Hacker News Reader

Comment by throwaway2016a | Hacker News Reader

throwaway2016a

I've been dying to replace my Echos with an open source smart speaker but half of them use AWS or Azure for test to speech and speech synthesis so really all you are in control of is the software that runs on the device itself. So this is a coo step in the right direction.

View on HN

The Rhasspy [0] author recently got hired by mycroft to work on satelites and fully local. Rhasspy requires a lot of manual work, but replacing Alexa is already possible. I’m somewhat stuck with the current hardware availability issues, but I have a Pi 3 satellite that does wakeword detection (this is supposed to be handled by Pi Zero 2 W in the future) and sends the voice to the MQTT server running on a PI 4, the data gets picked up by the Rhasspy instance also running there, it does STT, intent recognition, sends the intent to home assistant and then does TTS back to the satellite.

My main software issue is currently how to replicate the music functionality. Playing music at the satellite that requested it, lowering the volume when it recognizes the wakeword. Preselection of "commands" for band and genre names should be easily scriptable afterwards.

In a quiet room, I have no issues with wakeword detection using a playstation eye camera (I wanted the seed USB microhphone array, but between discovering it and starting with buying hardware the supply chain bit once again)

[0]: https://rhasspy.readthedocs.io/en/latest/

And how well does STT work when the room is not quiet anymore, e.g. when music is playing?

My understanding is, that the seeed array would work better than the PS eye, but for the volume I normally listen music at, it still works okay.

Yeah we aren't using the seeed array in the final Mark II. But we have used the same XMOS XVF-3510 to perform acoustic echo cancellation. That means, even with music blasting out of the speakers, you can still wake the device from across the room.

In a simple fashion you can think of it as subtracting the audio being output from the audio coming in from the microphone.

I was talking about Rhasspy though. As nice as your devices look, tiny satelites are mandatory for me to replace alexa.

Playing music from a Plex server is a major use case for me, and I have given up on Rhasspy because I couldn’t get all the pieces to work together (I have the mic array HAT and a Synology I can run recognition on). Do you have a write-up of your setup?

> Playing music from a Plex server is a major use case for me, and I have given up on Rhasspy because I couldn’t get all the pieces to work together (I have the mic array HAT and a Synology I can run recognition on). Do you have a write-up of your setup?

I have not yet managed / worked enough on it (the lack of HW making everything theoretical, which kills my motivation). The way I understand it, is that there’ll either be a casting server on the satelite, or a pulse audio/pipewire server reachable via network. But I have next to no experience with consumer linux, so the configuration of those parts is… hard.

But there are many tutorials for playing multi-room audio (with icecast or something), I just assumed it would be easier without multi-room as I don’t need it, but it turns out it’s not ;)

Didn't realize rhasspy already has satellite support. I shall have to check that out!

I've got a home server and a seed array so would be ideal to split that mic (rasp) and processing

They're also just really not great. I tested out Mycroft a couple years ago and found that the success rate for getting it to understand its wake word and listen for commands was under 10%. Maybe if you buy their prepackaged product, it works better, but that's not something I want to do. I just want to run it on a Pi 4 (which they claim works) with a mic array.

10% doesn't sound much worse than my Alexas' 30%...

And if it's anything like siri, it can barely do anything useful, so it doesn't matter if it understands you.

"I searched the web for 'shutup stop mute stop talking' and here is what I found on Wikipedia..."

counter point, I use my alexa daily and don't run into many issues with voice recognition, or it's lack of understanding.

Daily uses:

1) in the AM i set up all my needed reminders for 5 minutes before every meeting I have

2) it's connected to my hue bridge so I can turn off/on lights by asking while laying in bed, which is wonderful.

3) I play music all day.

4) It reminds me 10 minutes before every sunset to go outside for a walk.

I find Siri really useful - for a very limited set of tasks where recognition is about 100% and being hands free has a benefit. Typically this is starting exercise workouts and countdown timers. For more general tasks the recognition is still good (for me, seems to cary by voice) but even at 90% there will be one mistakes in most requests.

My Google home is pretty near 100%; I can count on one hand the number of times it hasn't "heard" me over the past year. That's my benchmark.

They’ve done a lot of work in the last year on the software side. Might be worth revisiting. They’re tentatively on track to (finally!) ship in September of this year.

Yeah I think there are two sides to this coin (and just for clarity - all of this relates to Picroft, not Mimic 3 the TTS engine that just launched). The audio hardware makes a huge difference to audio input which is why we've developed the custom SJ201 board that's in the Mark II. But even on DIY units we have been making big improvements on the wake word detection by better balancing our training data sets. Once the Mark II is shipping there are additional wake word improvements on the roadmap. Eventually the system will optimize for the users of each device. So the wake word model on your device wouldn't be exactly the same as the model on mine. We've also ported the Wake Word model to Tensorflow Lite which means it uses a small fraction of the system resources that it used to :D We're also about to make some bigger changes to mycroft-core that will help to support a broader range of hardware in a more consistent way. So whilst you could try it again today and I can guarantee it's better than the last time you used it, if you want a DIY system instead of a Mark II - I'd suggest adding a reminder to check it again in a couple of months once these bigger changes land.

That sounds fantastic! Thank you for replying; I'll definitely check back and give it another go.

Reply on news.ycombinator.com