I think it might have had something to do with the firmware, it’s a bit confusing how many versions there are, and my board was the standalone without Xiao soldered on.
I updated the XVF’s firmware (ask Claude/Cursor AI to help you out if needed), and then ran all their test arduino sketches (limited success, record/replay gave me noise blasts).
Then i ran an i2c scanner sketch a 0x2c started showing up.
It’s now working, but not as well as I hoped: it’s much more limited in range than the Home Assistant voice box I had before, I need to stand within 1m of it and speak very loud. The “sensitivity” for wake word is afaik not tied to gain, but rather how likely it is it will detect the wake word.
Is there anything i can do to boost mic gain? Right now it’s a bit disappointing, even my wife says it’s worse than before…
It shouldn’t be like this, and you shouldn’t have to raise gain. My device hears me from above 5 meters good, even when I speak normally.
Is it in a case? If so take it out. What wake word are you using? if not using “Okay Nabu” try that, it is by far the most trained and accurate wake word. Mine can easily hear me from 20 to 25 feet away in a pretty quiet environment (no TV). That is were is shines, long range accurate voice detection. It just tends to hold onto noises and keeps listening, TV in particular for me.
1 Like